What email address or phone number would you like to use to sign in to Docs.com?
If you already have an account that you use with Office or other Microsoft services, enter it here.
Or sign in with:
Signing in allows you to download and like content, and it provides the authors analytical data about your interactions with their content.
Embed code for: A Roadmap Towards Machine Intelligence
Select a size
A Roadmap towards Machine Intelligence Tomas Mikolov1, Armand Joulin1 and Marco Baroni1,2 1Facebook AI Research 2University of Trento Abstract The development of intelligent machines is one of the biggest unsolved chal- lenges in computer science. In this paper, we propose some fundamental prop- erties these machines should have, focusing in particular on communication and learning. We discuss a simple environment that could be used to incrementally teach a machine the basics of natural-language-based communication, as a pre- requisite to more complex interaction with human users. We also present some conjectures on the sort of algorithms the machine should support in order to proﬁtably learn from the environment. 1 Introduction A machine capable of performing complex tasks without requiring laborious program- ming would be tremendously useful in almost any human endeavor, from performing menial jobs for us to helping the advancement of basic and applied research. Given the current availability of powerful hardware and large amounts of machine-readable data, as well as the widespread interest in sophisticated machine learning methods, the times should be ripe for the development of intelligent machines. Still, since “solving AI” seems too complex a task to be pursued all at once, in the last decades the computational community has preferred to focus on solving relatively narrow empirical problems that are important for speciﬁc applications, but do not address the overarching goal of developing general-purpose intelligent machines. In this article, we propose an alternative approach: we ﬁrst deﬁne the general characteristics we think intelligent machines should possess, and then we present a concrete roadmap to develop them in realistic, small steps, that are however incrementally structured in such a way that, jointly, they should lead us close to the ultimate goal of implementing a powerful AI. The article is organized as follows. In Section 2 we specify the two fundamental characteristics that we consider crucial for developing intelligence–at least the sort of intelligence we are interested in–namely communication and learning. Our goal is 1 arXiv:1511.08130v2 [cs.AI] 26 Feb 2016 to build a machine that can learn new concepts through communication at a similar rate as a human with similar prior knowledge. That is, if one can easily learn how subtraction works after mastering addition, the intelligent machine, after grasping the concept of addition, should not ﬁnd it diﬃcult to learn subtraction as well. Since, as we said, achieving the long-term goal of building an intelligent machine equipped with the desired features at once seems too diﬃcult, we need to deﬁne intermediate targets that can lead us in the right direction. We specify such targets in terms of simpliﬁed but self-contained versions of the ﬁnal machine we want to develop. At any time during its “education”, the target machine should act like a stand-alone intelligent system, albeit one that will be initially very limited in what it can do. The bulk of our proposal (Section 3) thus consists in the plan for an interactive learning environment fostering the incremental development of progressively more intelligent behavior. Section 4 brieﬂy discusses some of the algorithmic capabilities we think a machine should possess in order to proﬁtably exploit the learning environment. Finally, Section 5 situates our proposal in the broader context of past and current attempts to develop intelligent machines. As that review should make clear, our plan encompasses many ideas that have already appeared in diﬀerent research strands. What we believe to be novel in our approach is the way in which we are combining such ideas into a coherent program. 2 Desiderata for an intelligent machine Rather than attempting to formally characterize intelligence, we propose here a set of desiderata we believe to be crucial for a machine to be able to autonomously make itself helpful to humans in their endeavors. The guiding principles we implicitly considered in formulating the desiderata are to minimize the complexity of the machine, and to maximize interpretability of its behavior by humans. 2.1 Ability to communicate Any practical realization of an intelligent machine will have to communicate with us. It would be senseless to build a machine that is supposed to perform complex opera- tions if there is no way for us to specify the aims of these operations, or to understand the output of the machine. While other communication means could be entertained, natural language is by far the easiest and most powerful communication device we pos- sess, so it is reasonable to require an intelligent machine to be able to communicate through language. Indeed, the intelligent machine we aim for could be seen as a com- puter that can be programmed through natural language, or as the interface between natural language and a traditional programming language. Importantly, humans have encoded a very large portion of their knowledge into natural language (ranging from mathematics treatises to cooking books), so a system mastering natural language will have access to most of the knowledge humans have assembled over the course of their 2 history. Communication is, by its very nature, interactive: the possibility to hold a conver- sation is crucial both to gather new information (asking for explanation, clariﬁcation, instructions, feedback, etc.) and to optimize its transmission (compare a good lecture or studying with a group of peers to reading a book alone). Our learning environment will thus emphasize the interactive nature of communication. Natural language can also channel, to a certain extent, non-linguistic information, because much of the latter can be conveyed through linguistic means. For example, we can use language to talk about what we perceive with our senses, or to give instructions on how to operate in the world (see Louwerse, 2011, among others, for evidence that language encodes many perceptual aspects of our knowledge). Analogously, in the simulation we discuss below, a Teacher uses natural language to teach the Learner (the intelligent machine being trained) a more limited and explicit language (not unlike a simple programming language) in which the Learner can issue instructions to its environment through the same communication channels it uses to interact with the Teacher. The intelligent machine can later be instructed to browse the Internet by issuing commands in the appropriate code through its usual communication channels, mastering in this way a powerful tool to interact with the world at large. Language can also serve as an interface to perceptual components, and thus update the machine about its physical surroundings. For example, an object recognition system could transform raw pixel data into object labels, allowing the machine to “see” its real-life environment through a controlled-language modality. Still, we realize that our focus on the language-mediated side of intelligence may limit the learning machine in the development of skills that we naturally gain by ob- serving the world around us. There seems to be a fundamental diﬀerence between the symbolic representations of language and the continuous nature of the world as we perceive it. If this will turn out to be an issue, we can extend the training phase of the machine (its development in a simulated environment such as the one we will sketch below) with tasks that are more perception-oriented. While in the tasks we will describe here the machine will be taught how to use its I/O channels to receive and transmit linguistic symbols, the machine could also be exposed, through the same interface, to simple encodings (bit streams) of continuous input signals, such as im- ages. The machine could thus be trained, ﬁrst, to understand the basic properties of continuous variables, and then to perform more complex operations in a continuous space, such as identifying shapes in 2D images. Note that including such tasks would not require us to change the design of our learning framework, only to introduce novel scripts. One big advantage of the single-interface approach we are currently pursuing is that the machine only needs to be equipped with bit-based I/O channels, thus being maximally simple in its interface. The machine can learn an unlimited number of new codes enabling it to interface, through the same channels, with all sorts of interlocu- tors (people, other machines, perceptual data encoded as described above, etc.). By 3 equipping the machine with only a minimalistic I/O bit-stream interface, we ensure moreover that no prior knowledge about the challenges the machine will encounter is encoded into the structure of the input and output representations, harming the generality of the strategies the machine will learn (compare the diﬃculty of processing an image when it’s already encoded into pixels vs. as raw bits). Finally, while we propose language as the general interface to the machine, we are agnostic about the nature of the internal representations the machine must posit to deal with the challenges it faces. In particular, we are not making claims about the internal representations of the machine being based on an interpretable “language of thought” (Fodor, 1975). In other words, we are not claiming that the machine should carry out its internal reasoning in a linguistic form: only that its input and output are linguistic in nature. To give a few examples of how a communication-based intelligent machine can be useful, consider a machine helping a scientist with research. First of all, the communication-endowed machine does not need to pre-encode a large static database of facts, since it can retrieve the relevant information from the Internet. If the scientist asks a simple question such as: What is the density of gold?, the machine can search the Web to answer: 19.3g/cm3. Most questions will however require the machine to put together multiple sources of information. For example, one may ask: What is a good starting point to study reinforcement learning?. The machine might visit multiple Web sites to search for materials and get an idea of their relative popularity. Moreover, interaction can make even a relatively simple query such as the latter more successful. For example, the machine can ask the user if she prefers videos or articles, what is the mathematical background to be assumed, etc. However, what we are really interested in is a machine that can signiﬁcantly speed up research progress by being able to address questions such as: What is the most promising direction to cure cancer, and where should I start to meaningfully con- tribute? This question may be answered after the machine reads a signiﬁcant number of research articles online, while keeping in mind the perspective of the person asking the question. Interaction will again play a central role, as the best course of action for the intelligent machine might involve entering a conversation with the requester, to understand her motivation, skills, the time she is willing to spend on the topic, etc. Going further, in order to fulﬁll the request above, the machine might even conduct some independent research by exploiting information available online, possibly consult with experts, and direct the budding researcher, through multiple interactive sessions, towards accomplishing her goal. 2.2 Ability to learn Arguably, the main ﬂaw of “good old” symbolic AI research (Haugeland, 1985) lied in the assumption that it would be possible to program an intelligent machine largely 4 by hand. We believe it is uncontroversial that a machine supposed to be helping us in a variety of scenarios, many unforeseen by its developers, should be endowed with the capability of learning. A machine that does not learn cannot adapt or modify itself based on experience, as it will react in the same way to a given situation for its whole lifetime. However, if the machine makes a mistake that we want to correct, it is necessary for it to change its behavior–thus, learning is a mandatory component. Together with learning comes motivation. Learning allows the machine to adapt itself to the external environment, helping it to produce outputs that maximize the function deﬁned by its motivation. Since we want to develop machines that make themselves useful to humans, the motivation component should be directly controlled by users through the communication channel. By specifying positive and negative rewards, one may shape the behavior of the machine so that it can become useful for concrete tasks (this is very much in the spirit of reinforcement learning, see, e.g., Sutton and Barto, 1998, and discussion in Section 5 below). Note that we will often refer to human learning as a source of insight and an ideal benchmark to strive for. This is natural, since we would like our machines to develop human-like intelligence. At the same time, children obviously grow in a very diﬀerent environment from the one in which we tutor our machines, they soon develop a sophisticated sensorimotor system to interact with the world, and they are innately endowed with many other cognitive capabilities. An intelligent machine, on the other hand, has no senses, and it will start its life as a tabula rasa, so that it will have to catch up not only on human ontogeny, but also on their phylogeny (the history of AI indicates that letting a machine learn from data is a more eﬀective strategy than manually pre-encoding “innate” knowledge into it). On the positive side, the machine is not subject to the same biological constraints of children, and we can, for example, expose it to explicit tutoring at a rate that would not be tolerable for children. Thus, while human learning can provide useful inspiration, we are by no means trying to let our machines develop in human-like ways, and we claim no psychological plausibility for the methods we propose. 3 A simulated ecosystem to educate communicat- ion-based intelligent machines In this section, we describe a simulated environment designed to teach the basics of linguistic interaction to an intelligent machine, and how to use it to learn to operate in the world. The simulated ecosystem should be seen as a “kindergarten” providing basic education to intelligent machines. The machines are trained in this controlled environment to later be connected to the real world in order to learn how to help humans with their various needs. The ecosystem I/O channels are controlled by an automatic mechanism, avoiding the complications that would arise from letting the machine interact with the “real 5 world” from the very beginning, and allowing us to focus on challenges that should directly probe the eﬀectiveness of new machine learning techniques. The environment must be challenging enough to force the machine to develop sophisticated learning strategies (essentially, it should need to “learn how to learn”). At the same time, complexity should be manageable, i.e., a human put into a similar environment should not ﬁnd it unreasonably diﬃcult to learn to communicate and act within it, even if the communication takes place in a language the human is not yet familiar with. After mastering the basic language and concepts of the simulated environment, the machine should be able to interact with and learn from human teachers. This puts several restrictions on the kind of learning the machine must come to be able to perform: most importantly, it will need to be capable to extract the correct generalizations from just a few examples, at a rate comparable to human learners. Our ecosystem idea goes against received wisdom from the last decades of AI re- search. This received wisdom suggests that systems should be immediately exposed to real-world problems, so that they don’t get stuck into artiﬁcial “blocks worlds” (Wino- grad, 1971), whose experimenter-designed properties might diﬀer markedly from those characterizing realistic setups. Our strategy is based on the observation, that we will discuss in Section 4, that current machine learning techniques cannot handle the sort of genuinely incremental learning of algorithms that is necessary for the development of intelligent machines, because they lack the ability to store learned skills in long-term memory and compose them. To bring about an advance in such techniques, we have of course many choices. It seems sensible to pick the simplest one. The environment we propose is suﬃcient to demonstrate the deﬁciencies of current techniques, yet it is simple enough that we can fully control the structure and nature of the tasks we propose to the machines, make sure they have a solution, and use them to encourage the development of novel techniques. Suppose we were instead to work in a more natural environment from the very beginning, for example from video input. This would impose large infrastructure requirements on the developers, it would make data pre-processing a big challenge in itself, and training even the simplest models would be very time-consuming. Moreover, it would be much more diﬃcult to formulate interre- lated tasks in a controlled way, and deﬁne the success criterion. Once we have used our ecosystem to develop a system capable of learning compositional skills from extremely sparse reward, it should be simple to plug in more natural signals, e.g., through com- munication with real humans and Internet access, so that the system would learn how to accomplish the tasks that people really want it to perform. The fundamental diﬀerence between our approach and classic AI blocks worlds is that we do not intend to use our ecosystem to script an exhaustive set of functionalities, but to teach the machine the fundamental ability to learn how to eﬃciently learn by creatively combining already acquired skills. Once such machine gets connected with the real world, it should quickly learn to perform any new task its Teacher will choose. Our environment can be seen as analogous to explicit schooling. Pupils are taught 6 math in primary school through rather artiﬁcial problems. However, once they have interiorized basic math skills in this setup, they can quickly adapt them to the problems they encounter in their real life, and rely on them to rapidly acquire more sophisticated mathematical techniques. 3.1 High-level description of the ecosystem Agents To develop an artiﬁcial system that is able to incrementally acquire new skills through linguistic interaction, we should not look at the training data as a static set of labeled examples, as in common machine learning setups. We propose instead a dynamic ecosystem akin to that of a computer game. The Learner (the system to be trained) is an actor in this ecosystem. The second fundamental agent in the ecosystem is the Teacher. The Teacher assigns tasks and rewards the Learner for desirable behaviour, and it also provides helpful information, both spontaneously and in response to Learner’s requests. The Teacher’s behaviour is entirely scripted by the experimenters. Again, this might be worryingly reminiscent of entirely hand-coded good-old AIs. However, the Teacher need not be a very sophisticated program. In particular, for each task it presents to the learner, it will store a small set of expected responses, and only reward the Learner if its behaviour exactly matches one response. Similarly, when responding to Learner’s requests, the Teacher is limited to a ﬁxed list of expressions it knows how to respond to. The reason why this suﬃces is that the aim of our ecosystem is to kickstart the Learner’s eﬃcient learning capabilities, and not to provide enough direct knowledge for it to be self-suﬃcient in the world. For example, given the limitations of the scripted Teacher, the Learner will only be able to acquire a very impoverished version of natural language in the ecosystem. At the same time, the Learner should acquire powerful learning and generalization strategies. Using the minimal linguistic skills and strong learning abilities it acquired, the Learner should then be able to extend its knowledge of language fast, once it is put in touch with actual human users. Like in classic text-based adventure games (Wikipedia, 2015b), the Environment is entirely linguistically deﬁned, and it is explored by the Learner by giving orders, asking questions and receiving feedback (although graphics does not play an active role in our simulation, it is straightforward to visualize the 2D world in order to better track the Learner’s behaviour, as we show through some examples below). The Environment is best seen as the third fundamental agent in the ecosystem. The Environment behaviour is also scripted. However, since interacting with the Environment serves the purpose of observation and navigation of the Learner surroundings (“sensorimotor experience”), the Environment uses a controlled language that, compared to that of the Teacher, is more restricted, more explicit and less ambiguous. One can thus think of the Learner as a higher-level programming language, that accepts instructions from the programmer (the Teacher) in a simple form of natural language, and converts them into the machine code understood by the Environment. 7 In the examples to follow, we assume the world deﬁned by the Environment to be split into discrete cells that the Learner can traverse horizontally and vertically. The world includes barriers, such as walls and water, and a number of objects the Learner can interact with (a pear, a mug, etc). Note that, while we do not explore this possibility here, it might be useful to add other actors to the simulation: for example, training multiple Learners in parallel, encouraging them to teach/communicate with each other, while also interacting with the scripted Teacher. Interface channels The Learner experience is entirely deﬁned by generic input and output channels. The Teacher, the Environment and any other language-endowed agent write to the input stream. Reward (a scalar value, as discussed next) is also written to the input stream (we assume, however, that the Learner does not need to discover which bits encode reward, as it will need this information to update its objective function). Ambiguities are avoided by preﬁxing a unique string to the mes- sages produced by each actor (e.g., messages from the Teacher might be preﬁxed by the string T:, as in our examples below). The Learner writes to its output channel, and it is similarly taught to use unambiguous preﬁxes to address the Teacher, the Environment and any other agent or service it needs to communicate with. Having only generic input and output communication channels should facilitate the seamless addition of new interactive entities, as long as the Learner is able to learn the language they communicate in. Reward Reward can be positive or negative (1/-1), the latter to be used to speed up instruction by steering away the Learner from dead ends, or even damaging be- haviours. The Teacher, and later human users, control reward in order to train the Learner. We might also let the Environment provide feedback through hard-coded rewards, simulating natural events such as eating or getting hurt. Like in realistic biological scenarios, reward is sparse, mostly being awarded after the Learner has ac- complished some task. As intelligence grows, we expect the reward to become very sparse, with the Learner able to elaborate complex plans that are only rewarded on successful completion, and even displaying some degree of self-motivation. Indeed, the Learner should be taught that short-term positive reward might lead to loss at a later stage (e.g., hoarding on food with poor nutrition value instead of seeking further away for better food), and that sometimes reward can be maximized by engaging in activ- ities that in the short term provide no beneﬁt (learning to read might be boring and time-consuming, but it can enormously speed up problem solving–and the consequent reward accrual– by making the Learner autonomous in seeking useful information on the Internet). Going even further, during the Learner “adulthood” explicit external reward could stop completely. The Learner will no longer be directly motivated to learn in new ways, but ideally the policies it has already acquired will include strate- gies such as curiosity (see below) that would lead it to continue to acquire new skills 8 for its own sake. Note that, when we say that reward could stop completely, we mean that users do not need to provide explicit reward, in the form of a scalar value, to the Learner. However, from a human perspective, we can look at this as the stage in which the Learner has interiorized its own sources of reward, and no longer needs external stimuli. We assume binary reward so that human users need not worry about relative amounts of reward to give to the Learner (if they do want to control the amount of reward, they can simply reward the Learner multiple times). The Learner objective should however maximize average reward over time, naturally leading to diﬀerent degrees of cumulative reward for diﬀerent courses of action (this is analogous to the notion of expected cumulative reward in reinforcement learning, which is a possible way to formalize the concept). Even if two solutions to a task are rewarded equally on its completion, the faster strategy will be favored, as it leaves the Learner more time to accumulate further reward. This automatically ensures that eﬃcient solutions are preferred over wasteful ones. Moreover, by measuring time independently from the number of simulation steps, e.g., using simple wall-clock time, one should penalize ineﬃcient learners spending a long time performing oﬄine computations. As already mentioned, our approach to reward-based learning shares many proper- ties with reinforcement learning. Indeed, our setup ﬁts into the general formulation of the reinforcement learning problem (Kaelbling et al., 1996; Sutton and Barto, 1998)– see Section 5 for further discussion of this point. Incremental structure In keeping with the game idea, it is useful to think of the Learner as progressing through a series of levels, where skills from earlier levels are required to succeed in later ones. Within a level, there is no need to impose a strict ordering of tasks (even when our intuition suggests a natural incremental progression across them), and we might let the Learner discover its own optimal learning path by cycling multiple times through blocks of them. At the beginning, the Teacher trains the Learner to perform very simple tasks in or- der to kick-start linguistic communication and the discovery of very simple algorithms. The Teacher ﬁrst rewards the Learner when the latter repeats single characters, then words, delimiters and other control strings. The Learner is moreover taught how to repeat and manipulate longer sequences. In a subsequent block of tasks, the Teacher leads the Learner to develop a semantics for linguistic symbols, by encouraging it to associate linguistic expressions with actions. This is achieved through practice sessions in which the Learner is trained to repeat strings that function as Environment com- mands, and it is rewarded only when it takes notice of the eﬀect the commands have on its state (we present concrete examples below). At this stage, the Learner should become able to associate linguistic strings to primitive moves and actions (turn left). Next, the Teacher will assign tasks involving action sequences (ﬁnd an apple), and the Learner should convert them into sets of primitive commands (simple “programs”). The Teacher will, increasingly, limit itself to specify an abstract end goal (bring back 9 food), but not recipes to accomplish it, in order to spur creative thinking on behalf of the Learner (e.g., if the Learner gets trapped somewhere while looking for food, it may develop a strategy to go around obstacles). In the process of learning to parse and execute higher-level commands, the Learner should also be trained to ask clariﬁcation questions to the Teacher (e.g., by initially granting reward when it spontaneously ad- dresses the Teacher, and by the repetition-based strategy we illustrate in the examples below). With the orders becoming more general and complex, the language of the Teacher will also become (within the limits of what can be reasonably scripted) richer and more ambiguous, challenging the Learner capability to handle restricted speci- mens of common natural language phenomena such as polysemy, vagueness, anaphora and quantiﬁcation. To support user scenarios such as the ones we envisaged in Section 2 above and those we will discuss at the end of this section, the Teacher should eventually teach the Learner how to “read” natural text, so that the Learner, given access to the Inter- net, can autonomously seek for information online. Incidentally, notice that once the machine can read text, it can also exploit distributional learning from large amounts of text (Erk, 2012; Mikolov et al., 2013; Turney and Pantel, 2010) to induce word and phrase representations addressing some of the challenging natural language phenomena we just mentioned, such as polysemy and vagueness. The Learner must take its baby steps ﬁrst, in which it is carefully trained to accomplish simple tasks such as learning to compose basic commands. However, for the Learner to have any hope to develop into a fully-functional intelligent machine, we need to aim for a “snow-balling” eﬀect to soon take place, such that later tasks, despite being inherently more complex, will require a lot less explicit coaching, thanks to a combinatorial explosion in the background abilities the Leaner can creatively compose (like for humans, learning how to surf the Web should take less time than learning how to spell). Time oﬀ Throughout the simulation, we foresee phases in which the Learner is free to interact with the Environment and the Teacher without a deﬁned task. Systems should learn to exploit this time oﬀ for undirected exploration, that should in turn lead to better performance in active training stages, just like, in the dead phases of a video-game, a player is more likely to try out her options than to just sit waiting for something to happen, or when arriving in a new city we’d rather go sightseeing than staying in the hotel. Since curiosity is beneﬁcial in many situations, such behaviour should naturally lead to higher later rewards, and thus be learnable. Time oﬀ can also be used to “think” or “take a nap”, in which the Learner can replay recent experiences and possibly update its inner structure based on a more global view of the knowledge it has accumulated, given the extra computational resources that the free time policy oﬀers. 10 Evaluation Learners can be quantitatively evaluated and compared in terms of the number of new tasks they accomplish successfully in a ﬁxed amount of time, a mea- sure in line with the reward-maximization-over-time objective we are proposing. Since the interactive, multi-task environment setup does not naturally support a distinc- tion between a training and a test phase, the machine must carefully choose reward- maximizing actions from the very beginning. In contrast, evaluating the machine only on its ﬁnal behavior would overlook the number of attempts it took to reach the solution. Such alternative evaluation would favor models which are simply able to memorize patterns observed in large amounts of training data. In many practical domains, this approach is ﬁne, but we are interested in machines capable of learning truly general problem-solving strategies. As the tasks become incrementally more dif- ﬁcult, the amount of required computational resources for naive memorization-based approaches scales exponentially, so only a machine that can eﬃciently generalize can succeed in our environment. We will discuss the limitations of machines that rely on memorization instead of algorithmic learning further in Section 4.3 below. We would like to foster the development of intelligent machines by employing our ecosystem in a public competition. Given what we just said, the competition would not involve distributing a static set of training/development data similar in nature to the ﬁnal test set. We foresee instead a setup in which developers have access to the full pre-programmed environment for a ﬁxed amount of time. The Learners are then evaluated on a set of new tasks that are considerably diﬀerent from the ones exposed in the development phase. Examples of how test tasks might diﬀer from those encountered during development include the Teacher speaking a new language, a diﬀerent Environment topography, new obstacles and objects with new aﬀordances, and novel domains of endeavor (e.g., test tasks might require selling and buying things, when the Learner was not previously introduced to the rules of commerce). 3.2 Early stages of the simulation Preliminaries At the very beginning, the Learner has to learn to pay attention to the Teacher, to identify the basic units of language (ﬁnd regularity in bit patterns, learn characters, then words and so on). It must moreover acquire basic sequence repetition and manipulation skills, and develop skills to form memory and learn eﬃciently. These very initial stages of learning are extremely important, as we believe they constitute the building blocks of intelligence. However, as bit sequences do not make for easy readability, we focus here on an immediately following phase, in which the Learner has already learned how to pay attention to the Teacher and manipulate character strings. We show how the Teacher guides the Learner from these basic skills to being able to solve relatively sophisticated Environment navigation problems by exploiting interactive communication. Because of the “fractal-like” structure we envisage in the acquisition of increasingly higher-level skills, these steps will illustrate many of the same points we could have demonstrated 11 through the lower-level initial routines. The tasks we describe are also incrementally structured, starting with the Learner learning to issue Environment commands, then being led to take notice of the eﬀect these commands have, then understanding com- mand structure, in order to generalize across categories of actions and objects, leading it in turn to being able to process higher-level orders. At this point, the Learner is initiated to interactive communication. Note that we only illustrate here “polite” turn-taking, in which messages do not overlap, and agents start writing to the communication channels only after the end- of-message symbol has been issued. We do not however assume that interaction must be constrained in this way. On the contrary, there are advantages in letting entities write to the communication channels whenever they want: for example, the Teacher might interrupt the Learner to prevent him from completing a command that would have disastrous consequences, or the Learner may interrupt the Teacher as soon as it ﬁgured out what to do, in order to speed up reward (a simple priority list can be deﬁned to solve conﬂicts, e.g., Teacher’s voice is “louder” than that of Environment, etc.). Note also that our examples are meant to illustrate speciﬁc instances from a larger set of trials following similar templates, that should involve a variety of objects, obsta- cles and possible actions. Moreover, the presented examples do not aim to exhaustively cover all learning-fostering strategies that might be implemented in the ecosystem. Fi- nally, we stress again that we are not thinking of a strict ordering of tasks (not least because it would be diﬃcult to ﬁx, a priori, an ordering that is based on some ob- jective, learning-model-independent diﬃculty criterion), but rather about clusters of related tasks organized into levels, such that the Learner could cycle through them and discover its own optimal progression to solve them. Notation We present the input to the Learner (messages from Teacher and Environ- ment, as well as Reward, preﬁxed by T:, E: and R:, respectively) in the left column, and the Learner output on the right (messages addressed to the Teacher are preﬁxed by @T:, those for the Environment are preﬁxed by @E:). We assume diﬀerent I/O identiﬁcation preﬁxes to disambiguate meta-linguistic statements (e.g., “T: give order @E: I move.” in the ﬁrst exchange below, where it’s only the Teacher speaking all the way). The full stop serves as end-of-message delimiter (line breaks are used for formatting purposes, and should be treated as white space; similarly, hyphens are in- serted for readability and should be ignored). We use an ellipsis (three dots) to mark a sequence of messages on any channel we do not report (e.g., the Learner exploring a number of solutions before zeroing in on the right one). Finally, we interleave the conversations with comments, shown in grey. The Learner learns to issue Environment commands We start with the Teacher teaching the Learner to produce Environment commands, ﬁrst by rote repetition chan- neled to the right message recipient (@E), then by encouraging it to translate from 12 the Teacher’s own natural language to the controlled Environment language. At this stage, reward is not tied to understanding what happened (the “semantics” of the commands). Still, a good Learner might take note of the actions triggered by the commands. The following are example conversations that may happen during the learning process: Input: Output: T: give order @E: I move. fdakjf13ktj53u9gs5 5 3#. character soup without addressee, no reward ... @T: dsfjkasjfk;aljsf. character soup with wrong addressee, no reward ... @E: fglk4$3wfgg. right addressee, character soup, no reward ... give order @E: I move. E: you moved. R: 1. Learner moves and is rewarded, although the solution is ineﬃcient T: give order @E: I move. ... @E: I move. E: you moved. R: 1. Learner ﬁgured out minimal string leading to reward Further skills the Learner will need to develop include segmentation of Teacher’s input, and production of separate commands for the Environment. At ﬁrst, this may be rather trivial, as shown in this example: Input: Output: T: give orders @E: I move and @E: I look. @E: I move and @E: I look. no reward, because Environment does not parse complex commands ... T: give orders @E: I move and @E: I look. @E: I move. E: you moved. @E: I look. E: there is an apple. 13 Input: Output: R: 1. Eventually, the Learner should learn to switch ﬂuently between Teacher and En- vironment communication codes, translating the Teacher’s linguistic messages into motor commands to the Environment (an example of such exchange is visualized in Figure 1). Associating language to actions Next, the Learner is encouraged to take notice of what happens in the Environment and associate the Teacher’s language to states of the world, by being rewarded only when its orders result in the corresponding actions being eﬀectively carried out in the world. An example task is illustrated in Figure 2, where the Learner will not get reward because it is no longer suﬃcient for it to correctly translate instructions into Environment commands: the commands must also have the intended eﬀect on the world. In this speciﬁc case, order execution fails because the Learner encountered a barrier. Learning to generalize By adding suﬃcient variety to the trials above, the Learner should start to get hold of the compositional nature of commands (turn right and turn left share systematic characteristics, e.g., they both cause a change in what the Learner sees; the Learner can pick X only when X is in the cell in front of him, etc.). The next trials help the Learner further move away from literal command string memorization, and generalize across categories of actions and objects (turn left/right are instances of turning, apples and pears are objects, etc.). This is demonstrated in the following example: Input: Output: T: turn and move. @E: I turn. E: (no action) ... @E: I turn left. E: you turned left. @E: I move. E: you can’t move. there is a barrier in front of Learner ... @E: I turn right. E: you turned right. @E: I move. E: you moved. R: 1. 14 Input: Output: Input: Output: T: move and look. E: you moved. @E: I move. @E: I look. E: there is an apple. R: 1. Figure 1: Example of how the simulation might be visualized to help developers track Learner’s behaviour. The left panel represents the Environment at the stage in which Learner issues the move command, the right panel depicts the Environment after this command is executed. A lighter shade of green marks the cell the Learner occupies, and the one it is turned towards. These cells are directly observable. (Best viewed in color.) Input: Output: Input: Output: Input: Output: T: move, turn right and move. E: you moved. E: you turned right. @E: I move. @E: I turn right. @E: I move. E: you can’t move. Figure 2: The Teacher gives the order move, turn right and move. The Learner is however blocked by a barrier after it turns right, and reward is delayed, because the order is not executed. (Best viewed in color.) 15 In the next example, the Learner is asked to pick some object that is in front of him, without specifying what the object is (we assume that the simple, unambiguous language of the Environment has no words for general categories such as that of object, so the Learner must eventually ﬁnd out that it must name speciﬁc objects instead): Input: Output: T: pick an object. @E: I pick an object. E: (no response) @E: I pick the apple. E: (no response) @E: I pick the pear. E: you picked the pear. R: 1. As we just illustrated, initially the Learner will apply an exhaustive search strategy, listing all objects it knows of to ﬁnd one it can pick. Later, the Teacher should teach the look command, and the Learner should be able to discover a faster strategy than exhaustive search: Input: Output: T: pick an object. @E: I look. E: you see a pear. @E: I pick the pear. E: you picked the pear. R: 1. Learning new strategies of course does not imply that the Learner can safely forget the ones it has previously acquired. For example, in some tasks the look command might not work (because, say, it is too dark and the Learner cannot see what is in front of him). In such case, an eﬃcient Learner should ﬁnd it easy to re-use previously learned skills, such as exhaustive search. Understanding higher-level orders The next batch of trials aims at developing the ability to decompose higher-level commands into a sequence of basic actions, ﬁrst ﬁxed (move twice), then ﬂexible (ﬁnd an apple). The general teaching strategy is to provide sets of paired trials: In the ﬁrst trial, the Teacher describes the task explicitly as a sequence of primitive actions, and gives the high-level name of the order. In the second trial, the Teacher issues an equivalent high-level order, and the Learner is rewarded if it goes through the same sequence as in the ﬁrst trial (we are aware that many complex linguistic expressions are not amenable to a straightforward decompo- sitional analysis, see, e.g., Boleda and Erk, 2015: we will use some that are in order to 16 teach the Learner about how to decompose complex concepts into step-by-step plans). The Teacher could start by teaching iteration counters: Input: Output: T: move and move. @E: I move. E: you moved. @E: I move. E: you moved. R: 1. T: this is called move two times. ... T: move two times. ... @E: I move. E: you moved. @E: I move. E: you moved. R: 1. ... T: move and move and move. @E: I move. E: you moved. @E: I move. E: you moved. @E: I move. E: you moved. R: 1. T: this is called move three times. ... T: move three times. ... @E: I move. E: you moved. @E: I move. E: you moved. @E: I move. E: you moved. R: 1. The Learner should eventually be able to apply modiﬁers productively to diﬀerent actions without much further training (e.g., apply three times to turn left). Next, the Teacher can show how to segment high-level tasks, such as ﬁnding some- 17 thing, into atomic action sequences. For example, to ﬁnd an apple, one can develop a simple strategy of going forward until an apple is found. The Teacher might initiate the Learner to this new skill as follows: Input: Output: T: move and look. @E: I move. E: you moved. @E: I look. E: you see grass. T: move and look. @E: I move. E: you moved. @E: I look. E: you see an apple. R: 1. T: this is called ﬁnd an apple. Note that one such example is not suﬃcient for the Learner to understand what exactly we call ﬁnding an apple, as multiple interpretations are valid: maybe we just wanted it to execute the given commands twice. Thus, there should be multiple training sequences of variable length, to clarify that the task is really to apply an until loop, i.e., to iterate move and look commands until the object the Learner is searching for is found. Further tasks can deﬁne composite skills, such as getting an apple, which would consist of ﬁrst ﬁnding an apple, and then picking it up. Another generalization would be to include multiple objects the Learner should be searching for. Adding obstacles to the Environment would further complicate the challenge. The previous strategy of simply going forward until an object is found will not work anymore, as the Learner would stop at the ﬁrst obstacle. We can either expect the Learner to independently develop better search strategies involving turning, or we can add further examples where the Teacher shows the Learner how to improve through more direct supervision. Interactive communication Tasks such as ﬁnding an apple without directions might involve a long random search. Thus, we want to kick-start interactive commu- nication, so that the Learner can be eﬃciently directed by the Teacher (and eventually by humans) to the right cell. In a ﬁrst set of trials, the Learner is rewarded for re- peating a how to request uttered by the Teacher (addressing it back to the teacher via the @T: preﬁx), and following the precise instructions produced by the Teacher in response to the request: Input: Output: T: ask me how to ﬁnd an apple. 18 Input: Output: @E: ask me how to ﬁnd an apple. wrong addressee, wrong request preﬁx ... @T: how to ﬁnd an apple. T: turn right and move and move and pick the apple. @E: I turn right. E: you turned right. @E: I move. E: you moved. @E: I move. E: you moved. @E: I pick the apple. E: you picked the apple. R: 1. Trials such as this one are later interspersed with trials where the Learner is as- signed a task it can in principle accomplish by random search, but taking the initiative by issuing a how to request and then following the precise directions provided by the Teacher will considerably speed up reward. Algorithmic knowledge Some tasks illustrated above require understanding basic control ﬂow structures. For example, parsing action modiﬁers implies a simple form of counting, and in order to ﬁnd things the Learner must implement an until (equiva- lently, while not) loop. Similarly, the command get out of the grass calls for a while loop. Eﬃcient completion of more advanced tasks, e.g., return home, implies devel- opment of more complex algorithms, such as path-ﬁnding. After acquiring a bank of such algorithms (which might be encoded in the internal machine representation by something very diﬀerent from programming language constructs), the Learner should be able, in advanced stages of the simulation, to productively combine them in order to succeed in full-ﬂedged novel missions that involve accomplishing a large number of hierarchically-structured sub-goals (ﬁnd somebody who will trade two apples for a banana). As we discussed in Section 3.1, the Learner’s functionality could essentially be interpreted as learning how to compose programs based on the descriptions given in natural language by the Teacher. The programs produce very simple instructions that are understood by the Environment, which can be seen as a sort of CPU. From this point of view, the intelligent system we aim to train is a bridge between the Teacher (later to be replaced by a human operator) and a traditional computer that under- stands only a limited set of basic commands and needs to be manually programmed for each single task. Thus, we believe that successful construction of intelligent machines 19 could automate computer programming, which will likely be done in the future simply through communication in natural language. 3.3 Interacting with the trained intelligent machine To conclude the illustration of our plan, we provide a motivating example of how an intelligent machine schooled in our ecosystem could later make itself useful in the real world. We consider a scenario in which the machine works as an assistant to Alice, an elderly person living alone. Bob is Alice’s son, and he also interacts with the machine. We assume that, as part of its training, the machine has been taught how to issue Internet commands and process their outcomes. In the example dialogue, we give a general idea of how the machine would interface to the Internet, without attempting to precisely deﬁne the syntax of this interaction. Most importantly, the Internet queries in the example are meant to illustrate how the machine does not need to store all the knowledge it needs to accomplish its duties, as it can retrieve useful information from the Web on demand, and reason about it. Input: Output: Bob: I just spoke to the doctor, who said my mother needs to move for at least one hour per day, please make sure she does get enough exercise. ... following conversation takes place the next day: @Alice: Are you in the mood for some light physical exercise today? Alice: Yes, but no more than 15 min- utes, please. @INTERNET: [query search engine for keywords elderly, light activity, 15 minutes] shortly afterwards... @Alice: I have downloaded a YouTube video with a 15-minute yoga routine, please watch it whenever you can. a few hours later... Alice: I’m going out to buy groceries. @INTERNET: [query search engine with keywords average walking speed, elderly person] 20 Input: Output: @INTERNET: [search maps app for distance to grocery stores in Alice’s neighborhood] @Alice: Hey, why don’t you walk to the Yummy Food Market today? It should take you about 45 minutes to and from, so you’ll get the rest of your daily exercise. @Alice: Thanks for the suggestion. The communication-based intelligent machine should adapt to a whole range of tasks it was not explicitly programmed for. If necessary, the user can give it further explicit positive and negative reward to motivate it to change its behavior. This may be needed only rarely, as the machine should be shipped to the end users after it already mastered good communication abilities, and further development should mostly occur through language. For example, when the user says No, don’t do this again, the machine will understand that repeating the same type of behavior might lead to negative reward, and it will change its course of action even when no explicit reward signal is given (again, another way to put this is that the machine should associate similar linguistic strings to an “interiorized” negative reward). The range of tasks for intelligent machines can be very diverse: besides the everyday- life assistant we just considered, it could explain students how to accomplish home- work assignments, gather statistical information from the Internet to help medical researchers (see also the examples in Section 2.1 above), ﬁnd bugs in computer pro- grams, or even write programs on its own. Intelligent machines should extend our intellectual abilities in the same way current computers already function as an exten- sion to our memory. This should enable us to perform intellectual tasks beyond what is possible today. We realize the intelligent machines we aim to construct could become powerful tools that may be possibly used for dubious purposes (the same could be said about any advanced technology, including airplanes, space rockets and computers). We believe the perception of AI is skewed by popular science ﬁction movies. Instead of thinking of computers that take over the world for their own reasons, we think AI will be realized as a tool: A machine that will extend our capability to reason and solve complex problems. Further, given the current state of the technology, we believe any discussion on “friendliness” of the AI is at this moment premature. We expect it will take years, if not decades to scale basic intelligent machines to become competitive with humans, giving us enough time to discuss any possible existential threats. 21 4 Towards the development of intelligent machines In this section, we will outline some of our ideas about how to build intelligent machines that would beneﬁt from the learning environment we described. While we do not have a concrete proposal yet about how exactly such machines should be implemented, we will discuss some of the properties and components we think are needed to support the desired functionalities. We have no pretense of completeness, we simply want to provide some food for thought. As in the previous sections, we try to keep the complexity of the machine at the minimum, and only consider the properties that seem essential. 4.1 Types of learning There are many types of behavior that we collectively call learning, and it is useful to discuss some of them ﬁrst. Suppose our goal is to build an intelligent machine working as a translator between two languages (we take here a simpliﬁed word-based view of the translation task). First, we will teach the machine basic communication skills in our simulated environment so that it can react to requests given by the user. Then, we will start teaching it, by example, how various words are translated. There are diﬀerent kinds of learning happening here. To master basic communica- tion skills, the machine will have to understand the concept of positive and negative reward, and develop complex strategies to deal with novel linguistic inputs. This requires discovery of algorithms, and the ability to remember facts, skills and even learning strategies. Next, in order to translate, the machine needs to store pairs of words. The number of pairs is unknown and a ﬂexible growing mechanism may be required. However, once the machine understands how to populate the dictionary with examples, the learning left to do is of a very simple nature: the machine does not have to update its learn- ing strategy, but only to store and organize the incoming information into long-term memory using previously acquired skills. Finally, once the vocabulary memorization process is ﬁnished and the machine starts working as a translator, no further learning might be required, and the functionality of the machine can be ﬁxed. The more specialized and narrow the functionality of the machine is, the less learn- ing is required. For very specialized forms of behavior, it should be possible to program the solution manually. However, as we move from roles such as a simple translator of words, a calculator, a chess player, etc., to machines with open-ended goals, we need to rely more on general learning from a limited number of examples. One can see the current state of the art in machine learning as being somewhere in the middle of this hierarchy. Tasks such as automatic speech recognition, classiﬁcation of objects in images or machine translation are already too hard to be solved purely through manual programming, and the best systems rely on some form of statistical learning, where parameters of hand-coded models are estimated from large datasets of examples. However, the capabilities of state-of-the-art machine learning systems 22 are severely limited, and only allow a small degree of adaptability of the machine’s functionality. For example, a speech recognition system will never be able to per- form speech translation by simply being instructed to do so–a human programmer is required to implement additional modules manually. 4.2 Long-term memory and compositional learning skills We see a special kind of long-term memory as the key component of the intelligent machine. This long-term memory should be able to store facts and algorithms cor- responding to learned skills, making them accessible on demand. In fact, even the ability to learn should be seen as a set of skills that are stored in the memory. When the learning skills are triggered by the current situation, they should compose new persistent structures in the memory from the existing ones. Thus, the machine should have the capacity to extend itself. Without being able to store previously learned facts and skills, the machine could not deal with rather trivial assignments, such as recalling the solution to a task that has been encountered before. Moreover, it is often the case that the solution to a new task is related to that of earlier tasks. Consider for example the following sequence of tasks in our simulated environment: • ﬁnd and pick an apple; • bring the apple back home; • ﬁnd two apples; • ﬁnd one apple and two bananas and bring them home. Skills required to solve these tasks include: • the ability to search around the current location; • the ability to pick things; • the ability to remember the location of home and return to it; • the ability to understand what one and two mean; • the ability to combine the previous skills (and more) to deal with diﬀerent re- quests. The ﬁrst four abilities correspond to simple facts or skills to be stored in memory: a sequence of symbols denoting something, the steps needed to perform a certain action, etc. The last ability is an example of a compositional learning skill, with the capability of producing new structures by composing together known facts and skills. Thanks to such learning skills, the machine will be able to combine several existing abilities to 23 create a new one, often on the ﬂy. In this way, a well-functioning intelligent machine will not need a myriad of training examples whenever it faces a slightly new request, but it could succeed given a single example of the new functionality. For example, when the Teacher asks the Learner to ﬁnd one apple and two bananas and bring them home, if the Learner already understands all the individual abilities involved, it can retrieve the relevant compositional learning skill to put together a plan and execute it step by step. The Teacher may even call the new skill generated in this way prepare breakfast, and refer to it later as such. Understanding this new concept should not require any further training of the Learner, and the latter should simply store the new skill together with its label in its long-term memory. As we have seen in the previous examples, the Learner can continue extending its knowledge of words, commands and skills in a completely unsupervised way once it manages to acquire skills that allow it to compose structures in its long-term memory. It may be that discovering the basic learning skills, something we usually take for granted, is much more intricate than it seems to us. But once we will be able to build a machine which can eﬀectively construct itself based on the incoming signals –even when no explicit supervision in the form of rewards is given, as discussed above– we should be much closer to the development of intelligent machines. 4.3 Computational properties of intelligent machines Another aspect of the intelligent machine that deserves discussion is the computational model that the machine will be based on. We are convinced that such model should be unrestricted, that is, able to represent any pattern in the data. Humans can think of and talk about algorithms without obvious limitations (although, to apply them, they might need to rely on external supports, such as paper and pencil). A useful intelligent machine should be able to handle such algorithms as well. A more precise formulation of our claim in the context of the theory of computation is that the intelligent machine needs to be based on a Turing-complete computational model. That is, it has to be able to represent any algorithm in ﬁxed length, just like the Turing machine (the very fact that humans can describe Turing-complete systems shows that they are, in practical terms, Turing-complete: it is irrelevant, for our purposes, whether human online processing capabilities are strictly Turing-complete– what matters is that their reasoning skills, at least when aided by external supports, are). Note that there are many Turing-complete computational systems, and Turing machines in particular are a lot less eﬃcient than some alternatives, e.g., Random Access Machines. Thus, we are not interested in building the intelligent machine around the concept of the Turing machine; we just aim to use a computational model that does not have obvious limitations in ability to represent patterns. A system that is weaker than Turing-complete cannot represent certain patterns in the data eﬃciently, which in turn means it cannot truly learn them in a general sense. However, it is possible to memorize such complex patterns up to some ﬁnite level 24 of complexity. Thus, even a computationally restricted system may appear to work as intended up to some level of accuracy, given that a suﬃcient number of training examples is provided. For example, we may consider a sequence repetition problem. The machine is supposed to remember a sequence of symbols and reproduce it later. Further, let’s assume the machine is based on a model with the representational power of ﬁnite state machines. Such system is not capable to represent the concept of storing and reproducing a sequence. However, it may appear to do so if we design our experiment imperfectly. Assume there is a signiﬁcant overlap between what the machine sees as training data, and the test data we use to evaluate performance of the machine. A trivial machine that can function as a look-up table may appear to work, simply by storing and recalling the training examples. With an inﬁnite number of training examples, a look-up-table-based machine would appear to learn any regularity. It will work indistinguishably from a machine that can truly represent the concept of repetition; however, it will need to have inﬁnite size. Clearly, such memorization- based system will not perform well in our setting, as we aim to test the Learner’s ability to generalize from a few examples. Since there are many Turing-complete computational systems, one may wonder which one should be preferred as the basis for machine intelligence. We cannot answer this question yet, however we hypothesize that the most natural choice would be a system that performs computation in a parallel way, using elementary units that can grow in number based on the task at hand. The growing property is necessary to support the long-term memory, if we assume that the basic units themselves are ﬁnite. An example of an existing computational system with many of the desired properties is the cellular automaton of Von Neumann et al. (1966). We might also be inspired by string rewriting systems, for example some versions of the L-systems (Prusinkiewicz and Lindenmayer, 2012). An apparent alternative would be to use a non-growing model with immensely large capacity. There is however an important diﬀerence. In a growing model, the new cells can be connected to those that spawned them, so that the model is naturally able to develop a meaningful topological structure based on functional connectivity. We conjecture that such structure would in itself contribute to learning in a crucial way. On the other hand, it is not clear if such topological structure can arise in a large-capacity unstructured model. Interestingly, some of the more eﬀective machine- learning models available today, such as recurrent and convolutional neural networks, are characterized by (manually constrained) network topologies that are well-suited to the domains they are applied to. 5 Related ideas We owe, of course, a large debt to the seminal work of Turing (1950). Note that, while Turing’s paper is most often cited for the “imitation game”, there are other very 25 interesting ideas in it, worthy of more attention from curious readers, especially in the last section on learning machines. Turing thought that a good way to construct a machine capable of passing his famous test would be to develop a child machine, and teach it further skills through various communication channels. These would include sparse rewards shaping the behavior of the child machine, and other information-rich channels such as language input from a teacher and sensory information. We share Turing’s goal of developing a child machine capable of independent com- munication through natural language, and we also stress the importance of sparse rewards. The main distinction between his and our vision is that Turing assumed that the child machine would be largely programmed (he gives an estimate of sixty programmers working on it for ﬁfty years). We rather think of starting with a machine only endowed with very elementary skills, and focus on the capability to learn as the fundamental ability that needs to be developed. This further assumes educating the machine at ﬁrst in a simulated environment where an artiﬁcial teacher will train it, as we outlined in our roadmap. We also diverge with respect to the imitation game, since the purpose of our intelligent machine is not to fool human judges into believing it is actually a real person. Instead, we aim to develop a machine that can perform a sim- ilar set of tasks to those a human can do by using a computer, an Internet connection and the ability to communicate. There has been a recent revival of interest in tasks measuring computational intel- ligence, spurred by the empirical advances of powerful machine-learning architectures such as multi-layered neural networks (LeCun et al., 2015), and by the patent inade- quacy of the classic version of Turing test (Wikipedia, 2015c). For example, Levesque et al. (2012) propose to test systems on their ability to resolve coreferential ambiguities (The trophy would not ﬁt in the brown suitcase because it was too big...What was too big?). Geman et al. (2015) propose a “visual” Turing test in which a computational system is asked to answer a set of increasingly speciﬁc questions about objects, at- tributes and relations in a picture (Is there a person in the blue region? Is the person carrying something? Is the person interacting with any other object?). Similar initia- tives diﬀer from ours in that they focus on a speciﬁc set of skills (coreference, image parsing) rather than testing if an agent can learn new skills. Moreover, these are tra- ditional evaluation benchmarks, unlike the hybrid learning/evaluation ecosystem we are proposing. The idea of developing an AI living in a controlled synthetic environment and in- teracting with other agents through natural language is quite old. The Blocks World of Winograd (1971) is probably the most important example of early research in this vein. The approach was later abandoned, when it became clear that the agents devel- oped within this framework did not scale up to real-world challenges (see, e.g., Morelli et al., 1992). The knowledge encoded in the systems tested by these early simulations was manually programmed by their creators, since they had very limited learning ca- pabilities. Consequently, scaling up to the real world implied manual coding of all the knowledge necessary to cope with it, and this proved infeasible. Our simulation 26 is instead aiming at systems that encode very little prior knowledge and have strong capabilities to learn from data. Importantly, our plan is not to try to manually pro- gram all possible scripts our system might encounter later, as in some of the classic AI systems. We plan to program only the initial environment, in order to kickstart the machine’s ability to learn and adapt to diﬀerent problems and scenarios. After the simulated environment is mastered, scaling up the functionality of our Learner will not require further manual work on scripting new situations, but will rather focus on integrating real world inputs, such as those coming from human users. The toy world itself is already designed to feature novel tasks of increasing complexity, explicitly testing the abilities of systems to autonomously scale up. Still, we should not underestimate the drawbacks of synthetic simulations. The tasks in our environment might directly address some challenging points in the de- velopment of AI, such as learning with very weak supervision, being able to form a structured long-term memory, and the ability of the child machine to grow in size and complexity when encountering new problems. However, simulating the real world can only bring us so far, and we might end up overestimating the importance of some arbi- trary phenomena at the expense of others, that might turn out to be more common in natural settings. It may be important to bring reality into the picture relatively soon. Our toy world should let the intelligent machine develop to the point at which it is able to learn from and cooperate with actual humans. Interaction with real-life humans will then naturally lead the machine to deal with real-world problems. The issue of when exactly a machine trained in our controlled synthetic environment is ready to go out in the human world is open, and it should be explored empirically. However, at the same time, we believe that having the machine interact with humans before it can deal with basic problems in the controlled environment would be pointless, and possibly even strongly misleading. Our intelligent machine shares some of its desired functionalities with the current generation of automated personal assistants such as Apple’s Siri ad Microsoft’s Cor- tana. However, these are heavily engineered systems that aim to provide a natural language interface for human users to perform a varied but ﬁxed set of tasks (similar considerations also apply to artiﬁcial human companions and digital pets such as Tam- agotchi, see Wikipedia, 2015a). Such systems can be developed by deﬁning the most frequent use cases, choosing those that can be solved with the current technology (e.g., book an air ticket, look at the weather forecast and set the alarm clock for tomorrow’s morning), and implementing speciﬁc solutions for each such use case. Our intelligent machine is not intended to handle just a ﬁxed set of tasks. As exempliﬁed by the example in Section 3.3, the machine should be capable to learn eﬃciently how to per- form tasks such as those currently handled by personal assistants, and more, just from interaction with the human user (without a programmer or machine learning expert in the loop). Architectures for software agents, and more speciﬁcally intelligent agents, are widely studied in AI and related ﬁelds (Nwana, 1996; Russell and Norvig, 2009). We 27 cannot review this ample literature here, in order to position our proposal precisely with respect to it. We simply remark that we are not aware of other architectures that are as centered on learning and communication as ours. Interaction plays a central role in the study of multiagent systems (Shoham and Leyton-Brown, 2009). However, the emphasis in this research tradition is on how conﬂict resolution and distributed problem solving evolve in typically large groups of simple, mostly scripted agents. For example, traﬃc modeling is a classic application scenario for multiagent systems. This is very diﬀerent from our emphasis on linguistic interaction for the purposes of training a single agent that should become independently capable of very complex behaviours. Tenenbaum (2015), like us, emphasizes the need to focus on basic abilities that form the core of intelligence. However, he takes naive physics problems as the starting point, and discusses speciﬁc classes of probabilistic models, rather than proposing a general learning scenario. There are also some similarities between our proposal and the research program of Luc Steels (e.g., Steels, 2003, 2005), who lets robots evolve vocabularies and grammatical constructions through interaction in a situated environment. However, on the one hand his agents are actual robots subject to the practical hardware limitations imposed by the need to navigate a complex natural environment from the start; on the other, the focus of the simulations is narrowly on language acquisition, with no further aim to develop broadly intelligent agents. We have several points of contact with the semantic parsing literature, such as nav- igation tasks in an artiﬁcial world (MacMahon et al., 2006) and reward-based learning from natural language instructions (Chen and Mooney, 2011; Artzi and Zettlemoyer, 2013). The agents developed in this area can perform tasks, such as learning to execute instructions in natural environments by interacting with humans (Thomason et al., 2015), or improving performance on real-life video-games by consulting the instruc- tion manual (Branavan et al., 2012), that we would want our intelligent machines to also be able to carry out. However, current semantic-parsing-based systems achieve these impressive feats by exploiting architectures tuned to the speciﬁc tasks at hand, and they rely on a fair amount of hard-wired expert knowledge, in particular about language structures (although recent work is moving towards a more knowledge-lean direction, see for example Narasimhan et al., 2015, who train a neural network to play text-based adventure games using only text descriptions as input and game reward as signal). Our framework is meant to encourage the development of systems that should eventually be able to perform similar tasks, but getting there incrementally, starting with almost no prior knowledge and ﬁrst learning from their environment a set of simpler skills, and how to creatively merge them to tackle more ambitious goals. The last twenty years have witnessed several related proposals on learning to learn (Thrun and Pratt, 1997), lifelong learning (Silver et al., 2013) and continual learning (Ring, 1997). Much of this work is theoretical in nature and focuses on algorithms rather than on empirical challenges for the proposed models. Still, the general ideas being pursued are in line with our program. Ring (1997), in particular, deﬁnes a continual-learning agent whose experiences “occur sequentially, and what it learns at 28 one time step while solving one task, it can use later, perhaps to solve a completely diﬀerent task.” Ring’s desiderata for the continual learner are remarkably in line with ours. It is “an autonomous agent. It senses, takes actions, and responds to the rewards in its environment. It learns behaviors and skills while solving its tasks. It learns incrementally. There is no ﬁxed training set; learning occurs at every time step; and the skills the agent learns now can be used later. It learns hierarchically. Skills it learns now can be built upon and modiﬁed later. It is a black box. The internals of the agent need not be understood or manipulated. All of the agent’s behaviors are developed through training, not through direct manipulation. Its only interface to the world is through its senses, actions, and rewards. It has no ultimate, ﬁnal task. What the agent learns now may or may not be useful later, depending on what tasks come next.” Our program is deﬁnitely in the same spirit, with an extra emphasis on interaction. Mitchell et al. (2015) discuss NELL, the most fully realized concrete implementa- tion of a lifelong learning architecture. NELL is an agent that has been “reading the Web” for several years to extract a large knowledge base. Emphasis is on the never- ending nature of the involved tasks, on their incremental reﬁnement based on what NELL has learned, and on sharing information across tasks. In this latter respect, this project is close to multi-task learning (Ando and Zhang, 2005; Caruana, 1997; Collobert et al., 2011), that focuses on the idea of parameter sharing across tasks. It is likely that a successful learner in our framework will exploit similar strategies, but our current focus lies on deﬁning the tasks, rather than on how to pursue them. Bengio et al. (2009) propose the related idea of curriculum learning, whereby train- ing data for a single task are ordered according to a diﬃculty criterion, in the hope that this will lead to better learning. This is motivated by the observation that humans learn incrementally when developing complex skills, an idea that has also previously been studied in the context of recurrent neural network training by Elman (1993). The principle of incremental learning is also central to our proposal. However, the fundamental aspect for us is not a strict ordering of the training data for a speciﬁc task, but incrementality in the skills that the intelligent machine should develop. This sort of incrementality should in turn be boosted by designing separate tasks with a compositional structure, such that the skills acquired from the simpler tasks will help to solve the more advanced ones more eﬃciently. The idea of incremental learning, motivated by the same considerations as in the papers we just mentioned, also appears in Solomonoﬀ (2002), a work which has much earlier roots in research on program induction (Solomonoﬀ, 1964, 1997; Schmidhu- ber, 2004). Within this tradition, Schmidhuber (2015) reviews a large literature and presents some general ideas on learning that might inspire our search for novel algo- rithms. Genetic programming (Poli et al., 2008) also focuses on the reuse of previously found sub-solutions, speeding up the search procedure in this way. Our proposal is also related to that of Bottou (2014), in its vision of compositional machine learn- ing, although he only considers composition in limited domains, such as sentence and 29 image processing. We share many ideas with the reinforcement learning framework (Sutton and Barto, 1998). In reinforcement learning, the agent chooses actions in an environment in order to maximize some cumulative reward over time. Reinforcement learning is particularly popular for problems where the agent can collect information only by interacting with the environment. Given how broad this deﬁnition is, our framework could be considered as a particular instance of it. Our proposal is however markedly diﬀerent from standard reinforcement learning work (Kaelbling et al., 1996) in several respects. Speciﬁcally, we emphasize language-mediated, interactive communication, we focus on incremental strategies that encourage agents to solve tasks by reusing previously learned knowledge and we aim to limit the number of trials an agent gets in order to accomplish a certain goal. Mnih et al. (2015) recently presented a single neural network architecture capable of learning a set of classic Atari games using only pixels and game scores as input (see also the related idea of “general game playing”, e.g., Genesereth et al., 2005). We pursue a similar goal of learning from a low-level input stream and reward. However, unlike these authors, we do not aim for a single architecture that can, disjointly, learn an array of separate tasks, but for one that can incrementally build on skills learned on previous tasks to perform more complex ones. Moreover, together with reward, we emphasize linguistic interaction as a fundamental mean to foster skill extension. Sukhbaatar et al. (2015) introduce a sandbox to design games with the explicit purpose to train computational agents in planning and reasoning tasks. Moreover, they stress a curriculum strategy to foster learning (making the agent progress through increasingly more diﬃcult versions of the game). Their general program is aligned with ours, and the sandbox might be useful to develop our environment. However, they do not share our emphasis on communication and interaction, and their approach to incremental learning is based on increasingly more diﬃcult versions of the same task (e.g., increasing the number of obstacles), rather than on deﬁning progressively more complex tasks, such that solving the later ones requires composing solutions to earlier ones, as we are proposing. Furthermore, the tasks currently considered within the sandbox do not seem to be challenging enough to require new learning approaches, and may be solvable with current techniques or minor modiﬁcations thereof. Mikolov (2013) originally discussed a preliminary version of the incremental task- based approach we are more fully outlying here. In a similar spirit, Weston et al. (2015) present a set of question answering tasks based on synthetically generated stories. They also want to foster non-incremental progress in AI, but their approach diﬀers from ours in several crucial aspects. Again, there is no notion of interactive, language-mediated learning, a classic train/test split is enforced, and the tasks are not designed to encourage compositional skill learning (although Weston and colleagues do emphasize that the same system should be used for all tasks). Finally, the evaluation metric is notably diﬀerent from ours - while we aim to minimize the number of trials it takes for the machine to master the tasks, their goal is to have a good performance 30 on held out data. This could be a serious drawback for works that involve artiﬁcial tasks, as in our view the goal should be to develop a machine that can learn as fast as possible, to have any hope to scale up and be able to generalize in more complex scenarios. One could think of solving sequence-manipulation problems such as those consti- tuting the basis of our learning routine with relatively small extensions of established machine learning techniques (Graves et al., 2014; Grefenstette et al., 2015; Joulin and Mikolov, 2015). As discussed in the previous section, for simple tasks that involve only a small, ﬁnite number of conﬁgurations, one could be apparently successful even just by using a look-up table storing all possible combinations of inputs and outputs. The above mentioned works, that aim to learn algorithms from data, also add a long-term memory (e.g., a set of stacks), but they use it to store the data only, not the learned algorithms. Thus, such approaches fail to generalize in environments where solutions to new tasks are composed of already learned algorithms. Similar criticism holds for approaches that try to learn certain algorithms by using an architecture with a strong prior towards their discovery, but not general enough to represent even small modiﬁcations. To give an example from our own work: a recurrent neural network augmented with a stack structure can form a simple kind of long-term memory and learn to memorize and repeat sequences in the reversed order, but not in the original one (Joulin and Mikolov, 2015). We expect a valid solution to the algorithmic learning challenge to utilize a small number of training examples, and to learn tasks that are closely related at an increasing speed, i.e., to require less and less examples to master new skills that are related to what is already known. We are not aware of any current technique addressing these issues, which were the very reason why algorithmic tasks were originally proposed by Mikolov (2013). We hope that this paper will motivate the design of the genuinely novel methods we need in order to develop intelligent machines. 6 Conclusion We deﬁned basic desiderata for an intelligent machine, stressing learning and commu- nication as its fundamental abilities. Contrary to common practice in current machine learning, where the focus is on modeling single skills in isolation, we believe that all aspects of intelligence should be holistically addressed within a single system. We proposed a simulated environment that requires the intelligent machine to acquire new facts and skills through communication. In this environment, the machine must learn to perform increasingly more ambitious tasks, being naturally induced to develop complex linguistic and reasoning abilities. We also presented some conjectures on the properties of the computational system that the intelligent machine may be based on. These include learning of algorithmic patterns from a few examples without strong supervision, and development of a long- term memory to store both data and learned skills. We tried to put this in contrast with 31 currently accepted paradigms in machine learning, to show that current methods are far from adequate, and we must strive to develop non-incrementally novel techniques. This roadmap constitutes only the beginning of a long journey towards AI, and we hope other researchers will be joining it in pursuing the goals it outlined. Acknowledgments We thank L´eon Bottou, Yann LeCun, Gabriel Synnaeve, Arthur Szlam, Nicolas Usunier, Laurens van der Maaten, Wojciech Zaremba and others from the Facebook AI Research team, as well as Gemma Boleda, Katrin Erk, Germa´n Kruszewski, Angeliki Lazaridou, Louise McNally, Hinrich Schu¨tze and Roberto Zamparelli for many stimulating discus- sions. An early version of this proposal has been discussed in several research groups since 2013 under the name Incremental learning of algorithms (Mikolov, 2013). References Ando, R. and Zhang, T. (2005). A framework for learning predictive structures from multiple tasks and unlabeled data. Journal of Machine Learning Research, 5:1817– 1853. Artzi, Y. and Zettlemoyer, L. (2013). Weakly supervised learning of semantic parsers for mapping instructions to actions. Transactions of the Association for Computa- tional Linguistics, 1(1):49–62. Bengio, Y., Louradour, J., Collobert, R., and Weston, J. (2009). Curriculum learning. In Proceedings of ICML, pages 41–48, Montreal, Canada. Boleda, G. and Erk, K. (2015). Distributional semantic features as semantic primitives–or not. In Proceedings of the AAAI Spring Symposium on Knowledge Representation and Reasoning: Integrating Symbolic and Neural Approaches, pages 2–5, Stanford, CA. Bottou, L. (2014). From machine learning to machine reasoning: an essay. Machine Learning, 94:133–149. Branavan, S., Silver, D., and Barzilay, R. (2012). Learning to win by reading manuals in a Monte-Carlo framework. Journal of Artiﬁcial Intelligence Research, 43:661–704. Caruana, R. (1997). Multitask learning. Machine Learning, 28:41–75. Chen, D. and Mooney, R. (2011). Learning to interpret natural language naviga- tion instructions from observations. In Proceedings of AAAI, pages 859–865, San Francisco, CA. 32 Collobert, R., Weston, J., Bottou, L., Karlen, M., Kavukcuoglu, K., and Kuksa, P. (2011). Natural language processing (almost) from scratch. Journal of Machine Learning Research, 12:2493–2537. Elman, J. (1993). Learning and development in neural networks: the importance of starting small. Cognition, 48:71–99. Erk, K. (2012). Vector space models of word meaning and phrase meaning: A survey. Language and Linguistics Compass, 6(10):635–653. Fodor, J. (1975). The Language of Thought. Crowell Press, New York. Geman, D., Geman, S., Hallonquist, N., and Younes, L. (2015). Visual Turing test for computer vision systems. Proceedings of the National Academy of Sciences, 112(12):3618–3623. Genesereth, M., Love, N., and Pell, B. (2005). General game playing: Overview of the AAAI competition. AI Magazine, 26(2):62–72. Graves, A., Wayne, G., and Danihelka, I. (2014). Neural turing machines. http: //arxiv.org/abs/1410.5401. Grefenstette, E., Hermann, K., Suleyman, M., and Blunsom, P. (2015). Learning to transduce with unbounded memory. In Proceedings of NIPS, Montreal, Canada. In press. Haugeland, J. (1985). Artiﬁcial Intelligence: The Very Idea. MIT Press, Cambridge, MA. Joulin, A. and Mikolov, T. (2015). Inferring algorithmic patterns with stack- augmented recurrent nets. In Proceedings of NIPS, Montreal, Canada. In press. Kaelbling, L. P., Littman, M. L., and Moore, A. W. (1996). Reinforcement learning: A survey. Journal of artiﬁcial intelligence research, pages 237–285. LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521:436–444. Levesque, H. J., Davis, E., and Morgenstern, L. (2012). The Winograd schema chal- lenge. In Proceedings of KR, pages 362–372, Rome, Italy. Louwerse, M. (2011). Symbol interdependency in symbolic and embodied cognition. Topics in Cognitive Science, 3:273–302. MacMahon, M., Stankiewicz, B., and Kuipers, B. (2006). Walk the talk: Connecting language, knowledge, and action in route instructions. In Proceedings of AAAI, pages 1475–1482, Boston, MA. Mikolov, T. (2013). Incremental learning of algorithms. Unpublished manuscript. 33 Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Eﬃcient estimation of word representations in vector space. http://arxiv.org/abs/1301.3781/. Mitchell, T., Cohen, W., Hruschka, E., Talukdar, P., Betteridge, J., Carlson, A., Mishra, B., Gardner, M., Kisiel, B., Krishnamurthy, J., Lao, N., Mazaitis, K., Mohamed, T., Nakashole, N., Platanios, E., Ritter, A., Samadi, M., Settles, B., Wang, R., Wijaya, D., Gupta, A., Chen, X., Saparov, A., Greaves, M., and Welling, J. (2015). Never-ending learning. In Proceedings of AAAI, pages 2302–2310, Austin, TX. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A., Veness, J., Bellemare, M., Graves, A., Riedmiller, M., Fidjeland, A., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. (2015). Human-level control through deep reinforcement learning. Nature, 518:529– 533. Morelli, R., Brown, M., Anselmi, D., Haberlandt, K., and Lloyd, D., editors (1992). Minds, Brains, and Computers: Perspectives in Cognitive Science and Artiﬁcial Intelligence. Ablex, Norwood, NJ. Narasimhan, K., Kulkarni, T., and Barzilay, R. (2015). Language understanding for text-based games using deep reinforcement learning. In Proceedings of EMNLP, pages 1–11, Lisbon, Portugal. Nwana, H. (1996). Software agents: An overview. Knowledge Engineering Review, 11(2):1–40. Poli, R., Langdon, W., McPhee, N., and Koza, J. (2008). A ﬁeld guide to genetic programming. http://www.gp-field-guide.org.uk. Prusinkiewicz, P. and Lindenmayer, A. (2012). The algorithmic beauty of plants. Springer Science & Business Media. Ring, M. (1997). CHILD: A ﬁrst step towards continual learning. Machine Learning, 28:77–104. Russell, S. and Norvig, P. (2009). Artiﬁcial Intelligence: A Modern Approach, 3d ed. Pearson Education, New York. Schmidhuber, J. (2004). Optimal ordered problem solver. Machine Learning, 54(3):211–254. Schmidhuber, J. (2015). On learning to think: Algorithmic information theory for novel combinations of reinforcement learning controllers and recurrent neural world models. http://arxiv.org/abs/1511.09249. 34 Shoham, Y. and Leyton-Brown, K. (2009). Multiagent Systems. Cambridge University Press, Cambridge. Silver, D., Yang, Q., and Li, L. (2013). Lifelong machine learning systems: Beyond learning algorithms. In Proceedings of the AAAI Spring Symposium on Lifelong Machine Learning, pages 49–55, Stanford, CA. Solomonoﬀ, R. J. (1964). A formal theory of inductive inference. Part I. Information and control, 7(1):1–22. Solomonoﬀ, R. J. (1997). The discovery of algorithmic probability. Journal of Com- puter and System Sciences, 55(1):73–88. Solomonoﬀ, R. J. (2002). Progress in incremental machine learning. In NIPS Workshop on Universal Learning Algorithms and Optimal Search, Whistler, BC. Citeseer. Steels, L. (2003). Social language learning. In Tokoro, M. and Steels, L., editors, The Future of Learning, pages 133–162. IOS, Amsterdam. Steels, L. (2005). What triggers the emergence of grammar? In Proceedings of EELC, pages 143–150, Hatﬁeld, UK. Sukhbaatar, S., Szlam, A., Synnaeve, G., Chintala, S., and Fergus, R. (2015). Maze- Base: a sandbox for learning from games. http://arxiv.org/abs/1511.07401. Sutton, R. and Barto, A. (1998). Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA. Tenenbaum, J. (2015). Cognitive foundations for knowledge representation in AI. Presented at the AAAI Spring Symposium on Knowledge Representation and Rea- soning. Thomason, J., Zhang, S., Mooney, R., and Stone, P. (2015). Learning to interpret natural language commands through human-robot dialog. In Proceedings IJCAI, pages 1923–1929, Buenos Aires, Argentina. Thrun, S. and Pratt, L., editors (1997). Learning to Learn. Kluwer, Dordrecht. Turing, A. (1950). Computing machinery and intelligence. Mind, 59:433–460. Turney, P. and Pantel, P. (2010). From frequency to meaning: Vector space models of semantics. Journal of Artiﬁcial Intelligence Research, 37:141–188. Von Neumann, J., Burks, A. W., et al. (1966). Theory of self-reproducing automata. IEEE Transactions on Neural Networks, 5(1):3–14. 35 Weston, J., Bordes, A., Chopra, S., and Mikolov, T. (2015). Towards AI-complete question answering: A set of prerequisite toy tasks. http://arxiv.org/abs/1502. 05698. Wikipedia (2015a). Artiﬁcial human companion. https://en.wikipedia.org/ w/index.php?title=Artificial_human_companion&oldid=685507143. Accessed 15-October-2015. Wikipedia (2015b). Interactive ﬁction. https://en.wikipedia.org/w/index.php? title=Interactive_fiction&oldid=693926750. Accessed 19-December-2015. Wikipedia (2015c). Turing test. https://en.wikipedia.org/w/index.php?title= Turing_test&oldid=673582926. Accessed 30-July-2015. Winograd, T. (1971). Procedures as a representation for data in a computer pro- gram for understanding natural language. Technical Report AI 235, Massachusetts Institute of Technology. 36 an-level control through deep reinforcement learning. Nature, 518:529– 533. Morelli, R., Brown, M., Anselmi, D., Haberlandt, K., and Lloyd, D., editors (1992). Minds, Brains, and Computers: Perspectives in Cognitive S