Research Scientist, Robotics
Ai2's Robotics team is hiring a research scientist to investigate foundation models for robotics and develop open methods, datasets, benchmarks, and models. The work combines large-scale simulation with real robots, spanning manipulation, planning and control, reinforcement or imitation learning, sim-to-real transfer, world models, and broadly competent robotic systems.
Towards a general model of human reward-based learning
Traditional work in the study of human reward-based learning involves designing an experimental task---often inspired by Reinforcement Learning (RL) theory---and fits a small set of computational models---often inspired by RL algorithms---to that dataset. For example, researchers often model human behavior on bandit tasks using variants of Q-learning. While this approach has been highly productive, leading to landmark discoveries such as the dopamine reward prediction error hypothesis, it also has limitations. This talk focuses on the lack of generalizability of such models: Even if they closely fit behavior on the original task, models derived from the one-task-one-model paradigm usually predict behavior on other tasks quite poorly. I argue that this lack of generalizability is a fundamental problem for the cognitive sciences: we intuitively expect our models to be robust to superficial task differences, such as variations in the number of choice options, reward probabilities, or the exact kind of non-stationarity. I will propose potential solutions to this problem along two dimensions: the behavioral dataset and the computational model. Regarding computational models, I will introduce work in which we moved beyond the limitations of hand-crafted one-off models by employing flexible, data-driven methods. These methods allowed us to compare classes of models instead of individual model instances, allowing us to cover the space of possible models more exhaustively, and innovate cognitive mechanisms very efficiently. For the behavioral dataset, we move from using single learning tasks to a comprehensive task space that encompasses most existing paradigms in the literature, while closing the gaps between them in a near-continuous fashion. Our results suggest that more general models in conjunction with broader datasets can pave the road toward increasingly general models of human reward-based learning and decision making, and a persistent departure from many aspects of RL theory. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-05-20. Recording duration: 00:51:19.
What is So Interesting About Reinforcement Learning?
This talk aims to answer these questions along four dimensions. First is history. RL was the basis of AI long before the term AI was introduced in 1956. The first machine learning (ML) systems were based on RL even before digital computers existed. Despite notable early successes of ML based on RL, RL essentially disappeared from ML until relatively recently. A second reason for renewed interest in RL is the clarification of some misunderstandings that have been prevalent in the ML community. A third, and most important, reason for this resurgence is that new, or rediscovered, algorithms and connections to well developed mathematical and engineering methods have been worked out. Finally, a fourth reason for the renewed interest in RL is its strong links to animal reward systems, in particular, to the role that dopamine plays in motivation and learning. VVTNS Sixth Season Opening Lecture. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2025-10-29. Recording duration: 00:55:21.
New methods for tracking and control of dynamic animal behavior during learning
The dynamics of learning in natural and artificial environments is a problem of great interest to both neuroscientists and artificial intelligence experts. However, standard analyses of animal training data either treat behavior as fixed, or track only coarse performance statistics (e.g., accuracy and bias), providing limited insight into the dynamic evolution of behavioral strategies over the course of learning. To overcome these limitations, we propose a dynamic psychophysical model that efficiently tracks trial-to-trial changes in behavior over the course of training. In this talk, I will describe recent work based on a dynamic logistic regression model that captures the time-varying dependencies of behavior on stimuli and other task covariates, which we applied to mouse training data from the International Brain Lab (IBL). Secondly, I will discuss efforts to infer animal learning rules from time-varying behavior in order to characterize how they adjust their policy in response to reward. Finally, I will describe recent work on adaptive optimal training, which combines ideas from reinforcement learning and adaptive experimental design to formulate methods for inferring animal learning rules from behavior, and using these rules to speed up animal training. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2025-01-15. Recording duration: 00:49:39.
Screen Savers : Protecting adolescent mental health in a digital world
In our rapidly evolving digital world, there is increasing concern about the impact of digital technologies and social media on the mental health of young people. Policymakers and the public are nervous. Psychologists are facing mounting pressures to deliver evidence that can inform policies and practices to safeguard both young people and society at large. However, research progress is slow while technological change is accelerating.My talk will reflect on this, both as a question of psychological science and metascience. Digital companies have designed highly popular environments that differ in important ways from traditional offline spaces. By revisiting the foundations of psychology (e.g. development and cognition) and considering digital changes' impact on theories and findings, we gain deeper insights into questions such as the following. (1) How do digital environments exacerbate developmental vulnerabilities that predispose young people to mental health conditions? (2) How do digital designs interact with cognitive and learning processes, formalised through computational approaches such as reinforcement learning or Bayesian modelling?However, we also need to face deeper questions about what it means to do science about new technologies and the challenge of keeping pace with technological advancements. Therefore, I discuss the concept of ‘fast science’, where, during crises, scientists might lower their standards of evidence to come to conclusions quicker. Might psychologists want to take this approach in the face of technological change and looming concerns? The talk concludes with a discussion of such strategies for 21st-century psychology research in the era of digitalization.
Decision and Behavior
This webinar addressed computational perspectives on how animals and humans make decisions, spanning normative, descriptive, and mechanistic models. Sam Gershman (Harvard) presented a capacity-limited reinforcement learning framework in which policies are compressed under an information bottleneck constraint. This approach predicts pervasive perseveration, stimulus‐independent “default” actions, and trade-offs between complexity and reward. Such policy compression reconciles observed action stochasticity and response time patterns with an optimal balance between learning capacity and performance. Jonathan Pillow (Princeton) discussed flexible descriptive models for tracking time-varying policies in animals. He introduced dynamic Generalized Linear Models (Sidetrack) and hidden Markov models (GLM-HMMs) that capture day-to-day and trial-to-trial fluctuations in choice behavior, including abrupt switches between “engaged” and “disengaged” states. These models provide new insights into how animals’ strategies evolve under learning. Finally, Kenji Doya (OIST) highlighted the importance of unifying reinforcement learning with Bayesian inference, exploring how cortical-basal ganglia networks might implement model-based and model-free strategies. He also described Japan’s Brain/MINDS 2.0 and Digital Brain initiatives, aiming to integrate multimodal data and computational principles into cohesive “digital brains.”
Maintaining Plasticity in Neural Networks
Nonstationarity presents a variety of challenges for machine learning systems. One surprising pathology which can arise in nonstationary learning problems is plasticity loss, whereby making progress on new learning objectives becomes more difficult as training progresses. Networks which are unable to adapt in response to changes in their environment experience plateaus or even declines in performance in highly non-stationary domains such as reinforcement learning, where the learner must quickly adapt to new information even after hundreds of millions of optimization steps. The loss of plasticity manifests in a cluster of related empirical phenomena which have been identified by a number of recent works, including the primacy bias, implicit under-parameterization, rank collapse, and capacity loss. While this phenomenon is widely observed, it is still not fully understood. This talk will present exciting recent results which shed light on the mechanisms driving the loss of plasticity in a variety of learning problems and survey methods to maintain network plasticity in non-stationary tasks, with a particular focus on deep reinforcement learning.
Neural mechanisms of adaptive behavior
Animals and humans rapidly adapt their behavior to dynamic environmental changes, such as predator threats or fluctuating food resources, often without immediate rewards. Existing literature posits that animals rely on internal representations of the environment, termed “beliefs”, for their decision policy. However, previous work ties belief updates to external reward signals, which does not explain adaptation in scenarios where trial-and-error approaches are inefficient or potentially perilous. In this work, we propose that the brain utilize dynamic representations that continuously infer the state of the environment, allowing it to update behavior rapidly. I will present a Bayesian theory for state inference in a partially observed Markov Decision Process with multiple interacting latent variables. Optimal behavior requires knowledge of hidden interactions between latent states. I will show that recurrent neural networks trained through reinforcement solve the task by learning the hidden interaction between latent states, and their activity encodes the dynamics of the optimal Bayesian estimators. The behavior of rodents trained on an identical task aligns with our theoretical model and neural network simulations, suggesting that the brain utilizes dynamic internal state representation and inference. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2024-01-31. Recording duration: 00:51:21.
Learning to Express Reward Prediction Error-like Dopaminergic Activity Requires Plastic Representations of Time
The dominant theoretical framework to account for reinforcement learning in the brain is temporal difference (TD) reinforcement learning. The TD framework predicts that some neuronal elements should represent the reward prediction error (RPE), which means they signal the difference between the expected future rewards and the actual rewards. The prominence of the TD theory arises from the observation that firing properties of dopaminergic neurons in the ventral tegmental area appear similar to those of RPE model-neurons in TD learning. Previous implementations of TD learning assume a fixed temporal basis for each stimulus that might eventually predict a reward. Here we show that such a fixed temporal basis is implausible and that certain predictions of TD learning are inconsistent with experiments. We propose instead an alternative theoretical framework, coined FLEX (Flexibly Learned Errors in Expected Reward). In FLEX, feature specific representations of time are learned, allowing for neural representations of stimuli to adjust their timing and relation to rewards in an online manner. In FLEX dopamine acts as an instructive signal which helps build temporal models of the environment. FLEX is a general theoretical framework that has many possible biophysical implementations. In order to show that FLEX is a feasible approach, we present a specific biophysically plausible model which implements the principles of FLEX. We show that this implementation can account for various reinforcement learning paradigms, and that its results and predictions are consistent with a preponderance of both existing and reanalyzed experimental data.
Richly structured reward predictions in dopaminergic learning circuits
Theories from reinforcement learning have been highly influential for interpreting neural activity in the biological circuits critical for animal and human learning. Central among these is the identification of phasic activity in dopamine neurons as a reward prediction error signal that drives learning in basal ganglia and prefrontal circuits. However, recent findings suggest that dopaminergic prediction error signals have access to complex, structured reward predictions and are sensitive to more properties of outcomes than learning theories with simple scalar value predictions might suggest. Here, I will present recent work in which we probed the identity-specific structure of reward prediction errors in an odor-guided choice task and found evidence for multiple predictive “threads” that segregate reward predictions, and reward prediction errors, according to the specific sensory features of anticipated outcomes. Our results point to an expanded class of neural reinforcement learning algorithms in which biological agents learn rich associative structure from their environment and leverage it to build reward predictions that include information about the specific, and perhaps idiosyncratic, features of available outcomes, using these to guide behavior in even quite simple reward learning tasks.
Off-policy learning in the basal ganglia
I will discuss work with Jack Lindsey modeling reinforcement learning for action selection in the basal ganglia. I will argue that the presence of multiple brain regions, in addition to the basal ganglia, that contribute to motor control motivates the need for an off-policy basal ganglia learning algorithm. I will then describe a biological implementation of such an algorithm that predicts tuning of dopamine neurons to a quantity we call "action surprise," in addition to reward prediction error. In the same model, an implementation of learning from a motor efference copy also predicts a novel solution to the problem of multiplexing feedforward and efference-related striatal activity. The solution exploits the difference between D1 and D2-expressing medium spiny neurons and leads to predictions about striatal dynamics.
Mapping learning and decision-making algorithms onto brain circuitry
In the first half of my talk, I will discuss our recent work on the midbrain dopamine system. The hypothesis that midbrain dopamine neurons broadcast an error signal for the prediction of reward is among the great successes of computational neuroscience. However, our recent results contradict a core aspect of this theory: that the neurons uniformly convey a scalar, global signal. I will review this work, as well as our new efforts to update models of the neural basis of reinforcement learning with our data. In the second half of my talk, I will discuss our recent findings of state-dependent decision-making mechanisms in the striatum.
Learning in/about/from the basal ganglia
The basal ganglia are a collection of brain areas that are connected by a variety of synaptic pathways and are a site of significant reward-related dopamine release. These properties suggest a possible role for the basal ganglia in action selection, guided by reinforcement learning. In this talk, I will discuss a framework for how this function might be performed and computational results using an upward mapping to identify putative low-dimensional control ensembles that may be involved in tuning decision policy. I will also present some recent experimental results and theory – related to effects of extracellular ion dynamics -- that run counter to the classical view of basal ganglia pathways and suggest a new interpretation of certain aspects of this framework. For those not so interested in the basal ganglia, I hope that the upward mapping approach and impact of extracellular ion dynamics will nonetheless be of interest!
Dissecting the role of accumbal D1 and D2 medium spiny neurons in information encoding
Nearly all motivated behaviors require the ability to associate outcomes with specific actions and make adaptive decisions about future behavior. The nucleus accumbens (NAc) is integrally involved in these processes. The NAc is a heterogeneous population primarily composed of D1 and D2 medium spiny projection (MSN) neurons that are thought to have opposed roles in behavior, with D1 MSNs promoting reward and D2 MSNs promoting aversion. Here we examined what types of information are encoded by the D1 and D2 MSNs using optogenetics, fiber photometry, and cellular resolution calcium imaging. First, we showed that mice responded for optical self-stimulation of both cell types, suggesting D2-MSN activation is not inherently aversive. Next, we recorded population and single cell activity patterns of D1 and D2 MSNs during reinforcement as well as Pavlovian learning paradigms that allow dissociation of stimulus value, outcome, cue learning, and action. We demonstrated that D1 MSNs respond to the presence and intensity of unconditioned stimuli – regardless of value. Conversely, D2 MSNs responded to the prediction of these outcomes during specific cues. Overall, these results provide foundational evidence for the discrete aspects of information that are encoded within the NAc D1 and D2 MSN populations. These results will significantly enhance our understanding of the involvement of the NAc MSNs in learning and memory as well as how these neurons contribute to the development and maintenance of substance use disorders.
NaV Long-term Inactivation Regulates Adaptation in Place Cells and Depolarization Block in Dopamine Neurons
In behaving rodents, CA1 pyramidal neurons receive spatially-tuned depolarizing synaptic input while traversing a specific location within an environment called its place. Midbrain dopamine neurons participate in reinforcement learning, and bursts of action potentials riding a depolarizing wave of synaptic input signal rewards and reward expectation. Interestingly, slice electrophysiology in vitro shows that both types of cells exhibit a pronounced reduction in firing rate (adaptation) and even cessation of firing during sustained depolarization. We included a five state Markov model of NaV1.6 (for CA1) and NaV1.2 (for dopamine neurons) respectively, in computational models of these two types of neurons. Our simulations suggest that long-term inactivation of this channel is responsible for the adaptation in CA1 pyramidal neurons, in response to triangular depolarizing current ramps. We also show that the differential contribution of slow inactivation in two subpopulations of midbrain dopamine neurons can account for their different dynamic ranges, as assessed by their responses to similar depolarizing ramps. These results suggest long-term inactivation of the sodium channel is a general mechanism for adaptation.
Input and target-selective plasticity in sensory neocortex during learning
Behavioral experience shapes neural circuits, adding and subtracting connections between neurons that will ultimately control sensation and perception. We are using natural sensory experience to uncover basic principles of information processing in the cerebral cortex, with a focus on how sensory learning can selectively alter synaptic strength. I will discuss recent findings that differentiate reinforcement learning from sensory experience, showing rapid and selective plasticity of thalamic and inhibitory synapses within primary sensory cortex.
Why would we need Cognitive Science to develop better Collaborative Robots and AI Systems?
While classical industrial robots are mostly designed for repetitive tasks, assistive robots will be challenged by a variety of different tasks in close contact with humans. Hereby, learning through the direct interaction with humans provides a potentially powerful tool for an assistive robot to acquire new skills and to incorporate prior human knowledge during the exploration of novel tasks. Moreover, an intuitive interactive teaching process may allow non-programming experts to contribute to robotic skill learning and may help to increase acceptance of robotic systems in shared workspaces and everyday life. In this talk, I will discuss recent research I did on interactive robot skill learning and the remaining challenges on the route to human-centered teaching of assistive robots. In particular, I will also discuss potential connections and overlap with cognitive science. The presented work covers learning a library of probabilistic movement primitives from human demonstrations, intention aware adaptation of learned skills in shared workspaces, and multi-channel interactive reinforcement learning for sequential tasks.
Network dynamics in the basal ganglia and possible implications for Parkinson’s disease
The basal ganglia are a collection of brain areas that are connected by a variety of synaptic pathways and are a site of significant reward-related dopamine release. These properties suggest a possible role for the basal ganglia in action selection, guided by reinforcement learning. In this talk, I will discuss a framework for how this function might be performed. I will also present some recent experimental results and theory that call for a re-evaluation of certain aspects of this framework. Next, I will turn to the changes in basal ganglia activity observed to occur with the dopamine depletion associated with Parkinson’s disease. I will discuss some of the potential functional implications of some of these changes and, if time permits, will conclude with some new results that focus on delta oscillations under dopamine depletion.
Higher cognitive resources for efficient learning
A central issue in reinforcement learning (RL) is the ‘curse-of-dimensionality’, arising when the degrees-of-freedom are much larger than the number of training samples. In such circumstances, the learning process becomes too slow to be plausible. In the brain, higher cognitive functions (such as abstraction or metacognition) may be part of the solution by generating low dimensional representations on which RL can operate. In this talk I will discuss a series of studies in which we used functional magnetic resonance imaging (fMRI) and computational modeling to investigate the neuro-computational basis of efficient RL. We found that people can learn remarkably complex task structures non-consciously, but also that - intriguingly - metacognition appears tightly coupled to this learning ability. Furthermore, when people use an explicit (conscious) policy to select relevant information, learning is accelerated by abstractions. At the neural level, prefrontal cortex subregions are differentially involved in separate aspects of learning: dorsolateral prefrontal cortex pairs with metacognitive processes, while ventromedial prefrontal cortex with valuation and abstraction. I will discuss the implications of these findings, in particular new questions on the function of metacognition in adaptive behavior and the link with abstraction.
Transforming task representations
Humans can adapt to a novel task on our first try. By contrast, artificial intelligence systems often require immense amounts of data to adapt. In this talk, I will discuss my recent work (https://www.pnas.org/content/117/52/32970) on creating deep learning systems that can adapt on their first try by exploiting relationships between tasks. Specifically, the approach is based on transforming a representation for a known task to produce a representation for the novel task, by inferring and then using a higher order function that captures a relationship between the tasks. This approach can be interpreted as a type of analogical reasoning. I will show that task transformation can allow systems to adapt to novel tasks on their first try in domains ranging from card games, to mathematical objects, to image classification and reinforcement learning. I will discuss the analogical interpretation of this approach, an analogy between levels of abstraction within the model architecture that I refer to as homoiconicity, and what this work might suggest about using deep-learning models to infer analogies more generally.
On cognitive maps and reinforcement learning in large-scale animal behaviour
Bats are extreme aviators and amazing navigators. Many bat species nightly commute dozens of kilometres in search of food, and some bat species annually migrate over thousands of kilometres. Studying bats in their natural environment has always been extremely challenging because of their small size (mostly <50 gr) and agile nature. We have recently developed novel miniature technology allowing us to GPS-tag small bats, thus opening a new window to document their behaviour in the wild. We have used this technology to track fruit-bats pups over 5 months from birth to adulthood. Following the bats’ full movement history allowed us to show that they use novel short-cuts which are typical for cognitive-map based navigation. In a second study, we examined how nectar-feeding bats make foraging decisions under competition. We show that by relying on a simple reinforcement learning strategy, the bats can divide the resource between them without aggression or communication. Together, these results demonstrate the power of the large scale natural approach for studying animal behavior.
Learning in pain: probabilistic inference and (mal)adaptive control
Pain is a major clinical problem affecting 1 in 5 people in the world. There are unresolved questions that urgently require answers to treat pain effectively, a crucial one being how the feeling of pain arises from brain activity. Computational models of pain consider how the brain processes noxious information and allow mapping neural circuits and networks to cognition and behaviour. To date, they have generally have assumed two largely independent processes: perceptual and/or predictive inference, typically modelled as an approximate Bayesian process, and action control, typically modelled as a reinforcement learning process. However, inference and control are intertwined in complex ways, challenging the clarity of this distinction. I will discuss how they may comprise a parallel hierarchical architecture that combines pain inference, information-seeking, and adaptive value-based control. Finally, I will discuss whether and how these learning processes might contribute to chronic pain.
Mental Simulation, Imagination, and Model-Based Deep RL
Mental simulation—the capacity to imagine what will or what could be—is a salient feature of human cognition, playing a key role in a wide range of cognitive abilities. In artificial intelligence, the last few years have seen the development of methods which are analogous to mental models and mental simulation. In this talk, I will discuss recent methods in deep learning for constructing such models from data and learning to use them via reinforcement learning, and compare such approaches to human mental simulation. While a number of challenges remain in matching the capacity of human mental simulation, I will highlight some recent progress on developing more compositional and efficient model-based algorithms through the use of graph neural networks and tree search.
Choice engineering and the modeling of operant learning
Organisms modify their behavior in response to its consequences, a phenomenon referred to as operant learning. Contemporary modeling of this learning behavior is based on reinforcement learning algorithms. I will discuss some of the challenges that these models face, and proposed a new approach to model-selection that is based on testing their ability to engineer behavior. Finally, I will present the results of The Choice Engineering Competition – an academic competition that compared the efficacies of qualitative and quantitative models of operant learning in shaping behavior.
Peril, Prudence and Planning as Risk, Avoidance and Worry
Risk occupies a central role in both the theory and practice of decision-making. Although it is deeply implicated in many conditions involving dysfunctional behavior and thought, modern theoretical approaches to understanding and mitigating risk in either one-shot or sequential settings, which are derived largely from finance and economics, have yet to permeate fully the fields of neural reinforcement learning and computational psychiatry. I will discuss the use of dynamic and static versions of one prominent approach, namely conditional value-at-risk, to examine both the nature of risk avoidant choices, encompassing such things as justified gambler's fallacies, and the optimal planning that can lead to consideration of such choices, with implications for offline, ruminative, thinking.
Navigation Turing Test: Toward Human-like RL
tbc
Uncertainty in learning and decision making
Uncertainty plays a critical role in reinforcement learning and decision making. However, exactly how subjective uncertainty influences behaviour remains unclear. Multi-armed bandits are a useful framework to gain more insight into this. Paired with computational tools such as Kalman filters, they allow us to closely characterize the interplay between trial-by-trial value, uncertainty, learning, and choice. In this talk, I will present recent research where we also measured participants visual fixations on the options in a multi-armed bandit task. The estimated value of each option, and the uncertainty in these estimations, influenced what subjects looked at in the period before making a choice and their subsequent choice, as additionally did fixation itself. Uncertainty also determined how long participants looked at the obtained outcomes. Our findings clearly show the importance of uncertainty in learning and decision making.
An inference perspective on meta-learning
While meta-learning algorithms are often viewed as algorithms that learn to learn, an alternative viewpoint frames meta-learning as inferring a hidden task variable from experience consisting of observations and rewards. From this perspective, learning to learn is learning to infer. This viewpoint can be useful in solving problems in meta-RL, which I’ll demonstrate through two examples: (1) enabling off-policy meta-learning, and (2) performing efficient meta-RL from image observations. I’ll also discuss how this perspective leads to an algorithm for few-shot image segmentation.
On cognitive maps and reinforcement learning in large-scale animal behaviour
Bats are extreme aviators and amazing navigators. Many bat species nightly com-mute dozens of kilometres in search of food, and some bat species annually migrate over thousands of kilometres. Studying bats in their natural environment has al-ways been extremely challenging because of their small size (mostly <50 gr) and agile nature. We have recently developed novel miniature technology allowing us to GPS-tag small bats, thus opening a new window to document their behaviour in the wild. We have used this technology to track fruit-bats pups over 5 months from birth to adulthood. Following the bats’ full movement history allowed us to show that they use novel short-cuts which are typical for cognitive-map based naviga-tion. In a second study, we examined how nectar-feeding bats make foraging deci-sions under competition. We show that by relying on a simple reinforcement learn-ing strategy, the bats can divide the resource between them without aggression or communication. Together, these results demonstrate the power of the large scale natural approach for studying animal behavior.
A journey through connectomics: from manual tracing to the first fully automated basal ganglia connectomes
The "mind of the worm", the first electron microscopy-based connectome of C. elegans, was an early sign of where connectomics is headed, followed by a long time of little progress in a field held back by the immense manual effort required for data acquisition and analysis. This changed over the last few years with several technological breakthroughs, which allowed increases in data set sizes by several orders of magnitude. Brain tissue can now be imaged in 3D up to a millimeter in size at nanometer resolution, revealing tissue features from synapses to the mitochondria of all contained cells. These breakthroughs in acquisition technology were paralleled by a revolution in deep-learning segmentation techniques, that equally reduced manual analysis times by several orders of magnitude, to the point where fully automated reconstructions are becoming useful. Taken together, this gives neuroscientists now access to the first wiring diagrams of thousands of automatically reconstructed neurons connected by millions of synapses, just one line of program code away. In this talk, I will cover these developments by describing the past few years' technological breakthroughs and discuss remaining challenges. Finally, I will show the potential of automated connectomics for neuroscience by demonstrating how hypotheses in reinforcement learning can now be tackled through virtual experiments in synaptic wiring diagrams of the songbird basal ganglia.
The geometry of abstraction in hippocampus and pre-frontal cortex
The curse of dimensionality plagues models of reinforcement learning and decision-making. The process of abstraction solves this by constructing abstract variables describing features shared by different specific instances, reducing dimensionality and enabling generalization in novel situations. Here we characterized neural representations in monkeys performing a task where a hidden variable described the temporal statistics of stimulus-response-outcome mappings. Abstraction was defined operationally using the generalization performance of neural decoders across task conditions not used for training. This type of generalization requires a particular geometric format of neural representations. Neural ensembles in dorsolateral pre-frontal cortex, anterior cingulate cortex and hippocampus, and in simulated neural networks, simultaneously represented multiple hidden and explicit variables in a format reflecting abstraction. Task events engaging cognitive operations modulated this format. These findings elucidate how the brain and artificial systems represent abstract variables, variables critical for generalization that in turn confers cognitive flexibility.
E-prop: A biologically inspired paradigm for learning in recurrent networks of spiking neurons
Transformative advances in deep learning, such as deep reinforcement learning, usually rely on gradient-based learning methods such as backpropagation through time (BPTT) as a core learning algorithm. However, BPTT is not argued to be biologically plausible, since it requires to a propagate gradients backwards in time and across neurons. Here, we propose e-prop, a novel gradient-based learning method with local and online weight update rules for recurrent neural networks, and in particular recurrent spiking neural networks (RSNNs). As a result, e-prop has the potential to provide a substantial fraction of the power of deep learning to RSNNs. In this presentation, we will motivate e-prop from the perspective of recent insights in neuroscience and show how these have to be combined to form an algorithm for online gradient descent. The mathematical results will be supported by empirical evidence in supervised and reinforcement learning tasks. We will also discuss how limitations that are inherited from gradient-based learning methods, such as sample-efficiency, can be addressed by considering an evolution-like optimization that enhances learning on particular task families. The emerging learning architecture can be used to learn tasks by a single demonstration, hence enabling one-shot learning.
Working memory transforms goals into rewards
Humans continuously need to learn to make good choices – be it using a new video-conferencing set up, figuring out what questions to ask to successfully secure a reliable babysitter, or just selecting which location in a house is least likely to be interrupted by toddlers during work calls. However, the goals we seek to attain – such as using zoom successfully – are often vaguely defined and previously unexperienced, and in that sense cannot be known by us as being rewarding. We hypothesized that learning to make good choices in such situations nevertheless leverages reinforcement learning processes, and that executive functions in general, and working memory in particular, play a crucial role in defining the reward function for arbitrary outcomes in such a way that they become reinforcing. I will show results from a novel behavioral protocol, as well as preliminary computational and imaging evidence supporting our hypothesis.
Reward foraging task, and model-based analysis reveal how fruit flies learn the value of available options
Understanding what drives foraging decisions in animals requires careful manipulation of the value of available options while monitoring animal choices. Value-based decision-making tasks, in combination with formal learning models, have provided both an experimental and theoretical framework to study foraging decisions in lab settings. While these approaches were successfully used in the past to understand what drives choices in mammals, very little work has been done on fruit flies. This is even though fruit flies have served as a model organism for many complex behavioural paradigms. To fill this gap we developed a single-animal, trial-based decision-making task, where freely walking flies experienced optogenetic sugar-receptor neuron stimulation. We controlled the value of available options by manipulating the probabilities of optogenetic stimulation. We show that flies integrate a reward history of chosen options and forget value of unchosen options. We further discover that flies assign higher values to rewards experienced early in the behavioural session, consistent with formal reinforcement learning models. Finally, we show that the probabilistic rewards affect walking trajectories of flies, suggesting that accumulated value is controlling the navigation vector of flies in a graded fashion. These findings establish the fruit fly as a model organism to explore the genetic and circuit basis of value-based decisions.
Deep reinforcement learning and its neuroscientific implications
The last few years have seen some dramatic developments in artificial intelligence research. What implications might these have for neuroscience? Investigations of this question have, to date, 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 more profound neuroscientific implications: Deep reinforcement learning. Deep RL offers a rich 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. I’ll provide a high level introduction to deep RL, discuss some recent neuroscience-oriented investigations from my group at DeepMind, and survey some wider implications for research on brain and behavior.
Striatal circuits for reward learning and decision-making
How are actions linked with subsequent outcomes to guide choices? The nucleus accumbens (NAc), which is implicated in this process, receives glutamatergic inputs from the prelimbic cortex (PL) and midline regions of the thalamus (mTH). However, little is known about what is represented in PL or mTH neurons that project to NAc (PL-NAc and mTH-NAc). By comparing these inputs during a reinforcement learning task in mice, we discovered that i) PL-NAc preferentially represents actions and choices, ii) mTH-NAc preferentially represents cues, iii) choice-selective activity in PL-NAc is organized in sequences that persist beyond the outcome. Through computational modelling, we demonstrate that these sequences can support the neural implementation of temporal difference learning, a powerful algorithm to connect actions and outcomes across time. Finally, we test and confirm predictions of our circuit model by direct manipulation of PL-NAc neurons. Thus, we integrate experiment and modelling to suggest a neural solution for credit assignment.
The geometry of abstraction in artificial and biological neural networks
The curse of dimensionality plagues models of reinforcement learning and decision-making. The process of abstraction solves this by constructing abstract variables describing features shared by different specific instances, reducing dimensionality and enabling generalization in novel situations. We characterized neural representations in monkeys performing a task where a hidden variable described the temporal statistics of stimulus-response-outcome mappings. Abstraction was defined operationally using the generalization performance of neural decoders across task conditions not used for training. This type of generalization requires a particular geometric format of neural representations. Neural ensembles in dorsolateral pre-frontal cortex, anterior cingulate cortex and hippocampus, and in simulated neural networks, simultaneously represented multiple hidden and explicit variables in a format reflecting abstraction. Task events engaging cognitive operations modulated this format. These findings elucidate how the brain and artificial systems represent abstract variables, variables critical for generalization that in turn confers cognitive flexibility.