The neural basis of swap errors in working memory.

When making decisions in a cluttered world, humans and other animals often have to hold multiple items in memory at once - such as the different items on a shopping list. Psychophysical experiments in humans and other animals have shown remembered stimuli can sometimes become confused, with participants reporting chimeric stimuli composed of features from different stimuli. In particular, subjects will often make "swap errors" where they misattribute a feature from one object as belonging to another object. While swap errors have been described behaviorally, their neural mechanisms are unknown. Here, we elucidate these neural mechanisms through trial-by-trial analysis of neural population recordings from posterior and frontal brain regions while monkeys perform two multi-stimulus working memory tasks. In these tasks, monkeys were cued to report the color of an item that either was previously shown at a corresponding location (requiring selection from working memory) or will be shown at the corresponding location (requiring attention to a position). Animals made swap errors in both tasks. In the neural data, we find evidence that the neural correlates of swap errors emerged when correctly remembered information is selected incorrectly from working memory. This led to a representation of the distractor color as if it were the target color, underlying the eventual swap error. We did not find consistent evidence that swap errors arose from misinterpretation of the cue or errors during encoding or storage in working memory. These results suggest an alternative to established views on the neural origins of swap errors, and highlight selection from and manipulation in working memory as crucial - yet surprisingly brittle - neural processes.

Semi-orthogonal subspaces for value mediate a tradeoff between binding and generalization.

When choosing between options, we must associate their values with the action needed to select them. We hypothesize that the brain solves this binding problem through neural population subspaces. To test this hypothesis, we examined neuronal responses in five reward-sensitive regions in macaques performing a risky choice task with sequential offers. Surprisingly, in all areas, the neural population encoded the values of offers presented on the left and right in distinct subspaces. We show that the encoding we observe is sufficient to bind the values of the offers to their respective positions in space while preserving abstract value information, which may be important for rapid learning and generalization to novel contexts. Moreover, after both offers have been presented, all areas encode the value of the first and second offers in orthogonal subspaces. In this case as well, the orthogonalization provides binding. Our binding-by-subspace hypothesis makes two novel predictions borne out by the data. First, behavioral errors should correlate with putative spatial (but not temporal) misbinding in the neural representation. Second, the specific representational geometry that we observe across animals also indicates that behavioral errors should increase when offers have low or high values, compared to when they have medium values, even when controlling for value difference. Together, these results support the idea that the brain makes use of semi-orthogonal subspaces to bind features together.

Redundant representations are required to disambiguate simultaneously presented complex stimuli.

A pedestrian crossing a street during rush hour often looks and listens for potential danger. When they hear several different horns, they localize the cars that are honking and decide whether or not they need to modify their motor plan. How does the pedestrian use this auditory information to pick out the corresponding cars in visual space? The integration of distributed representations like these is called the assignment problem, and it must be solved to integrate distinct representations across but also within sensory modalities. Here, we identify and analyze a solution to the assignment problem: the representation of one or more common stimulus features in pairs of relevant brain regions-for example, estimates of the spatial position of cars are represented in both the visual and auditory systems. We characterize how the reliability of this solution depends on different features of the stimulus set (e.g., the size of the set and the complexity of the stimuli) and the details of the split representations (e.g., the precision of each stimulus representation and the amount of overlapping information). Next, we implement this solution in a biologically plausible receptive field code and show how constraints on the number of neurons and spikes used by the code force the brain to navigate a tradeoff between local and catastrophic errors. We show that, when many spikes and neurons are available, representing stimuli from a single sensory modality can be done more reliably across multiple brain regions, despite the risk of assignment errors. Finally, we show that a feedforward neural network can learn the optimal solution to the assignment problem, even when it receives inputs in two distinct representational formats. We also discuss relevant results on assignment errors from the human working memory literature and show that several key predictions of our theory already have support.

Subspace orthogonalization as a mechanism for binding values to space.

When choosing between options, we must solve an important binding problem. The values of the options must be associated with information about the action needed to select them. We hypothesize that the brain solves this binding problem through use of distinct population subspaces. To test this hypothesis, we examined the responses of single neurons in five reward-sensitive regions in rhesus macaques performing a risky choice task. In all areas, neurons encoded the value of the offers presented on both the left and the right side of the display in semi-orthogonal subspaces, which served to bind the values of the two offers to their positions in space. Supporting the idea that this orthogonalization is functionally meaningful, we observed a session-to-session covariation between choice behavior and the orthogonalization of the two value subspaces: trials with less orthogonalized subspaces were associated with greater likelihood of choosing the less valued option. Further inspection revealed that these semi-orthogonal subspaces arose from a combination of linear and nonlinear mixed selectivity in the neural population. We show this combination of selectivity balances reliable binding with an ability to generalize value across different spatial locations. These results support the hypothesis that semi-orthogonal subspaces support reliable binding, which is essential to flexible behavior in the face of multiple options.

Abstract representations emerge naturally in neural networks trained to perform multiple tasks.

Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability to generalize arises from a specific representational geometry, that we call abstract and that is referred to as disentangled in machine learning. These abstract representations have been observed in recent neurophysiological studies. However, it is unknown how they emerge. Here, using feedforward neural networks, we demonstrate that the learning of multiple tasks causes abstract representations to emerge, using both supervised and reinforcement learning. We show that these abstract representations enable few-sample learning and reliable generalization on novel tasks. We conclude that abstract representations of sensory and cognitive variables may emerge from the multiple behaviors that animals exhibit in the natural world, and, as a consequence, could be pervasive in high-level brain regions. We also make several specific predictions about which variables will be represented abstractly.

Primate superior colliculus is engaged in abstract higher-order cognition.

Categorization is a fundamental cognitive process by which the brain assigns stimuli to behaviorally meaningful groups. Investigations of visual categorization in primates have identified a hierarchy of cortical areas that are involved in the transformation of sensory information into abstract category representations. However, categorization behaviors are ubiquitous across diverse animal species, even those without a neocortex, motivating the possibility that subcortical regions may contribute to abstract cognition in primates. One candidate structure is the superior colliculus (SC), an evolutionarily conserved midbrain region that, although traditionally thought to mediate only reflexive spatial orienting, is involved in cognitive tasks that require spatial orienting. Here, we reveal a novel role of the primate SC in abstract, higher-order visual cognition. We compared neural activity in the SC and the posterior parietal cortex (PPC), a region previously shown to causally contribute to category decisions, while monkeys performed a visual categorization task in which they report their decisions with a hand movement. The SC exhibits stronger and shorter-latency category encoding than the PPC, and inactivation of the SC markedly impairs monkeys' category decisions. These results extend SC's established role in spatial orienting to abstract, non-spatial cognition.

Nonlinear mixed selectivity supports reliable neural computation.

Neuronal activity in the brain is variable, yet both perception and behavior are generally reliable. How does the brain achieve this? Here, we show that the conjunctive coding of multiple stimulus features, commonly known as nonlinear mixed selectivity, may be used by the brain to support reliable information transmission using unreliable neurons. Nonlinearly mixed feature representations have been observed throughout primary sensory, decision-making, and motor brain areas. In these areas, different features are almost always nonlinearly mixed to some degree, rather than represented separately or with only additive (linear) mixing, which we refer to as pure selectivity. Mixed selectivity has been previously shown to support flexible linear decoding for complex behavioral tasks. Here, we show that it has another important benefit: in many cases, it makes orders of magnitude fewer decoding errors than pure selectivity even when both forms of selectivity use the same number of spikes. This benefit holds for sensory, motor, and more abstract, cognitive representations. Further, we show experimental evidence that mixed selectivity exists in the brain even when it does not enable behaviorally useful linear decoding. This suggests that nonlinear mixed selectivity may be a general coding scheme exploited by the brain for reliable and efficient neural computation.