Beyond Reward Prediction Errors: Human Striatum Updates Rule Values During Learning

Distinct hippocampal mechanisms support concept formation and updating

Learning systems must constantly decide whether to create new representations or update existing ones. For example, a child learning that a bat is a mammal and not a bird would be best served by creating a new representation, whereas updating may be best when encountering a second similar bat. Characterizing the neural dynamics that underlie these complementary memory operations requires identifying the exact moments when each operation occurs. We address this challenge by interrogating fMRI brain activation with a computational learning model that predicts trial-by-trial when memories are created versus updated. We found distinct neural engagement in anterior hippocampus and ventral striatum for model-predicted memory create and update events during early learning. Notably, the degree of this effect in hippocampus, but not ventral striatum, significantly related to learning outcome. Hippocampus additionally showed distinct patterns of functional coactivation with ventromedial prefrontal cortex and angular gyrus during memory creation and premotor cortex during memory updating. These findings suggest that complementary memory functions, as formalized in computational learning models, underlie the rapid formation of novel conceptual knowledge, with the hippocampus and its interactions with frontoparietal circuits playing a crucial role in successful learning.

Significance statement

How do we reconcile new experiences with existing knowledge? Prominent theories suggest that novel information is either captured by creating new memories or leveraged to update existing memories, yet empirical support of how these distinct memory operations unfold during learning is limited. Here, we combine computational modeling of human learning behaviour with functional neuroimaging to identify moments of memory formation and updating and characterize their neural signatures. We find that both hippocampus and ventral striatum are distinctly engaged when memories are created versus updated; however, it is only hippocampus activation that is associated with learning outcomes. Our findings motivate a key theoretical revision that positions hippocampus is a key player in building organized memories from the earliest moments of learning.

Choice Type Impacts Human Reinforcement Learning

In reinforcement learning (RL) experiments, participants learn to make rewarding choices in response to different stimuli; RL models use outcomes to estimate stimulus-response values that change incrementally. RL models consider any response type indiscriminately, ranging from more concretely defined motor choices (pressing a key with the index finger), to more general choices that can be executed in a number of ways (selecting dinner at the restaurant). However, does the learning process vary as a function of the choice type? In Experiment 1, we show that it does: Participants were slower and less accurate in learning correct choices of a general format compared with learning more concrete motor actions. Using computational modeling, we show that two mechanisms contribute to this. First, there was evidence of irrelevant credit assignment: The values of motor actions interfered with the values of other choice dimensions, resulting in more incorrect choices when the correct response was not defined by a single motor action; second, information integration for relevant general choices was slower. In Experiment 2, we replicated and further extended the findings from Experiment 1 by showing that slowed learning was attributable to weaker working memory use, rather than slowed RL. In both experiments, we ruled out the explanation that the difference in performance between two condition types was driven by difficulty/different levels of complexity. We conclude that defining a more abstract choice space used by multiple learning systems for credit assignment recruits executive resources, limiting how much such processes then contribute to fast learning.

Humans combine value learning and hypothesis testing strategically in multi-dimensional probabilistic reward learning

Realistic and complex decision tasks often allow for many possible solutions. How do we find the correct one? Introspection suggests a process of trying out solutions one after the other until success. However, such methodical serial testing may be too slow, especially in environments with noisy feedback. Alternatively, the underlying learning process may involve implicit reinforcement learning that learns about many possibilities in parallel. Here we designed a multi-dimensional probabilistic active-learning task tailored to study how people learn to solve such complex problems. Participants configured three-dimensional stimuli by selecting features for each dimension and received probabilistic reward feedback. We manipulated task complexity by changing how many feature dimensions were relevant to maximizing reward, as well as whether this information was provided to the participants. To investigate how participants learn the task, we examined models of serial hypothesis testing, feature-based reinforcement learning, and combinations of the two strategies. Model comparison revealed evidence for hypothesis testing that relies on reinforcement-learning when selecting what hypothesis to test. The extent to which participants engaged in hypothesis testing depended on the instructed task complexity: people tended to serially test hypotheses when instructed that there were fewer relevant dimensions, and relied more on gradual and parallel learning of feature values when the task was more complex. This demonstrates a strategic use of task information to balance the costs and benefits of the two methods of learning.

Striatal reactivity during emotion and reward relates to approach-avoidance conflict behaviour and is altered in adults with anxiety or depression

BACKGROUND

We have previously reported activation in reward, salience and executive control regions during functional MRI (fMRI) using an approach-avoidance conflict (AAC) decision-making task with healthy adults. Further investigations into how anxiety and depressive disorders relate to differences in neural responses during AAC can inform their understanding and treatment. We tested the hypothesis that people with anxiety or depression have altered neural activation during AAC.

METHODS

We compared 118 treatment-seeking adults with anxiety or depression and 58 healthy adults using linear mixed-effects models to examine group-level differences in neural activation (fMRI) during AAC decision-making. Correlational analyses examined relationships between behavioural and neural measures.

RESULTS

Adults with anxiety or depression had greater striatal engagement when reacting to affective stimuli (p = 0.008, d = 0.31) regardless of valence, and weaker striatal engagement during reward feedback (p = 0.046, d = -0.27) regardless of the presence of monetary reward. They also had blunted amygdala activity during decision-making (p = 0.023, d = -0.32) regardless of the presence of conflict. Across groups, approach behaviour during conflict decision-making was inversely correlated with striatal activation during affective stimuli (p < 0.001, r = -0.28) and positively related to striatal activation during reward feedback (p < 0.001, r = 0.27).

LIMITATIONS

Our transdiagnostic approach did not allow for comparisons between specific anxiety disorders, and our cross-sectional approach did not allow for causal inference.

CONCLUSION

Anxiety and depression were associated with altered neural responses to AAC. Findings were consistent with the role of the striatum in action selection and reward responsivity, and they point toward striatal reactivity as a future treatment target. Blunting of amygdala activity in anxiety or depression may indicate a compensatory response to inhibit affective salience and maintain approach.

Rational arbitration between statistics and rules in human sequence processing

Detecting and learning temporal regularities is essential to accurately predict the future. A long-standing debate in cognitive science concerns the existence in humans of a dissociation between two systems, one for handling statistical regularities governing the probabilities of individual items and their transitions, and another for handling deterministic rules. Here, to address this issue, we used finger tracking to continuously monitor the online build-up of evidence, confidence, false alarms and changes-of-mind during sequence processing. All these aspects of behaviour conformed tightly to a hierarchical Bayesian inference model with distinct hypothesis spaces for statistics and rules, yet linked by a single probabilistic currency. Alternative models based either on a single statistical mechanism or on two non-commensurable systems were rejected. Our results indicate that a hierarchical Bayesian inference mechanism, capable of operating over distinct hypothesis spaces for statistics and rules, underlies the human capability for sequence processing.

Decision uncertainty during hypothesis testing enhances memory accuracy for incidental information

Humans actively seek information to reduce uncertainty, providing insight on how our decisions causally affect the world. While we know that episodic memories can help support future goal-oriented behaviors, little is known about how hypothesis testing during exploration influences episodic memory. To investigate this question, we designed a hypothesis testing paradigm, in which participants figured out rules to unlock treasure chests. Using this paradigm, we characterized how hypothesis testing during exploration influenced memory for the contents of the treasure chests. We found that there was an inverted U-shaped relationship between decision uncertainty and memory, such that memory was best when decision uncertainty was moderate. An exploratory analysis also showed that surprising outcomes lead to lower memory confidence independent of accuracy. These findings support a model in which moderate decision uncertainty during hypothesis testing enhances incidental information encoding.

Advances in modeling learning and decision-making in neuroscience

An organism's survival depends on its ability to learn about its environment and to make adaptive decisions in the service of achieving the best possible outcomes in that environment. To study the neural circuits that support these functions, researchers have increasingly relied on models that formalize the computations required to carry them out. Here, we review the recent history of computational modeling of learning and decision-making, and how these models have been used to advance understanding of prefrontal cortex function. We discuss how such models have advanced from their origins in basic algorithms of updating and action selection to increasingly account for complexities in the cognitive processes required for learning and decision-making, and the representations over which they operate. We further discuss how a deeper understanding of the real-world complexities in these computations has shed light on the fundamental constraints on optimal behavior, and on the complex interactions between corticostriatal pathways to determine such behavior. The continuing and rapid development of these models holds great promise for understanding the mechanisms by which animals adapt to their environments, and what leads to maladaptive forms of learning and decision-making within clinical populations.

Category learning in a recurrent neural network with reinforcement learning

It is known that humans and animals can learn and utilize category information quickly and efficiently to adapt to changing environments, and several brain areas are involved in learning and encoding category information. However, it is unclear that how the brain system learns and forms categorical representations from the view of neural circuits. In order to investigate this issue from the network level, we combine a recurrent neural network with reinforcement learning to construct a deep reinforcement learning model to demonstrate how the category is learned and represented in the network. The model consists of a policy network and a value network. The policy network is responsible for updating the policy to choose actions, while the value network is responsible for evaluating the action to predict rewards. The agent learns dynamically through the information interaction between the policy network and the value network. This model was trained to learn six stimulus-stimulus associative chains in a sequential paired-association task that was learned by the monkey. The simulated results demonstrated that our model was able to learn the stimulus-stimulus associative chains, and successfully reproduced the similar behavior of the monkey performing the same task. Two types of neurons were found in this model: one type primarily encoded identity information about individual stimuli; the other type mainly encoded category information of associated stimuli in one chain. The two types of activity-patterns were also observed in the primate prefrontal cortex after the monkey learned the same task. Furthermore, the ability of these two types of neurons to encode stimulus or category information was enhanced during this model was learning the task. Our results suggest that the neurons in the recurrent neural network have the ability to form categorical representations through deep reinforcement learning during learning stimulus-stimulus associations. It might provide a new approach for understanding neuronal mechanisms underlying how the prefrontal cortex learns and encodes category information.

Multitask learning over shared subspaces

This paper uses constructs from machine learning to define pairs of learning tasks that either shared or did not share a common subspace. Human subjects then learnt these tasks using a feedback-based approach and we hypothesised that learning would be boosted for shared subspaces. Our findings broadly supported this hypothesis with either better performance on the second task if it shared the same subspace as the first, or positive correlations over task performance for shared subspaces. These empirical findings were compared to the behaviour of a Neural Network model trained using sequential Bayesian learning and human performance was found to be consistent with a minimal capacity variant of this model. Networks with an increased representational capacity, and networks without Bayesian learning, did not show these transfer effects. We propose that the concept of shared subspaces provides a useful framework for the experimental study of human multitask and transfer learning.

How does knowledge gained from previous experience affect learning of new tasks? This question of “Transfer Learning” has been addressed by teachers, psychologists, and more recently by researchers in the fields of neural networks and machine learning. Leveraging constructs from machine learning, we designed pairs of learning tasks that either shared or did not share a common subspace. We compared the dynamics of transfer learning in humans with those of a multitask neural network model, finding that human performance was consistent with a minimal capacity variant of the model. Learning was boosted in the second task if the same subspace was shared between tasks. Additionally, accuracy between tasks was positively correlated but only when they shared the same subspace. Our results highlight the roles of subspaces, showing how they could act as a learning boost if shared, and be detrimental if not.

Confirmation of Interpersonal Expectations is Intrinsically Rewarding

People want to interact successfully with other individuals, and they invest significant efforts in attempting to do so. Decades of research have demonstrated that to simplify the dauntingly complex task of interpersonal communication, perceivers predict the responses of individuals in their environment using stereotypes and other sources of prior knowledge. Here, we show that these top-down expectations can also shape the subjective value of expectation-consistent and expectation-violating targets. Specifically, in two neuroimaging experiments (n = 58), we observed increased activation in brain regions associated with reward processing-including the nucleus accumbens-when perceivers observed information consistent with their social expectations. In two additional behavioral experiments (n = 704), we observed that perceivers were willing to forgo money to encounter an expectation-consistent target and avoid an expectation-violating target. Together, these findings suggest that perceivers value having their social expectations confirmed, much like food or monetary rewards.

Human Representation Learning

The central theme of this review is the dynamic interaction between information selection and learning. We pose a fundamental question about this interaction: How do we learn what features of our experiences are worth learning about? In humans, this process depends on attention and memory, two cognitive functions that together constrain representations of the world to features that are relevant for goal attainment. Recent evidence suggests that the representations shaped by attention and memory are themselves inferred from experience with each task. We review this evidence and place it in the context of work that has explicitly characterized representation learning as statistical inference. We discuss how inference can be scaled to real-world decisions by approximating beliefs based on a small number of experiences. Finally, we highlight some implications of this inference process for human decision-making in social environments.

Hippocampal pattern separation supports reinforcement learning

Animals rely on learned associations to make decisions. Associations can be based on relationships between object features (e.g., the three leaflets of poison ivy leaves) and outcomes (e.g., rash). More often, outcomes are linked to multidimensional states (e.g., poison ivy is green in summer but red in spring). Feature-based reinforcement learning fails when the values of individual features depend on the other features present. One solution is to assign value to multi-featural conjunctive representations. Here, we test if the hippocampus forms separable conjunctive representations that enables the learning of response contingencies for stimuli of the form: AB+, B-, AC-, C+. Pattern analyses on functional MRI data show the hippocampus forms conjunctive representations that are dissociable from feature components and that these representations, along with those of cortex, influence striatal prediction errors. Our results establish a novel role for hippocampal pattern separation and conjunctive representation in reinforcement learning.

Holistic Reinforcement Learning: The Role of Structure and Attention

Compact representations of the environment allow humans to behave efficiently in a complex world. Reinforcement learning models capture many behavioral and neural effects but do not explain recent findings showing that structure in the environment influences learning. In parallel, Bayesian cognitive models predict how humans learn structured knowledge but do not have a clear neurobiological implementation. We propose an integration of these two model classes in which structured knowledge learned via approximate Bayesian inference acts as a source of selective attention. In turn, selective attention biases reinforcement learning towards relevant dimensions of the environment. An understanding of structure learning will help to resolve the fundamental challenge in decision science: explaining why people make the decisions they do.

Moral contagion: Devaluation effect of immorality on hypothetical judgments of economic value

Moral contagion is a phenomenon in which individuals or objects take on the moral essence of the people who are associated with them. Previous studies have found that individuals value objects associated with moral and likable people more than those associated with immoral and dislikable people. However, the neural mechanisms underlying this "moral contagion effect" have not yet been explored. In the present study, we combined a novel "Second-hand Goods Pricing" paradigm with functional magnetic resonance imaging to (a) confirm the existence of the moral contagion effect on the hypothetical economic valuation of objects and (b) determine the neural substrates underlying it. Participants were shown second-hand goods, information regarding the moral valence of the previous owner, and an initial price assigned to the object by computer. The participants were then asked to adjust the initial price to one they deemed most reasonable. Behaviorally, we found a significant devaluation effect for immoral owners and a weaker reverse effect for moral owners. Imaging data showed that the devaluation effect was primarily driven by neural responses in the dorsal striatum (mainly the caudate nucleus) that were triggered by high initial prices assigned to the "contaminated" objects. Dynamic causal modeling revealed that the high initial price assigned to "contaminated" objects led to increased effective connectivity from the caudate nucleus to the ventromedial prefrontal cortex-the brain area that integrates values during decision making.

Predicting change: Approximate inference under explicit representation of temporal structure in changing environments

In our daily lives timing of our actions plays an essential role when we navigate the complex everyday environment. It is an open question though how the representations of the temporal structure of the world influence our behavior. Here we propose a probabilistic model with an explicit representation of state durations which may provide novel insights in how the brain predicts upcoming changes. We illustrate several properties of the behavioral model using a standard reversal learning design and compare its task performance to standard reinforcement learning models. Furthermore, using experimental data, we demonstrate how the model can be applied to identify participants' beliefs about the latent temporal task structure. We found that roughly one quarter of participants seem to have learned the latent temporal structure and used it to anticipate changes, whereas the remaining participants' behavior did not show signs of anticipatory responses, suggesting a lack of precise temporal expectations. We expect that the introduced behavioral model will allow, in future studies, for a systematic investigation of how participants learn the underlying temporal structure of task environments and how these representations shape behavior.

Dopaminergic basis for signaling belief updates, but not surprise, and the link to paranoia

Distinguishing between meaningful and meaningless sensory information is fundamental to forming accurate representations of the world. Dopamine is thought to play a central role in processing the meaningful information content of observations, which motivates an agent to update their beliefs about the environment. However, direct evidence for dopamine's role in human belief updating is lacking. We addressed this question in healthy volunteers who performed a model-based fMRI task designed to separate the neural processing of meaningful and meaningless sensory information. We modeled participant behavior using a normative Bayesian observer model and used the magnitude of the model-derived belief update following an observation to quantify its meaningful information content. We also acquired PET imaging measures of dopamine function in the same subjects. We show that the magnitude of belief updates about task structure (meaningful information), but not pure sensory surprise (meaningless information), are encoded in midbrain and ventral striatum activity. Using PET we show that the neural encoding of meaningful information is negatively related to dopamine-2/3 receptor availability in the midbrain and dexamphetamine-induced dopamine release capacity in the striatum. Trial-by-trial analysis of task performance indicated that subclinical paranoid ideation is negatively related to behavioral sensitivity to observations carrying meaningful information about the task structure. The findings provide direct evidence implicating dopamine in model-based belief updating in humans and have implications for understating the pathophysiology of psychotic disorders where dopamine function is disrupted.

Cross-decoding supramodal information in the human brain

Perceptual decision making is the cognitive process wherein the brain classifies stimuli into abstract categories for more efficient downstream processing. A system that, during categorization, can process information regardless of the information's original sensory modality (i.e., a supramodal system) would have a substantial advantage over a system with dedicated processes for specific sensory modalities. While many studies have probed decision processes through the lens of one sensory modality, it remains unclear whether there are such supramodal brain areas that can flexibly process task-relevant information regardless of the original "format" of the information. To investigate supramodality, one must ensure that supramodal information exists somewhere within the functional architecture by rendering information from multiple sensory systems necessary but insufficient for categorization. To this aim, we tasked participants with categorizing auditory and tactile frequency-modulated sweeps according to learned, supramodal categories in a delayed match-to-category paradigm while we measured their blood-oxygen-level dependent signal with functional MRI. To detect supramodal information, we implemented a set of cross-modality pattern classification analyses, which demonstrated that the left caudate nucleus encodes category-level information but not stimulus-specific information (such as spatial directions and stimulus modalities), while the right inferior frontal gyrus, showing the opposite pattern, encodes stimulus-specific information but not category-level information. Given our paradigm, these results reveal abstract representations in the brain that are independent of motor, semantic, and sensory-specific processing, instead reflecting supramodal, categorical information, which points to the caudate nucleus as a locus of cognitive processes involved in complex behavior.

Modeling subjective relevance in schizophrenia and its relation to aberrant salience

In schizophrenia, increased aberrant salience to irrelevant events and reduced learning of relevant information may relate to an underlying deficit in relevance detection. So far, subjective estimates of relevance have not been probed in schizophrenia patients. The mechanisms underlying belief formation about relevance and their translation into decisions are unclear. Using novel computational methods, we investigated relevance detection during implicit learning in 42 schizophrenia patients and 42 healthy individuals. Participants underwent functional magnetic resonance imaging while detecting the outcomes in a learning task. These were preceded by cues differing in color and shape, which were either relevant or irrelevant for outcome prediction. We provided a novel definition of relevance based on Bayesian precision and modeled reaction times as a function of relevance weighted unsigned prediction errors (UPE). For aberrant salience, we assessed responses to subjectively irrelevant cue manifestations. Participants learned the contingencies and slowed down their responses following unexpected events. Model selection revealed that individuals inferred the relevance of cue features and used it for behavioral adaption to the relevant cue feature. Relevance weighted UPEs correlated with dorsal anterior cingulate cortex activation and hippocampus deactivation. In patients, the aberrant salience bias to subjectively task-irrelevant information was increased and correlated with decreased striatal UPE activation and increased negative symptoms. This study shows that relevance estimates based on Bayesian precision can be inferred from observed behavior. This underscores the importance of relevance detection as an underlying mechanism for behavioral adaptation in complex environments and enhances the understanding of aberrant salience in schizophrenia.

Schizophrenia patients display deficits in the appropriate attribution of meaningfulness to stimuli; such as aberrantly increased processing of irrelevant and insufficient processing of relevant information. We aimed to investigate the subjective nature of relevance detection and its deficit in schizophrenia and developed an implicit learning paradigm that allowed for parallel learning from relevant and irrelevant information. Based on the idea that subjective relevance might be captured by Bayesian precision we set up different computational models of how subjective relevance guides learning and behavioral adaptation. We found that subjects use Bayesian precision to estimate stimulus relevance in order to integrate multidimensional information and adapt more to the subjectively relevant stimuli. This relevance weighted adaptation correlated with brain activation within the salience network. Further, schizophrenia patients displayed an increased aberrant tendency to irrelevant events which related to decreased striatal coding of the relevant learning signal. To conclude, our findings demonstrate how individual beliefs about relevance can be inferred from computational models. Furthermore, we suggest that aberrant salience observed in patients with schizophrenia reflects an idiosyncratic bias in states of high subjective uncertainty.