How learning enhances neural representations for behaviorally relevant stimuli via activity changes of cortical cell types remains unclear. We simultaneously imaged responses of pyramidal cells (PYR) along with parvalbumin (PV), somatostatin (SOM), and vasoactive intestinal peptide (VIP) inhibitory interneurons in primary visual cortex while mice learned to discriminate visual patterns. Learning increased selectivity for task-relevant stimuli of PYR, PV and SOM subsets but not VIP cells. Strikingly, PV neurons became as selective as PYR cells, and their functional interactions reorganized, leading to the emergence of stimulus-selective PYR-PV ensembles. Conversely, SOM activity became strongly decorrelated from the network, and PYR-SOM coupling before learning predicted selectivity increases in individual PYR cells. Thus, learning differentially shapes the activity and interactions of multiple cell classes: while SOM inhibition may gate selectivity changes, PV interneurons become recruited into stimulus-specific ensembles and provide more selective inhibition as the network becomes better at discriminating behaviorally relevant stimuli.

Inference of action potentials (‘spikes’) from neuronal calcium signals is complicated by the scarcity of simultaneous measurements of action potentials and calcium signals (‘ground truth’). We compiled a large, diverse ground-truth database from publicly available and newly performed recordings in zebrafish and mice covering a broad range of calcium indicators, cell types, and signal-to-noise ratios, comprising a total of >35 recording hours from 298 neurons. We developed an algorithm for spike inference (CASCADE) that is based on supervised deep networks, takes advantage of the ground-truth database, infers absolute spike rates, and outperforms existing model-based algorithms. To optimize performance for unseen imaging data, CASCADE retrains itself by resampling ground-truth data to match the respective sampling rate and noise level; therefore, no parameters need to be adjusted by the user. In addition, we developed systematic performance assessments for unseen data, openly release a resource toolbox, and provide a user-friendly cloud-based implementation.

Lead contact: t.mrsic-flogel@ucl.ac.uk 7 8In motor neocortex, preparatory activity predictive of specific movements is maintained by a positive 9 feedback loop with the thalamus. Motor thalamus receives excitatory input from the cerebellum, 10 which learns to generate predictive signals for motor control. The contribution of this pathway to 11 neocortical preparatory signals remains poorly understood. Here we show that in a virtual reality 12 conditioning task, cerebellar output neurons in the dentate nucleus exhibit preparatory activity 13 similar to that in anterolateral motor cortex prior to reward acquisition. Silencing activity in dentate 14 nucleus by photoactivating inhibitory Purkinje cells in the cerebellar cortex caused robust, short-15 latency suppression of preparatory activity in anterolateral motor cortex. Our results suggest that 16 preparatory activity is controlled by a learned decrease of Purkinje cell firing in advance of reward 17 under supervision of climbing fibre inputs signalling reward delivery. Thus, cerebellar computations 18 exert a powerful influence on preparatory activity in motor neocortex. 19 20 21 to upcoming movements or salient events such as reward (Giovannucci et al., 2017;Huang et al., 2013; 48 Kennedy et al., 2014; Wagner et al., 2017). For instance, DN neurons exhibit ramping activity predictive 49 of the timing and direction of the self-initiated saccades (Ashmore and Sommer, 2013; Ohmae et al., 50 2017). Moreover, inactivation of IPN activity reduces persistent activity in a region of medial prefrontal 51 cortex involved in trace eyeblink conditioning (Siegel and Mauk, 2013). Finally, a recent study has 52 established the existence of a loop between ALM and the cerebellum necessary for the maintenance of 53 preparatory activity (Gao et al., 2018). These results suggest that the cerebellum participates in 54 programming future actions, but the details of how it may contribute to preparatory activity in the 55 neocortex during goal-directed behaviour remain to be determined. 56 57 RESULTS 58 59 Preparatory activity in ALM prior to reward acquisition in a virtual corridor 60We developed a visuomotor task in which mice ran through a virtual corridor comprising salient visual 61 cues to reach a defined location where a reward was delivered (80 cm from the appearance of the 62 second checkerboard pattern; Figure 1A; see Methods). Within a week of training, mice learned to 63 128The cerebellar dentate nucleus exhibits preparatory activity 129Since the DN sends excitatory projections to the motor thalamus (Guo et al., 2017;Ichinohe et al., 2000; 130 Middleton and Strick, 1997; Thach and Jones, 1979), which has been shown to participate in the 131 maintenance of preparatory activity in mouse ALM neocortex (Guo et al., 2017), we investigated 132 whether DN activity could influence ALM processing. We first recorded the activity of DN neurons to 133 determine how their activity was modulated during the task (Figure 2A). Forty four percent of all 134 recorded DN neurons (n =...

Excitatory synaptic activity in the brain is shaped and balanced by inhibition. Because inhibition cannot propagate, it is often recruited with a synaptic delay by incoming excitation. Cerebellar Purkinje cells are driven by long-range excitatory parallel fiber inputs, which also recruit local inhibitory basket cells. The axon initial segment of each Purkinje cell is ensheathed by basket cell axons in a structure called the pinceau, which is largely devoid of chemical synapses. In mice, we found at the single-cell level that the pinceau mediates ephaptic inhibition of Purkinje cell firing at the site of spike initiation. The reduction of firing rate was synchronous with the presynaptic action potential, eliminating a synaptic delay and allowing granule cells to inhibit Purkinje cells without a preceding phase of excitation. Axon-axon ephaptic intercellular signaling can therefore mediate near-instantaneous feedforward and lateral inhibition.

Summary In motor neocortex, preparatory activity predictive of specific movements is maintained by a positive feedback loop with the thalamus. Motor thalamus receives excitatory input from the cerebellum, which learns to generate predictive signals for motor control. The contribution of this pathway to neocortical preparatory signals remains poorly understood. Here, we show that, in a virtual reality conditioning task, cerebellar output neurons in the dentate nucleus exhibit preparatory activity similar to that in anterolateral motor cortex prior to reward acquisition. Silencing activity in dentate nucleus by photoactivating inhibitory Purkinje cells in the cerebellar cortex caused robust, short-latency suppression of preparatory activity in anterolateral motor cortex. Our results suggest that preparatory activity is controlled by a learned decrease of Purkinje cell firing in advance of reward under supervision of climbing fiber inputs signaling reward delivery. Thus, cerebellar computations exert a powerful influence on preparatory activity in motor neocortex.

Visual intracortical and transthalamic pathways carry distinct information to cortical areasHighlights d Transthalamic pathway through pulvinar indirectly connects lower to higher cortical areas d This pathway combines input from V1 with that of many cortical and subcortical areas d Pulvinar conveys distinct visual and motor information to different higher visual areas d Direct intracortical and transthalamic pathways convey different information

The cerebellum aids the learning of fast, coordinated movements. According to current consensus, erroneously active parallel fibre synapses are depressed by complex spikes signalling movement errors. However, this theory cannot solve the credit assignment problem of processing a global movement evaluation into multiple cell-specific error signals. We identify a possible implementation of an algorithm solving this problem, whereby spontaneous complex spikes perturb ongoing movements, create eligibility traces and signal error changes guiding plasticity. Error changes are extracted by adaptively cancelling the average error. This framework, stochastic gradient descent with estimated global errors (SGDEGE), predicts synaptic plasticity rules that apparently contradict the current consensus but were supported by plasticity experiments in slices from mice under conditions designed to be physiological, highlighting the sensitivity of plasticity studies to experimental conditions. We analyse the algorithm’s convergence and capacity. Finally, we suggest SGDEGE may also operate in the basal ganglia.

Learning and attention increase visual response selectivity through distinct mechanismsHighlights d Learning and attention both enhance sensory processing at different timescales d The effects of learning and attention are uncorrelated at the single-cell level d Learning is driven by response suppression, attention by enhancement and suppression d Model with cell class-specific top-down inputs can account for effects of attention