Dynamic diversity is the answer to proxy failure.

We argue that a diverse and dynamic pool of agents mitigates proxy failure. Proxy modularity plays a key role in the ongoing production of diversity. We review examples from a range of scales.

Distributional reinforcement learning in prefrontal cortex.

The prefrontal cortex is crucial for learning and decision-making. Classic reinforcement learning (RL) theories center on learning the expectation of potential rewarding outcomes and explain a wealth of neural data in the prefrontal cortex. Distributional RL, on the other hand, learns the full distribution of rewarding outcomes and better explains dopamine responses. In the present study, we show that distributional RL also better explains macaque anterior cingulate cortex neuronal responses, suggesting that it is a common mechanism for reward-guided learning.

Generative replay underlies compositional inference in the hippocampal-prefrontal circuit.

Human reasoning depends on reusing pieces of information by putting them together in new ways. However, very little is known about how compositional computation is implemented in the brain. Here, we ask participants to solve a series of problems that each require constructing a whole from a set of elements. With fMRI, we find that representations of novel constructed objects in the frontal cortex and hippocampus are relational and compositional. With MEG, we find that replay assembles elements into compounds, with each replay sequence constituting a hypothesis about a possible configuration of elements. The content of sequences evolves as participants solve each puzzle, progressing from predictable to uncertain elements and gradually converging on the correct configuration. Together, these results suggest a computational bridge between apparently distinct functions of hippocampal-prefrontal circuitry and a role for generative replay in compositional inference and hypothesis testing.

Replay and compositional computation.

Replay in the brain has been viewed as rehearsal or, more recently, as sampling from a transition model. Here, we propose a new hypothesis: that replay is able to implement a form of compositional computation where entities are assembled into relationally bound structures to derive qualitatively new knowledge. This idea builds on recent advances in neuroscience, which indicate that the hippocampus flexibly binds objects to generalizable roles and that replay strings these role-bound objects into compound statements. We suggest experiments to test our hypothesis, and we end by noting the implications for AI systems which lack the human ability to radically generalize past experience to solve new problems.

Investigating associations of delay discounting with brain structure, working memory, and episodic memory.

INTRODUCTION: Delay discounting (DD), the preference for smaller and sooner rewards over larger and later ones, is an important behavioural phenomenon for daily functioning of increasing interest within psychopathology. The neurobiological mechanisms behind DD are not well understood and the literature on structural correlates of DD shows inconsistencies. METHODS: Here we leveraged a large openly available dataset (n = 1196) to investigate associations with memory performance and gray and white matter correlates of DD using linked independent component analysis. RESULTS: Greater DD was related to smaller anterior temporal gray matter volume. Associations of DD with total cortical volume, subcortical volumes, markers of white matter microscopic organization, working memory, and episodic memory scores were not significant after controlling for education and income. CONCLUSION: Effects of size comparable to the one we identified would be unlikely to be replicated with sample sizes common in many previous studies in this domain, which may explain the incongruities in the literature. The paucity and small size of the effects detected in our data underscore the importance of using large samples together with methods that accommodate their statistical structure and appropriate control for confounders, as well as the need to devise paradigms with improved task parameter reliability in studies relating brain structure and cognitive abilities with DD.

Dreading the pain of others? Altruistic responses to others' pain underestimate dread.

A dislike of waiting for pain, aptly termed 'dread', is so great that people will increase pain to avoid delaying it. However, despite many accounts of altruistic responses to pain in others, no previous studies have tested whether people take delay into account when attempting to ameliorate others' pain. We examined the impact of delay in 2 experiments where participants (total N = 130) specified the intensity and delay of pain either for themselves or another person. Participants were willing to increase the experimental pain of another participant to avoid delaying it, indicative of dread, though did so to a lesser extent than was the case for their own pain. We observed a similar attenuation in dread when participants chose the timing of a hypothetical painful medical treatment for a close friend or relative, but no such attenuation when participants chose for a more distant acquaintance. A model in which altruism is biased to privilege pain intensity over the dread of pain parsimoniously accounts for these findings. We refer to this underestimation of others' dread as a 'Dread Empathy Gap'.

Impaired neural replay of inferred relationships in schizophrenia.

An ability to build structured mental maps of the world underpins our capacity to imagine relationships between objects that extend beyond experience. In rodents, such representations are supported by sequential place cell reactivations during rest, known as replay. Schizophrenia is proposed to reflect a compromise in structured mental representations, with animal models reporting abnormalities in hippocampal replay and associated ripple activity during rest. Here, utilizing magnetoencephalography (MEG), we tasked patients with schizophrenia and control participants to infer unobserved relationships between objects by reorganizing visual experiences containing these objects. During a post-task rest session, controls exhibited fast spontaneous neural reactivation of presented objects that replayed inferred relationships. Replay was coincident with increased ripple power in hippocampus. Patients showed both reduced replay and augmented ripple power relative to controls, convergent with findings in animal models. These abnormalities are linked to impairments in behavioral acquisition and subsequent neural representation of task structure.

Temporally delayed linear modelling (TDLM) measures replay in both animals and humans.

There are rich structures in off-task neural activity which are hypothesized to reflect fundamental computations across a broad spectrum of cognitive functions. Here, we develop an analysis toolkit - temporal delayed linear modelling (TDLM) - for analysing such activity. TDLM is a domain-general method for finding neural sequences that respect a pre-specified transition graph. It combines nonlinear classification and linear temporal modelling to test for statistical regularities in sequences of task-related reactivations. TDLM is developed on the non-invasive neuroimaging data and is designed to take care of confounds and maximize sequence detection ability. Notably, as a linear framework, TDLM can be easily extended, without loss of generality, to capture rodent replay in electrophysiology, including in continuous spaces, as well as addressing second-order inference questions, for example, its temporal and spatial varying pattern. We hope TDLM will advance a deeper understanding of neural computation and promote a richer convergence between animal and human neuroscience.

Overcoming Pavlovian bias in semantic space.

Action is invigorated in the presence of reward-predicting stimuli and inhibited in the presence of punishment-predicting stimuli. Although valuable as a heuristic, this Pavlovian bias can also lead to maladaptive behaviour and is implicated in addiction. Here we explore whether Pavlovian bias can be overcome through training. Across five experiments, we find that Pavlovian bias is resistant to unlearning under most task configurations. However, we demonstrate that when subjects engage in instrumental learning in a verbal semantic space, as opposed to a motoric space, not only do they exhibit the typical Pavlovian bias, but this Pavlovian bias diminishes with training. Our results suggest that learning within the semantic space is necessary, but not sufficient, for subjects to unlearn their Pavlovian bias, and that other task features, such as gamification and spaced stimulus presentation may also be necessary. In summary, we show that Pavlovian bias, whilst robust, is susceptible to change with experience, but only under specific environmental conditions.

Replay bursts in humans coincide with activation of the default mode and parietal alpha networks.

Our brains at rest spontaneously replay recently acquired information, but how this process is orchestrated to avoid interference with ongoing cognition is an open question. Here we investigated whether replay coincided with spontaneous patterns of whole-brain activity. We found, in two separate datasets, that replay sequences were packaged into transient bursts occurring selectively during activation of the default mode network (DMN) and parietal alpha networks. These networks are believed to support inwardly oriented attention and inhibit bottom-up sensory processing and were characterized by widespread synchronized oscillations coupled to increases in high frequency power, mechanisms thought to coordinate information flow between disparate cortical areas. Our data reveal a tight correspondence between two widely studied phenomena in neural physiology and suggest that the DMN may coordinate replay bursts in a manner that minimizes interference with ongoing cognition.

Social discounting of pain.

Impatience can be formalized as a delay discount rate, describing how the subjective value of reward decreases as it is delayed. By analogy, selfishness can be formalized as a social discount rate, representing how the subjective value of rewarding another person decreases with increasing social distance. Delay and social discount rates for reward are correlated across individuals. However no previous work has examined whether this relationship also holds for aversive outcomes. Neither has previous work described a functional form for social discounting of pain in humans. This is a pertinent question, since preferences over aversive outcomes formally diverge from those for reward. We addressed this issue in an experiment in which healthy adult participants (N = 67) chose the timing and intensity of hypothetical pain for themselves and others. In keeping with previous studies, participants showed a strong preference for immediate over delayed pain. Participants showed greater concern for pain in close others than for their own pain, though this hyperaltruism was steeply discounted with increasing social distance. Impatience for pain and social discounting of pain were weakly correlated across individuals. Our results extend a link between impatience and selfishness to the aversive domain.

Deep Reinforcement Learning and Its Neuroscientific Implications.

The emergence of powerful artificial intelligence (AI) 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 that has so far received less attention from neuroscientists but that may have profound neuroscientific implications: deep reinforcement learning (RL). 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.

The value of what's to come: Neural mechanisms coupling prediction error and the utility of anticipation.

Having something to look forward to is a keystone of well-being. Anticipation of future reward, such as an upcoming vacation, can often be more gratifying than the experience itself. Theories suggest the utility of anticipation underpins various behaviors, ranging from beneficial information-seeking to harmful addiction. However, how neural systems compute anticipatory utility remains unclear. We analyzed the brain activity of human participants as they performed a task involving choosing whether to receive information predictive of future pleasant outcomes. Using a computational model, we show three brain regions orchestrate anticipatory utility. Specifically, ventromedial prefrontal cortex tracks the value of anticipatory utility, dopaminergic midbrain correlates with information that enhances anticipation, while sustained hippocampal activity mediates a functional coupling between these regions. Our findings suggest a previously unidentified neural underpinning for anticipation's influence over decision-making and unify a range of phenomena associated with risk and time-delay preference.

Social training reconfigures prediction errors to shape Self-Other boundaries.

Selectively attributing beliefs to specific agents is core to reasoning about other people and imagining oneself in different states. Evidence suggests humans might achieve this by simulating each other's computations in agent-specific neural circuits, but it is not known how circuits become agent-specific. Here we investigate whether agent-specificity adapts to social context. We train subjects on social learning tasks, manipulating the frequency with which self and other see the same information. Training alters the agent-specificity of prediction error (PE) circuits for at least 24 h, modulating the extent to which another agent's PE is experienced as one's own and influencing perspective-taking in an independent task. Ventromedial prefrontal myelin density, indexed by magnetisation transfer, correlates with the strength of this adaptation. We describe a frontotemporal learning network, which exploits relationships between different agents' computations. Our findings suggest that Self-Other boundaries are learnable variables, shaped by the statistical structure of social experience.

A distributional code for value in dopamine-based reinforcement learning.

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.

Human Replay Spontaneously Reorganizes Experience.

Knowledge abstracted from previous experiences can be transferred to aid new learning. Here, we asked whether such abstract knowledge immediately guides the replay of new experiences. We first trained participants on a rule defining an ordering of objects and then presented a novel set of objects in a scrambled order. Across two studies, we observed that representations of these novel objects were reactivated during a subsequent rest. As in rodents, human "replay" events occurred in sequences accelerated in time, compared to actual experience, and reversed their direction after a reward. Notably, replay did not simply recapitulate visual experience, but followed instead a sequence implied by learned abstract knowledge. Furthermore, each replay contained more than sensory representations of the relevant objects. A sensory code of object representations was preceded 50 ms by a code factorized into sequence position and sequence identity. We argue that this factorized representation facilitates the generalization of a previously learned structure to new objects.

Reinforcement Learning, Fast and Slow.

Deep reinforcement learning (RL) methods have driven impressive advances in artificial intelligence in recent years, exceeding human performance in domains ranging from Atari to Go to no-limit poker. This progress has drawn the attention of cognitive scientists interested in understanding human learning. However, the concern has been raised that deep RL may be too sample-inefficient - that is, it may simply be too slow - to provide a plausible model of how humans learn. In the present review, we counter this critique by describing recently developed techniques that allow deep RL to operate more nimbly, solving problems much more quickly than previous methods. Although these techniques were developed in an AI context, we propose that they may have rich implications for psychology and neuroscience. A key insight, arising from these AI methods, concerns the fundamental connection between fast RL and slower, more incremental forms of learning.

Computing Value from Quality and Quantity in Human Decision-Making.

How organisms learn the value of single stimuli through experience is well described. In many decisions, however, value estimates are computed "on the fly" by combining multiple stimulus attributes. The neural basis of this computation is poorly understood. Here we explore a common scenario in which decision-makers must combine information about quality and quantity to determine the best option. Using fMRI, we examined the neural representation of quality, quantity, and their integration into an integrated subjective value signal in humans of both genders. We found that activity within inferior frontal gyrus (IFG) correlated with offer quality, while activity in the intraparietal sulcus (IPS) specifically correlated with offer quantity. Several brain regions, including the anterior cingulate cortex (ACC), were sensitive to an interaction of quality and quantity. However, the ACC was uniquely activated by quality, quantity, and their interaction, suggesting that this region provides a substrate for flexible computation of value from both quality and quantity. Furthermore, ACC signals across subjects correlated with the strength of quality and quantity signals in IFG and IPS, respectively. ACC tracking of subjective value also correlated with choice predictability. Finally, activity in the ACC was elevated for choice trials, suggesting that ACC provides a nexus for the computation of subjective value in multiattribute decision-making.SIGNIFICANCE STATEMENT Would you prefer three apples or two oranges? Many choices we make each day require us to weigh up the quality and quantity of different outcomes. Using fMRI, we show that option quality is selectively represented in the inferior frontal gyrus, while option quantity correlates with areas of the intraparietal sulcus that have previously been associated with numerical processing. We show that information about the two is integrated into a value signal in the anterior cingulate cortex, and the fidelity of this integration predicts choice predictability. Our results demonstrate how on-the-fly value estimates are computed from multiple attributes in human value-based decision-making.

What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior.

It is proposed that a cognitive map encoding the relationships between entities in the world supports flexible behavior, but the majority of the neural evidence for such a system comes from studies of spatial navigation. Recent work describing neuronal parallels between spatial and non-spatial behaviors has rekindled the notion of a systematic organization of knowledge across multiple domains. We review experimental evidence and theoretical frameworks that point to principles unifying these apparently disparate functions. These principles describe how to learn and use abstract, generalizable knowledge and suggest that map-like representations observed in a spatial context may be an instance of general coding mechanisms capable of organizing knowledge of all kinds. We highlight how artificial agents endowed with such principles exhibit flexible behavior and learn map-like representations observed in the brain. Finally, we speculate on how these principles may offer insight into the extreme generalizations, abstractions, and inferences that characterize human cognition.

Prefrontal cortex as a meta-reinforcement learning system.

Over the past 20 years, neuroscience research on reward-based learning has converged on a canonical model, under which the neurotransmitter dopamine 'stamps in' associations between situations, actions and rewards by modulating the strength of synaptic connections between neurons. However, a growing number of recent findings have placed this standard model under strain. We now draw on recent advances in artificial intelligence to introduce a new theory of reward-based learning. Here, the dopamine system trains another part of the brain, the prefrontal cortex, to operate as its own free-standing learning system. This new perspective accommodates the findings that motivated the standard model, but also deals gracefully with a wider range of observations, providing a fresh foundation for future research.