Human learning in complex environments critically dependson the ability to perform model selection, that is to assess com-peting hypotheses about the structure of the environment. Im-portantly, information is accumulated continuously, which ne-cessitates an online process for model selection. While modelselection in human learning has been explored extensively, it isunclear how memory systems support learning in an online set-ting. We formulate a semantic learner and demonstrate that on-line learning on open model spaces results in a delicate choicebetween either tracking a possibly infinite number of compet-ing models or retaining experiences in an intact form. Sincenone of these choices is feasible for a bounded-resource mem-ory system, we propose an episodic learner that retains an op-timised subset of experiences in addition to semantic memory.On a simple model system we demonstrate that this norma-tive theory of episodic memory can effectively circumvent thechallenge of online model selection.

Recent research has proposed that systematic biases in human memory -- while seemingly highlighting a proclivity for failure -- can be understood as hallmarks of optimised lossy compression. Specifically, a form of compression termed semantic compression whereby an internal model of the environment is recruited to encode memories. Semantic compression casts memory errors in the normative framework of information theory, describing how limited memory resources should be distributed to optimise recall performance. Notably, the theory does not define a single best compression, rather a continuum of trade-offs between utilised capacity and expected distortion is possible. However, possible consequences of this characteristic feature have not been tested explicitly. Here we test the idea that gradual degradation of memories with time corresponds to a decrease in the amount of resources allocated to store memories. We apply the general framework to remembering synthetic words in a delayed recognition experiment and find that subjects are indeed less sensitive to intrusions generated by our model than generic distortions, and that delay length modulates recall rates in line with the predictions of the theory.

Effective task execution requires the representation of multiple task-related variables that determine how stimuli lead to correct responses. Even the primary visual cortex (V1) represents other task-related variables such as expectations, choice, and context. However, it is unclear how V1 can flexibly accommodate these variables without interfering with visual representations. We trained mice on a context-switching cross-modal decision task, where performance depends on inferring task context. We found that the context signal that emerged in V1 was behaviorally relevant as it strongly covaried with performance, independent from movement. Importantly, this signal was integrated into V1 representation by multiplexing visual and context signals into orthogonal subspaces. In addition, auditory and choice signals were also multiplexed as these signals were orthogonal to the context representation. Thus, multiplexing allows V1 to integrate visual inputs with other sensory modalities and cognitive variables to avoid interference with the visual representation while ensuring the maintenance of task-relevant variables.

Attention supports decision making by selecting the features that are relevant for decisions. Selective enhancement of the relevant features and inhibition of distractors has been proposed as potential neural mechanisms driving this selection process. Yet, how attention operates when relevance cannot be directly determined, and the attention signal needs to be internally constructed is less understood. Here we recorded from populations of neurons in the anterior cingulate cortex (ACC) of mice in an attention-shifting task where relevance of stimulus modalities changed across blocks of trials. In contrast with V1 recordings, decoding of the irrelevant modality gradually declined in ACC after an initial transient. Our analytical proof and a recurrent neural network model of the task revealed mutually inhibiting connections that produced context-gated suppression as observed in mice. Using this RNN model we predicted a correlation between contextual modulation of individual neurons and their stimulus drive, which we confirmed in ACC but not in V1.