If deep learning is the answer, what is the question?

Neuroscience research is undergoing a minor revolution. Recent advances in machine learning and artificial intelligence research have opened up new ways of thinking about neural computation. Many researchers are excited by the possibility that deep neural networks may offer theories of perception, cognition and action for biological brains. This approach has the potential to radically reshape our approach to understanding neural systems, because the computations performed by deep networks are learned from experience, and not endowed by the researcher. If so, how can neuroscientists use deep networks to model and understand biological brains? What is the outlook for neuroscientists who seek to characterize computations or neural codes, or who wish to understand perception, attention, memory and executive functions? In this Perspective, our goal is to offer a road map for systems neuroscience research in the age of deep learning. We discuss the conceptual and methodological challenges of comparing behaviour, learning dynamics and neural representations in artificial and biological systems, and we highlight new research questions that have emerged for neuroscience as a direct consequence of recent advances in machine learning.

Computational and systems neuroscience: The next 20 years.

Over the past 20 years, neuroscience has been propelled forward by theory-driven experimentation. We consider the future outlook for the field in the age of big neural data and powerful artificial intelligence models.

Using the Veil of Ignorance to align AI systems with principles of justice.

The philosopher John Rawls proposed the Veil of Ignorance (VoI) as a thought experiment to identify fair principles for governing a society. Here, we apply the VoI to an important governance domain: artificial intelligence (AI). In five incentive-compatible studies (N = 2, 508), including two preregistered protocols, participants choose principles to govern an Artificial Intelligence (AI) assistant from behind the veil: that is, without knowledge of their own relative position in the group. Compared to participants who have this information, we find a consistent preference for a principle that instructs the AI assistant to prioritize the worst-off. Neither risk attitudes nor political preferences adequately explain these choices. Instead, they appear to be driven by elevated concerns about fairness: Without prompting, participants who reason behind the VoI more frequently explain their choice in terms of fairness, compared to those in the Control condition. Moreover, we find initial support for the ability of the VoI to elicit more robust preferences: In the studies presented here, the VoI increases the likelihood of participants continuing to endorse their initial choice in a subsequent round where they know how they will be affected by the AI intervention and have a self-interested motivation to change their mind. These results emerge in both a descriptive and an immersive game. Our findings suggest that the VoI may be a suitable mechanism for selecting distributive principles to govern AI.

Abrupt and spontaneous strategy switches emerge in simple regularised neural networks.

Humans sometimes have an insight that leads to a sudden and drastic performance improvement on the task they are working on. Sudden strategy adaptations are often linked to insights, considered to be a unique aspect of human cognition tied to complex processes such as creativity or meta-cognitive reasoning. Here, we take a learning perspective and ask whether insight-like behaviour can occur in simple artificial neural networks, even when the models only learn to form input-output associations through gradual gradient descent. We compared learning dynamics in humans and regularised neural networks in a perceptual decision task that included a hidden regularity to solve the task more efficiently. Our results show that only some humans discover this regularity, and that behaviour is marked by a sudden and abrupt strategy switch that reflects an aha-moment. Notably, we find that simple neural networks with a gradual learning rule and a constant learning rate closely mimicked behavioural characteristics of human insight-like switches, exhibiting delay of insight, suddenness and selective occurrence in only some networks. Analyses of network architectures and learning dynamics revealed that insight-like behaviour crucially depended on a regularised gating mechanism and noise added to gradient updates, which allowed the networks to accumulate "silent knowledge" that is initially suppressed by regularised gating. This suggests that insight-like behaviour can arise from gradual learning in simple neural networks, where it reflects the combined influences of noise, gating and regularisation. These results have potential implications for more complex systems, such as the brain, and guide the way for future insight research.

Curriculum learning for human compositional generalization.

Generalization (or transfer) is the ability to repurpose knowledge in novel settings. It is often asserted that generalization is an important ingredient of human intelligence, but its extent, nature, and determinants have proved controversial. Here, we examine this ability with a paradigm that formalizes the transfer learning problem as one of recomposing existing functions to solve unseen problems. We find that people can generalize compositionally in ways that are elusive for standard neural networks and that human generalization benefits from training regimes in which items are axis aligned and temporally correlated. We describe a neural network model based around a Hebbian gating process that can capture how human generalization benefits from different training curricula. We additionally find that adult humans tend to learn composable functions asynchronously, exhibiting discontinuities in learning that resemble those seen in child development.

A recurrent neural network model of prefrontal brain activity during a working memory task.

When multiple items are held in short-term memory, cues that retrospectively prioritise one item over another (retro-cues) can facilitate subsequent recall. However, the neural and computational underpinnings of this effect are poorly understood. One recent study recorded neural signals in the macaque lateral prefrontal cortex (LPFC) during a retro-cueing task, contrasting delay-period activity before (pre-cue) and after (post-cue) retrocue onset. They reported that in the pre-cue delay, the individual stimuli were maintained in independent subspaces of neural population activity, whereas in the post-cue delay, the prioritised items were rotated into a common subspace, potentially allowing a common readout mechanism. To understand how such representational transitions can be learnt through error minimisation, we trained recurrent neural networks (RNNs) with supervision to perform an equivalent cued-recall task. RNNs were presented with two inputs denoting conjunctive colour-location stimuli, followed by a pre-cue memory delay, a location retrocue, and a post-cue delay. We found that the orthogonal-to-parallel geometry transformation observed in the macaque LPFC emerged naturally in RNNs trained to perform the task. Interestingly, the parallel geometry only developed when the cued information was required to be maintained in short-term memory for several cycles before readout, suggesting that it might confer robustness during maintenance. We extend these findings by analysing the learning dynamics and connectivity patterns of the RNNs, as well as the behaviour of models trained with probabilistic cues, allowing us to make predictions for future studies. Overall, our findings are consistent with recent theoretical accounts which propose that retrocues transform the prioritised memory items into a prospective, action-oriented format.

Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signals.

Humans can learn several tasks in succession with minimal mutual interference but perform more poorly when trained on multiple tasks at once. The opposite is true for standard deep neural networks. Here, we propose novel computational constraints for artificial neural networks, inspired by earlier work on gating in the primate prefrontal cortex, that capture the cost of interleaved training and allow the network to learn two tasks in sequence without forgetting. We augment standard stochastic gradient descent with two algorithmic motifs, so-called "sluggish" task units and a Hebbian training step that strengthens connections between task units and hidden units that encode task-relevant information. We found that the "sluggish" units introduce a switch-cost during training, which biases representations under interleaved training towards a joint representation that ignores the contextual cue, while the Hebbian step promotes the formation of a gating scheme from task units to the hidden layer that produces orthogonal representations which are perfectly guarded against interference. Validating the model on previously published human behavioural data revealed that it matches performance of participants who had been trained on blocked or interleaved curricula, and that these performance differences were driven by misestimation of the true category boundary.

Optimal utility and probability functions for agents with finite computational precision.

When making economic choices, such as those between goods or gambles, humans act as if their internal representation of the value and probability of a prospect is distorted away from its true value. These distortions give rise to decisions which apparently fail to maximize reward, and preferences that reverse without reason. Why would humans have evolved to encode value and probability in a distorted fashion, in the face of selective pressure for reward-maximizing choices? Here, we show that under the simple assumption that humans make decisions with finite computational precision--in other words, that decisions are irreducibly corrupted by noise--the distortions of value and probability displayed by humans are approximately optimal in that they maximize reward and minimize uncertainty. In two empirical studies, we manipulate factors that change the reward-maximizing form of distortion, and find that in each case, humans adapt optimally to the manipulation. This work suggests an answer to the longstanding question of why humans make "irrational" economic choices.

Quo vadis, planning?

Deep meta-learning is the driving force behind advances in contemporary AI research, and a promising theory of flexible cognition in natural intelligence. We agree with Binz et al. that many supposedly "model-based" behaviours may be better explained by meta-learning than by classical models. We argue that this invites us to revisit our neural theories of problem solving and goal-directed planning.

Model Sharing in the Human Medial Temporal Lobe.

Effective planning involves knowing where different actions take us. However, natural environments are rich and complex, leading to an exponential increase in memory demand as a plan grows in depth. One potential solution is to filter out features of the environment irrelevant to the task at hand. This enables a shared model of transition dynamics to be used for planning over a range of different input features. Here, we asked human participants (13 male, 16 female) to perform a sequential decision-making task, designed so that knowledge should be integrated independently of the input features (visual cues) present in one case but not in another. Participants efficiently switched between using a low-dimensional (cue independent) and a high-dimensional (cue specific) representation of state transitions. fMRI data identified the medial temporal lobe as a locus for learning state transitions. Within this region, multivariate patterns of BOLD responses were less correlated between trials with differing input features but similar state associations in the high dimensional than in the low dimensional case, suggesting that these patterns switched between separable (specific to input features) and shared (invariant to input features) transition models. Finally, we show that transition models are updated more strongly following the receipt of positive compared with negative outcomes, a finding that challenges conventional theories of planning. Together, these findings propose a computational and neural account of how information relevant for planning can be shared and segmented in response to the vast array of contextual features we encounter in our world.SIGNIFICANCE STATEMENT Effective planning involves maintaining an accurate model of which actions take us to which locations. But in a world awash with information, mapping actions to states with the right level of complexity is critical. Using a new decision-making "heist task" in conjunction with computational modeling and fMRI, we show that patterns of BOLD responses in the medial temporal lobe-a brain region key for prospective planning-become less sensitive to the presence of visual features when these are irrelevant to the task at hand. By flexibly adapting the complexity of task-state representations in this way, state-action mappings learned under one set of features can be used to plan in the presence of others.

How do (perceptual) distracters distract?

When a target stimulus occurs in the presence of distracters, decisions are less accurate. But how exactly do distracters affect choices? Here, we explored this question using measurement of human behaviour, psychophysical reverse correlation and computational modelling. We contrasted two models: one in which targets and distracters had independent influence on choices (independent model) and one in which distracters modulated choices in a way that depended on their similarity to the target (interaction model). Across three experiments, participants were asked to make fine orientation judgments about the tilt of a target grating presented adjacent to an irrelevant distracter. We found strong evidence for the interaction model, in that decisions were more sensitive when target and distracter were consistent relative to when they were inconsistent. This consistency bias occurred in the frame of reference of the decision, that is, it operated on decision values rather than on sensory signals, and surprisingly, it was independent of spatial attention. A normalization framework, where target features are normalized by the expectation and variability of the local context, successfully captures the observed pattern of results.

Normative Principles for Decision-Making in Natural Environments.

The decisions we make are shaped by a lifetime of learning. Past experience guides the way that we encode information in neural systems for perception and valuation, and determines the information we retrieve when making decisions. Distinct literatures have discussed how lifelong learning and local context shape decisions made about sensory signals, propositional information, or economic prospects. Here, we build bridges between these literatures, arguing for common principles of adaptive rationality in perception, cognition, and economic choice. We discuss how a single common framework, based on normative principles of efficient coding and Bayesian inference, can help us understand a myriad of human decision biases, including sensory illusions, adaptive aftereffects, choice history biases, central tendency effects, anchoring effects, contrast effects, framing effects, congruency effects, reference-dependent valuation, nonlinear utility functions, and discretization heuristics. We describe a simple computational framework for explaining these phenomena.

A map of decoy influence in human multialternative choice.

Human decisions can be biased by irrelevant information. For example, choices between two preferred alternatives can be swayed by a third option that is inferior or unavailable. Previous work has identified three classic biases, known as the attraction, similarity, and compromise effects, which arise during choices between economic alternatives defined by two attributes. However, the reliability, interrelationship, and computational origin of these three biases have been controversial. Here, a large cohort of human participants made incentive-compatible choices among assets that varied in price and quality. Instead of focusing on the three classic effects, we sampled decoy stimuli exhaustively across bidimensional multiattribute space and constructed a full map of decoy influence on choices between two otherwise preferred target items. Our analysis reveals that the decoy influence map is highly structured even beyond the three classic biases. We identify a very simple model that can fully reproduce the decoy influence map and capture its variability in individual participants. This model reveals that the three decoy effects are not distinct phenomena but are all special cases of a more general principle, by which attribute values are repulsed away from the context provided by rival options. The model helps us understand why the biases are typically correlated across participants and allows us to validate a prediction about their interrelationship. This work helps to clarify the origin of three of the most widely studied biases in human decision-making.

Thinking beyond the ventral stream: Comment on Bowers et al.

Bowers et al. rightly emphasise that deep learning models often fail to capture constraints on visual perception that have been discovered by previous research. However, the solution is not to discard deep learning altogether, but to design stimuli and tasks that more closely reflect the problems that biological vision evolved to solve, such as understanding scenes and preparing skilled action.

A practical guide for studying human behavior in the lab.

In the last few decades, the field of neuroscience has witnessed major technological advances that have allowed researchers to measure and control neural activity with great detail. Yet, behavioral experiments in humans remain an essential approach to investigate the mysteries of the mind. Their relatively modest technological and economic requisites make behavioral research an attractive and accessible experimental avenue for neuroscientists with very diverse backgrounds. However, like any experimental enterprise, it has its own inherent challenges that may pose practical hurdles, especially to less experienced behavioral researchers. Here, we aim at providing a practical guide for a steady walk through the workflow of a typical behavioral experiment with human subjects. This primer concerns the design of an experimental protocol, research ethics, and subject care, as well as best practices for data collection, analysis, and sharing. The goal is to provide clear instructions for both beginners and experienced researchers from diverse backgrounds in planning behavioral experiments.

A Normative Account of Confirmation Bias During Reinforcement Learning.

Reinforcement learning involves updating estimates of the value of states and actions on the basis of experience. Previous work has shown that in humans, reinforcement learning exhibits a confirmatory bias: when the value of a chosen option is being updated, estimates are revised more radically following positive than negative reward prediction errors, but the converse is observed when updating the unchosen option value estimate. Here, we simulate performance on a multi-arm bandit task to examine the consequences of a confirmatory bias for reward harvesting. We report a paradoxical finding: that confirmatory biases allow the agent to maximize reward relative to an unbiased updating rule. This principle holds over a wide range of experimental settings and is most influential when decisions are corrupted by noise. We show that this occurs because on average, confirmatory biases lead to overestimating the value of more valuable bandits and underestimating the value of less valuable bandits, rendering decisions overall more robust in the face of noise. Our results show how apparently suboptimal learning rules can in fact be reward maximizing if decisions are made with finite computational precision.

Hybrid computing using a neural network with dynamic external memory.

Artificial neural networks are remarkably adept at sensory processing, sequence learning and reinforcement learning, but are limited in their ability to represent variables and data structures and to store data over long timescales, owing to the lack of an external memory. Here we introduce a machine learning model called a differentiable neural computer (DNC), which consists of a neural network that can read from and write to an external memory matrix, analogous to the random-access memory in a conventional computer. Like a conventional computer, it can use its memory to represent and manipulate complex data structures, but, like a neural network, it can learn to do so from data. When trained with supervised learning, we demonstrate that a DNC can successfully answer synthetic questions designed to emulate reasoning and inference problems in natural language. We show that it can learn tasks such as finding the shortest path between specified points and inferring the missing links in randomly generated graphs, and then generalize these tasks to specific graphs such as transport networks and family trees. When trained with reinforcement learning, a DNC can complete a moving blocks puzzle in which changing goals are specified by sequences of symbols. Taken together, our results demonstrate that DNCs have the capacity to solve complex, structured tasks that are inaccessible to neural networks without external read-write memory.

Selective Integration during Sequential Sampling in Posterior Neural Signals.

Decisions are typically made after integrating information about multiple attributes of alternatives in a choice set. Where observers are obliged to consider attributes in turn, a computational framework known as "selective integration" can capture salient biases in human choices. The model proposes that successive attributes compete for processing resources and integration is biased towards the alternative with the locally preferred attribute. Quantitative analysis shows that this model, although it discards choice-relevant information, is optimal when the observers' decisions are corrupted by noise that occurs beyond the sensory stage. Here, we used electroencephalography (EEG) to test a neural prediction of the model: that locally preferred attributes should be encoded with higher gain in neural signals over the posterior cortex. Over two sessions, human observers judged which of the two simultaneous streams of bars had the higher (or lower) average height. The selective integration model fits the data better than a rival model without bias. Single-trial analysis showed that neural signals contralateral to the preferred attribute covaried more steeply with the decision information conferred by locally preferred attributes. These findings provide neural evidence in support of selective integration, complementing existing behavioral work.

Gain control explains the effect of distraction in human perceptual, cognitive, and economic decision making.

When making decisions, humans are often distracted by irrelevant information. Distraction has a different impact on perceptual, cognitive, and value-guided choices, giving rise to well-described behavioral phenomena such as the tilt illusion, conflict adaptation, or economic decoy effects. However, a single, unified model that can account for all these phenomena has yet to emerge. Here, we offer one such account, based on adaptive gain control, and additionally show that it successfully predicts a range of counterintuitive new behavioral phenomena on variants of a classic cognitive paradigm, the Eriksen flanker task. We also report that blood oxygen level-dependent signals in a dorsal network prominently including the anterior cingulate cortex index a gain-modulated decision variable predicted by the model. This work unifies the study of distraction across perceptual, cognitive, and economic domains.

Comparing continual task learning in minds and machines.

Humans can learn to perform multiple tasks in succession over the lifespan ("continual" learning), whereas current machine learning systems fail. Here, we investigated the cognitive mechanisms that permit successful continual learning in humans and harnessed our behavioral findings for neural network design. Humans categorized naturalistic images of trees according to one of two orthogonal task rules that were learned by trial and error. Training regimes that focused on individual rules for prolonged periods (blocked training) improved human performance on a later test involving randomly interleaved rules, compared with control regimes that trained in an interleaved fashion. Analysis of human error patterns suggested that blocked training encouraged humans to form "factorized" representation that optimally segregated the tasks, especially for those individuals with a strong prior bias to represent the stimulus space in a well-structured way. By contrast, standard supervised deep neural networks trained on the same tasks suffered catastrophic forgetting under blocked training, due to representational interference in the deeper layers. However, augmenting deep networks with an unsupervised generative model that allowed it to first learn a good embedding of the stimulus space (similar to that observed in humans) reduced catastrophic forgetting under blocked training. Building artificial agents that first learn a model of the world may be one promising route to solving continual task performance in artificial intelligence research.