Highlights d Random recurrent connections can support flexible working memory d Overlap of connections causes interference between memories, limiting capacity d Model captures many behavioral and physiological characteristics of working memory d Structured sensory networks can constrain high-dimensional random representations

To adapt to a changing world, we must be able to switch between rules already learned and, at other times, learn rules anew. Often we must do both at the same time, switching between known rules while also constantly re-estimating them. Here, we show these two processes, rule switching and rule learning, rely on distinct but intertwined computations, namely fast inference and slower incremental learning. To this end, we studied how monkeys switched between three rules. Each rule was compositional, requiring the animal to discriminate one of two features of a stimulus and then respond with an associated eye movement along one of two different response axes. By modeling behavior we found the animals learned the axis of response using fast inference (rule switching) while continuously re-estimating the stimulus-response associations within an axis (rule learning). Our results shed light on the computational interactions between rule switching and rule learning, and make testable neural predictions for these interactions.

Depending on environmental demands, humans can learn and exploit multiple concurrent sets of stimulus-response associations. Mechanisms underlying the learning of such task-sets remain unknown. Here we investigate the hypothesis that task-set learning relies on unsupervised chunking of stimulus-response associations that occur in temporal proximity. We examine behavioral and neural data from a task-set learning experiment using a network model. We first show that task-set learning can be achieved provided the timescale of chunking is slower than the timescale of stimulus-response learning. Fitting the model to behavioral data on a subject-by-subject basis confirmed this expectation and led to specific predictions linking chunking and task-set retrieval that were borne out by behavioral performance and reaction times. Comparing the model activity with BOLD signal allowed us to identify neural correlates of task-set retrieval in a functional network involving ventral and dorsal prefrontal cortex, with the dorsal system preferentially engaged when retrievals are used to improve performance.

Depending on environmental demands, humans can learn and exploit multiple concurrent sets of stimulus-response associations. Mechanisms underlying the learning of such task-sets remain unknown. Here we investigate the hypothesis that task-set learning relies on unsupervised chunking of stimulus-response associations that occur in temporal proximity. We examine behavioral and neural data from a task-set learning experiment using a network model. We first show that task-set learning can be achieved provided the timescale of chunking is slower than the timescale of stimulus-response learning. Fitting the model to behavioral data confirmed this expectation and led to specific predictions linking chunking and task-set retrieval that were borne out by behavioral performance and reaction times. Comparing the model activity with BOLD signal allowed us to identify neural correlates of task-set retrieval in a functional network involving ventral and dorsal prefrontal cortex, with the dorsal system preferentially engaged when retrievals are used to improve performance.Keywords task-set learning, Hebbian plasticity, cognitive control, mixed selectivity, computational model. 8 Rioult-Pedotti et al., 2000;Rumpel et al., 2005;Schultz and Dickinson, 2000;Xiong et al., 2015]. Understanding how synaptic changes at individual synapses are related to learn-10 ing in behaving animals remains however a daunting challenge, especially for complex cognitive tasks that go beyond simple stimulus-response associations. 12Depending on environmental demands, humans engaged in a given task are capable of learning and exploiting multiple concurrent strategies. For instance, in the classical 14 Stroop task [MacLeod, 1991;Stroop, 1935], an identical stimulus like a colored word leads 2 to di↵erent responses depending on whether the current requirement is to read the word 16 or identify its color. Human subjects are able to learn to flexibly switch between these two di↵erent stimulus-response associations, often called task-sets [Sakai, 2008]. Studies of 18 task-sets learning predominantly rely on abstract models that describe behavioral learning without any physiological constraint [Botvinick et al., 2009;Collins and Koechlin, 2012; 20 Collins and Frank, 2013;Daw et al., 2005Daw et al., , 2011Dayan and Daw, 2008;Franklin and Frank, 2018;Russek et al., 2017]. While these models are able to capture computational 22 aspects of behavior, and correlate them with physiological measurements [Daw et al., 2011;Donoso et al., 2014;Koechlin and Hyafil, 2007;Niv, 2009;Wilson et al., 2014], 24 understanding the underlying biophysical mechanisms is an open issue.One hypothesis [Rigotti et al., 2010b] states that learning of task-sets, and more gener-26 ally rule-based behavior, relies on unsupervised learning of temporal contiguity between events. Events that occur repeatedly after each other are automatically associated as 28 demonstrated in classical conditioning experiments [Hawkins et al., 1983;Kahana, 1996;Rescorla, 1988;Rescorla et al., 1972;Sakai...

Working memory is fundamental to cognition, allowing one to hold information 'in mind' and use it to guide behavior. A defining characteristic of working memory is its flexibility: we can hold anything in mind. However, typical models of working memory rely on finely tuned, content-specific, attractors to persistently maintain neural activity and therefore do not allow for the flexibility observed in behavior. Here we present a flexible model of working memory that maintains representations through random recurrent connections between two layers of neurons: a structured 'sensory' layer and a randomly connected, unstructured, layer. As the interactions are untuned with respect to the content being stored, the network is able to maintain any arbitrary input. However, this flexibility comes at a cost: the random connections overlap, leading to interference between representations and limiting the memory capacity of the network. Additionally, our model captures several other key behavioral and neurophysiological characteristics of working memory.