Asymmetric and adaptive reward coding via normalized reinforcement learning

Learning is widely modeled in psychology, neuroscience, and computer science by prediction error-guided reinforcement learning (RL) algorithms. While standard RL assumes linear reward functions, reward-related neural activity is a saturating, nonlinear function of reward; however, the computational and behavioral implications of nonlinear RL are unknown. Here, we show that nonlinear RL incorporating the canonical divisive normalization computation introduces an intrinsic and tunable asymmetry in prediction error coding. At the behavioral level, this asymmetry explains empirical variability in risk preferences typically attributed to asymmetric learning rates. At the neural level, diversity in asymmetries provides a computational mechanism for recently proposed theories of distributional RL, allowing the brain to learn the full probability distribution of future rewards. This behavioral and computational flexibility argues for an incorporation of biologically valid value functions in computational models of learning and decision-making.

Asymmetric Weights and Retrieval Practice in an Autoassociative Neural Network Model of Paired-Associate Learning

Rizzuto and Kahana (2001) applied an autoassociative Hopfield network to a paired-associate word learning experiment in which (1) participants studied word pairs (e.g., ABSENCE-HOLLOW), (2) were tested in one direction (ABSENCE-?) on a first test, and (3) were tested in the same direction again or in the reverse direction (?-HOLLOW) on a second test. The model contained a correlation parameter to capture the dependence between forward versus backward learning between the two words of a word pair, revealing correlation values close to 1.0 for all participants, consistent with neural network models that use the same weight for communication in both directions between nodes. We addressed several limitations of the model simulations and proposed two new models incorporating retrieval practice learning (e.g., the effect of the first test on the second) that fit the accuracy data more effectively, revealing substantially lower correlation values (average of .45 across participants, with zero correlation for some participants). In addition, we analyzed recall latencies, finding that second test recall was faster in the same direction after a correct first test. Only a model with stochastic retrieval practice learning predicted this effect. In conclusion, recall accuracy and recall latency suggest asymmetric learning, particularly in light of retrieval practice effects.

Asymmetric learning facilitates human inference of transitive relations

Humans and other animals are capable of inferring never-experienced relations (e.g., A>C) from other relational observations (e.g., A>B and B>C). The processes behind such transitive inference are subject to intense research. Here, we demonstrate a new aspect of relational learning, building on previous evidence that transitive inference can be accomplished through simple reinforcement learning mechanisms. We show in simulations that inference of novel relations benefits from an asymmetric learning policy, where observers update only their belief about the winner (or loser) in a pair. Across 4 experiments (n=145), we find substantial empirical support for such asymmetries in inferential learning. The learning policy favoured by our simulations and experiments gives rise to a compression of values which is routinely observed in psychophysics and behavioural economics. In other words, a seemingly biased learning strategy that yields well-known cognitive distortions can be beneficial for transitive inferential judgments.

Confirmation bias optimizes reward learning

ABSTRACTConfirmation bias—the tendency to overweight information that matches prior beliefs or choices—has been shown to manifest even in simple reinforcement learning. In line with recent work, we find that participants learned significantly more from choice-confirming outcomes in a reward-learning task. What is less clear is whether asymmetric learning rates somehow benefit the learner. Here, we combine data from human participants and artificial agents to examine how confirmation-biased learning might improve performance by counteracting decisional and environmental noise. We evaluate one potential mechanism for such noise reduction: visual attention—a demonstrated driver of both value-based choice and predictive learning. Surprisingly, visual attention showed the opposite pattern to confirmation bias, as participants were most likely to fixate on “missed opportunities”, slightly dampening the effects of the confirmation bias we observed. Several million simulated experiments with artificial agents showed this bias to be a reward-maximizing strategy compared to several alternatives, but only if disconfirming feedback is not completely ignored—a condition that visual attention may help to enforce.

Dissociation between asymmetric value updating and perseverance in human reinforcement learning

The learning rate is a key parameter in reinforcement learning that determines the extent to which novel information (outcome) is incorporated in guiding subsequent actions. Numerous studies have reported that the magnitude of the learning rate in human reinforcement learning is biased depending on the sign of the reward prediction error. However, this asymmetry can be observed as a statistical bias if the fitted model ignores the choice autocorrelation (perseverance), which is independent of the outcomes. Therefore, to investigate the genuine process underlying human choice behavior using empirical data, one should dissociate asymmetry in learning and perseverance from choice behavior. The present study addresses this issue by using a Hybrid model incorporating asymmetric learning rates and perseverance. First, by conducting simulations, we demonstrate that the Hybrid model can identify the true underlying process. Second, using the Hybrid model, we show that empirical data collected from a web-based experiment are governed by perseverance rather than asymmetric learning. Finally, we apply the Hybrid model to two open datasets in which asymmetric learning was reported. As a result, the asymmetric learning rate was validated in one dataset but not another.

SSRIs modulate asymmetric learning from reward and punishment

Human instrumental learning is driven by a history of outcome success and failure. We demonstrate that week-long treatment with a serotonergic antidepressant modulates a valence-dependent asymmetry in learning from reinforcement. In particular, we show that prolonged boosting of central serotonin reduces reward learning, and enhances punishment learning. This treatment induced learning asymmetry can result in lowered positive and enhanced negative expectations. A consequential effect is more rewarding, and less disappointing, experiences and this may, in part, explain the slow temporal evolution of serotonin’s well-established antidepressant effects.