Showing papers by "Matthew Botvinick published in 2020"

A distributional code for value in dopamine-based reinforcement learning

Abstract: Since its introduction, the reward prediction error theory of dopamine has explained a wealth of empirical phenomena, providing a unifying framework for understanding the representation of reward and value in the brain1–3. According to the now canonical theory, reward predictions are represented as a single scalar quantity, which supports learning about the expectation, or mean, of stochastic outcomes. Here we propose an account of dopamine-based reinforcement learning inspired by recent artificial intelligence research on distributional reinforcement learning4–6. We hypothesized that the brain represents possible future rewards not as a single mean, but instead as a probability distribution, effectively representing multiple future outcomes simultaneously and in parallel. This idea implies a set of empirical predictions, which we tested using single-unit recordings from mouse ventral tegmental area. Our findings provide strong evidence for a neural realization of distributional reinforcement learning. Analyses of single-cell recordings from mouse ventral tegmental area are consistent with a model of reinforcement learning in which the brain represents possible future rewards not as a single mean of stochastic outcomes, as in the canonical model, but instead as a probability distribution.

V-MPO: On-Policy Maximum a Posteriori Policy Optimization for Discrete and Continuous Control

Abstract: Some of the most successful applications of deep reinforcement learning to challenging domains in discrete and continuous control have used policy gradient methods in the on-policy setting. However, policy gradients can suffer from large variance that may limit performance, and in practice require carefully tuned entropy regularization to prevent policy collapse. As an alternative to policy gradient algorithms, we introduce V-MPO, an on-policy adaptation of Maximum a Posteriori Policy Optimization (MPO) that performs policy iteration based on a learned state-value function. We show that V-MPO surpasses previously reported scores for both the Atari-57 and DMLab-30 benchmark suites in the multi-task setting, and does so reliably without importance weighting, entropy regularization, or population-based tuning of hyperparameters. On individual DMLab and Atari levels, the proposed algorithm can achieve scores that are substantially higher than has previously been reported. V-MPO is also applicable to problems with high-dimensional, continuous action spaces, which we demonstrate in the context of learning to control simulated humanoids with 22 degrees of freedom from full state observations and 56 degrees of freedom from pixel observations, as well as example OpenAI Gym tasks where V-MPO achieves substantially higher asymptotic scores than previously reported.

Environmental drivers of systematicity and generalization in a situated agent.

Abstract: The question of whether deep neural networks are good at generalising beyond their immediate training experience is of critical importance for learning-based approaches to AI. Here, we consider tests of out-of-sample generalisation that require an agent to respond to never-seen-before instructions by manipulating and positioning objects in a 3D Unity simulated room. We first describe a comparatively generic agent architecture that exhibits strong performance on these tests. We then identify three aspects of the training regime and environment that make a significant difference to its performance: (a) the number of object/word experiences in the training set; (b) the visual invariances afforded by the agent's perspective, or frame of reference; and (c) the variety of visual input inherent in the perceptual aspect of the agent's perception. Our findings indicate that the degree of generalisation that networks exhibit can depend critically on particulars of the environment in which a given task is instantiated. They further suggest that the propensity for neural networks to generalise in systematic ways may increase if, like human children, those networks have access to many frames of richly varying, multi-modal observations as they learn.

Stabilizing Transformers for Reinforcement Learning

Abstract: Owing to their ability to both effectively integrate information over long time horizons and scale to massive amounts of data, self-attention architectures have recently shown breakthrough success in natural language processing (NLP), achieving state-of-the-art results in domains such as language modeling and machine translation. Harnessing the transformer's ability to process long time horizons of information could provide a similar performance boost in partially observable reinforcement learning (RL) domains, but the large-scale transformers used in NLP have yet to be successfully applied to the RL setting. In this work we demonstrate that the standard transformer architecture is difficult to optimize, which was previously observed in the supervised learning setting but becomes especially pronounced with RL objectives. We propose architectural modifications that substantially improve the stability and learning speed of the original Transformer and XL variant. The proposed architecture, the Gated Transformer-XL (GTrXL), surpasses LSTMs on challenging memory environments and achieves state-of-the-art results on the multi-task DMLab-30 benchmark suite, exceeding the performance of an external memory architecture. We show that the GTrXL, trained using the same losses, has stability and performance that consistently matches or exceeds a competitive LSTM baseline, including on more reactive tasks where memory is less critical. GTrXL offers an easy-to-train, simple-to-implement but substantially more expressive architectural alternative to the standard multi-layer LSTM ubiquitously used for RL agents in partially observable environments.

Unsupervised deep learning identifies semantic disentanglement in single inferotemporal neurons

Abstract: Deep supervised neural networks trained to classify objects have emerged as popular models of computation in the primate ventral stream These models represent information with a high-dimensional distributed population code, implying that inferotemporal (IT) responses are also too complex to interpret at the single-neuron level We challenge this view by modelling neural responses to faces in the macaque IT with a deep unsupervised generative model, beta-VAE Unlike deep classifiers, beta-VAE "disentangles" sensory data into interpretable latent factors, such as gender or hair length We found a remarkable correspondence between the generative factors discovered by the model and those coded by single IT neurons Moreover, we were able to reconstruct face images using the signals from just a handful of cells This suggests that the ventral visual stream may be optimising the disentangling objective, producing a neural code that is low-dimensional and semantically interpretable at the single-unit level

Deep Reinforcement Learning and its Neuroscientific Implications

Abstract: The emergence of powerful artificial intelligence is defining new research directions in neuroscience. To date, this research has focused largely on deep neural networks trained using supervised learning, in tasks such as image classification. However, there is another area of recent AI work which has so far received less attention from neuroscientists, but which may have profound neuroscientific implications: deep reinforcement learning. Deep RL offers a comprehensive framework for studying the interplay among learning, representation and decision-making, offering to the brain sciences a new set of research tools and a wide range of novel hypotheses. In the present review, we provide a high-level introduction to deep RL, discuss some of its initial applications to neuroscience, and survey its wider implications for research on brain and behavior, concluding with a list of opportunities for next-stage research.

Object-based attention for spatio-temporal reasoning: Outperforming neuro-symbolic models with flexible distributed architectures.

Abstract: Neural networks have achieved success in a wide array of perceptual tasks, but it is often stated that they are incapable of solving tasks that require higher-level reasoning. Two new task domains, CLEVRER and CATER, have recently been developed to focus on reasoning, as opposed to perception, in the context of spatio-temporal interactions between objects. Initial experiments on these domains found that neuro-symbolic approaches, which couple a logic engine and language parser with a neural perceptual front-end, substantially outperform fully-learned distributed networks, a finding that was taken to support the above thesis. Here, we show on the contrary that a fully-learned neural network with the right inductive biases can perform substantially better than all previous neural-symbolic models on both of these tasks, particularly on questions that most emphasize reasoning over perception. Our model makes critical use of both self-attention and learned "soft" object-centric representations, as well as BERT-style semi-supervised predictive losses. These flexible biases allow our model to surpass the previous neuro-symbolic state-of-the-art using less than 60% of available labelled data. Together, these results refute the neuro-symbolic thesis laid out by previous work involving these datasets, and they provide evidence that neural networks can indeed learn to reason effectively about the causal, dynamic structure of physical events.

Emergent Systematic Generalization In a Situated Agent

Abstract: The question of whether deep neural networks are good at generalising beyond their immediate training experience is of critical importance for learning-based approaches to AI. Here, we demonstrate strong emergent systematic generalisation in a neural network agent and isolate the factors that support this ability. In environments ranging from a grid-world to a rich interactive 3D Unity room, we show that an agent can correctly exploit the compositional nature of a symbolic language to interpret never-seen-before instructions. We observe this capacity not only when instructions refer to object properties (colors and shapes) but also verb-like motor skills (lifting and putting) and abstract modifying operations (negation). We identify three factors that can contribute to this facility for systematic generalisation: (a) the number of object/word experiences in the training set; (b) the invariances afforded by a first-person, egocentric perspective; and (c) the variety of visual input experienced by an agent that perceives the world actively over time. Thus, while neural nets trained in idealised or reduced situations may fail to exhibit a compositional or systematic understanding of their experience, this competence can readily emerge when, like human learners, they have access to many examples of richly varying, multi-modal observations as they learn.

The Variational Bandwidth Bottleneck: Stochastic Evaluation on an Information Budget

Abstract: In many applications, it is desirable to extract only the relevant information from complex input data, which involves making a decision about which input features are relevant. The information bottleneck method formalizes this as an information-theoretic optimization problem by maintaining an optimal tradeoff between compression (throwing away irrelevant input information), and predicting the target. In many problem settings, including the reinforcement learning problems we consider in this work, we might prefer to compress only part of the input. This is typically the case when we have a standard conditioning input, such as a state observation, and a ``privileged'' input, which might correspond to the goal of a task, the output of a costly planning algorithm, or communication with another agent. In such cases, we might prefer to compress the privileged input, either to achieve better generalization (e.g., with respect to goals) or to minimize access to costly information (e.g., in the case of communication). Practical implementations of the information bottleneck based on variational inference require access to the privileged input in order to compute the bottleneck variable, so although they perform compression, this compression operation itself needs unrestricted, lossless access. In this work, we propose the variational bandwidth bottleneck, which decides for each example on the estimated value of the privileged information before seeing it, i.e., only based on the standard input, and then accordingly chooses stochastically, whether to access the privileged input or not. We formulate a tractable approximation to this framework and demonstrate in a series of reinforcement learning experiments that it can improve generalization and reduce access to computationally costly information.

MEMO: A Deep Network for Flexible Combination of Episodic Memories

Abstract: Recent research developing neural network architectures with external memory have often used the benchmark bAbI question and answering dataset which provides a challenging number of tasks requiring reasoning. Here we employed a classic associative inference task from the human neuroscience literature in order to more carefully probe the reasoning capacity of existing memory-augmented architectures. This task is thought to capture the essence of reasoning -- the appreciation of distant relationships among elements distributed across multiple facts or memories. Surprisingly, we found that current architectures struggle to reason over long distance associations. Similar results were obtained on a more complex task involving finding the shortest path between nodes in a path. We therefore developed a novel architecture, MEMO, endowed with the capacity to reason over longer distances. This was accomplished with the addition of two novel components. First, it introduces a separation between memories/facts stored in external memory and the items that comprise these facts in external memory. Second, it makes use of an adaptive retrieval mechanism, allowing a variable number of ‘memory hops’ before the answer is produced. MEMO is capable of solving our novel reasoning tasks, as well as all 20 tasks in bAbI.

Rapid Task-Solving in Novel Environments

Abstract: We propose the challenge of rapid task-solving in novel environments (RTS), wherein an agent must solve a series of tasks as rapidly as possible in an unfamiliar environment. An effective RTS agent must balance between exploring the unfamiliar environment and solving its current task, all while building a model of the new environment over which it can plan when faced with later tasks. While modern deep RL agents exhibit some of these abilities in isolation, none are suitable for the full RTS challenge. To enable progress toward RTS, we introduce two challenge domains: (1) a minimal RTS challenge called the Memory&Planning Game and (2) One-Shot StreetLearn Navigation, which introduces scale and complexity from real-world data. We demonstrate that state-of-the-art deep RL agents fail at RTS in both domains, and that this failure is due to an inability to plan over gathered knowledge. We develop Episodic Planning Networks (EPNs) and show that deep-RL agents with EPNs excel at RTS, outperforming the nearest baseline by factors of 2-3 and learning to navigate held-out StreetLearn maps within a single episode. We show that EPNs learn to execute a value iteration-like planning algorithm and that they generalize to situations beyond their training experience. algorithm and that they generalize to situations beyond their training experience.

MEMO: A Deep Network for Flexible Combination of Episodic Memories

Abstract: Recent research developing neural network architectures with external memory have often used the benchmark bAbI question and answering dataset which provides a challenging number of tasks requiring reasoning. Here we employed a classic associative inference task from the memory-based reasoning neuroscience literature in order to more carefully probe the reasoning capacity of existing memory-augmented architectures. This task is thought to capture the essence of reasoning -- the appreciation of distant relationships among elements distributed across multiple facts or memories. Surprisingly, we found that current architectures struggle to reason over long distance associations. Similar results were obtained on a more complex task involving finding the shortest path between nodes in a path. We therefore developed MEMO, an architecture endowed with the capacity to reason over longer distances. This was accomplished with the addition of two novel components. First, it introduces a separation between memories (facts) stored in external memory and the items that comprise these facts in external memory. Second, it makes use of an adaptive retrieval mechanism, allowing a variable number of "memory hops" before the answer is produced. MEMO is capable of solving our novel reasoning tasks, as well as match state of the art results in bAbI.

Human dorsal anterior cingulate neurons signal conflict by amplifying task-relevant information

Abstract: SUMMARY Hemodynamic activity in dorsal anterior cingulate cortex (dACC) correlates with conflict, suggesting it contributes to conflict processing. This correlation could be explained by multiple neural processes that can be disambiguated by population firing rates patterns. We used targeted dimensionality reduction to characterize activity of populations of single dACC neurons as humans performed a task that manipulates two forms of conflict. Although conflict enhanced firing rates, this enhancement did not come from a discrete population of domain-general conflict-encoding neurons, nor from a distinct conflict-encoding response axis. Nor was it the epiphenomenal consequence of simultaneous coactivation of action plans. Instead, conflict amplified the task-relevant information encoded across the neuronal population. Effects of conflict were weaker and more heterogeneous in the dorsolateral prefrontal cortex (dlPFC), suggesting that dACC’s role in conflict processing may be somewhat specialized. Overall, these results support the theory that conflict biases competition between sensorimotor transformation processes occurring in dACC.

Model-free conventions in multi-agent reinforcement learning with heterogeneous preferences.

Abstract: Game theoretic views of convention generally rest on notions of common knowledge and hyper-rational models of individual behavior. However, decades of work in behavioral economics have questioned the validity of both foundations. Meanwhile, computational neuroscience has contributed a modernized 'dual process' account of decision-making where model-free (MF) reinforcement learning trades off with model-based (MB) reinforcement learning. The former captures habitual and procedural learning while the latter captures choices taken via explicit planning and deduction. Some conventions (e.g. international treaties) are likely supported by cognition that resonates with the game theoretic and MB accounts. However, convention formation may also occur via MF mechanisms like habit learning; though this possibility has been understudied. Here, we demonstrate that complex, large-scale conventions can emerge from MF learning mechanisms. This suggests that some conventions may be supported by habit-like cognition rather than explicit reasoning. We apply MF multi-agent reinforcement learning to a temporo-spatially extended game with incomplete information. In this game, large parts of the state space are reachable only by collective action. However, heterogeneity of tastes makes such coordinated action difficult: multiple equilibria are desirable for all players, but subgroups prefer a particular equilibrium over all others. This creates a coordination problem that can be solved by establishing a convention. We investigate start-up and free rider subproblems as well as the effects of group size, intensity of intrinsic preference, and salience on the emergence dynamics of coordination conventions. Results of our simulations show agents establish and switch between conventions, even working against their own preferred outcome when doing so is necessary for effective coordination.

Misdirected vigor: Differentiating the control of value from the value of control

Abstract: Theories of dACC function have to contend with an increasingly long and diverse list of signals that have been tied to this region. To account for this apparent heterogeneity, we recently proposed a theory of dACC function that embraces that heterogeneity and offers a unifying function focused on the evaluation, motivation and allocation of cognitive control. Our theory proposes that dACC integrates information relevant to determining the Expected Value of Control (EVC) – including signals pertaining to demands for control (i.e., how much and what kind of control is needed) and those pertaining to the potential incentives for control (i.e., what are the consequences of exerting control) – and uses this to determine how best to allocate cognitive control. We have shown that this theory accounts for variability in behavior and control states across a variety of experiments and task conditions, and that it can account for the confluence of seemingly unrelated task-related variables within this region, including those related to reward, decision conflict, and the unexpectedness of a stimulus or outcome (i.e., surprise). Vassena and colleagues (2020) recently sought to test predictions of the EVC model, and concluded that it was unable to account for patterns of task-related dACC activity that they observed. Specifically, they show that dACC activity increased with levels of conflict and surprise during a choice task, and argue that the EVC model predicts the opposite pattern of results. Here, we identify misconstruals in analysis and interpretation that led to this conclusion.

The Variational Bandwidth Bottleneck: Stochastic Evaluation on an Information Budget

Abstract: In many applications, it is desirable to extract only the relevant information from complex input data, which involves making a decision about which input features are relevant. The information bottleneck method formalizes this as an information-theoretic optimization problem by maintaining an optimal tradeoff between compression (throwing away irrelevant input information), and predicting the target. In many problem settings, including the reinforcement learning problems we consider in this work, we might prefer to compress only part of the input. This is typically the case when we have a standard conditioning input, such as a state observation, and a "privileged" input, which might correspond to the goal of a task, the output of a costly planning algorithm, or communication with another agent. In such cases, we might prefer to compress the privileged input, either to achieve better generalization (e.g., with respect to goals) or to minimize access to costly information (e.g., in the case of communication). Practical implementations of the information bottleneck based on variational inference require access to the privileged input in order to compute the bottleneck variable, so although they perform compression, this compression operation itself needs unrestricted, lossless access. In this work, we propose the variational bandwidth bottleneck, which decides for each example on the estimated value of the privileged information before seeing it, i.e., only based on the standard input, and then accordingly chooses stochastically, whether to access the privileged input or not. We formulate a tractable approximation to this framework and demonstrate in a series of reinforcement learning experiments that it can improve generalization and reduce access to computationally costly information.

Attention over learned object embeddings enables complex visual reasoning

Abstract: Neural networks have achieved success in a wide array of perceptual tasks but often fail at tasks involving both perception and higher-level reasoning. On these more challenging tasks, bespoke approaches (such as modular symbolic components, independent dynamics models or semantic parsers) targeted towards that specific type of task have typically performed better. The downside to these targeted approaches, however, is that they can be more brittle than general-purpose neural networks, requiring significant modification or even redesign according to the particular task at hand. Here, we propose a more general neural-network-based approach to dynamic visual reasoning problems that obtains state-of-the-art performance on three different domains, in each case outperforming bespoke modular approaches tailored specifically to the task. Our method relies on learned object-centric representations, self-attention and self-supervised dynamics learning, and all three elements together are required for strong performance to emerge. The success of this combination suggests that there may be no need to trade off flexibility for performance on problems involving spatio-temporal or causal-style reasoning. With the right soft biases and learning objectives in a neural network we may be able to attain the best of both worlds.