Neurons distribute synaptic inputs across their dendritic tree. In layer 2/3 pyramidal cells of primary visual cortex, spines on distal dendrites share somatic orientation preference but have receptive fields displaced in retinotopic space, which supports tuning to visual edges. However, it is not known how synaptic plasticity rules can lead to specialization of tuning properties across dendritic compartments. We demonstrate an experimentally grounded model of compartment-specific spike-timing dependent plasticity (STDP) that accounts for the enrichment of retinotopically-displaced inputs on distal branches. Our previous experimental work revealed compartment-specific calcium signals that predict reduced STDP-mediated depression but preserved potentiation. Based on these findings, we built an STDP model with compartment-specific properties, in which some distal branches are relatively resistant to STDPmediated depression. Synapses on these branches are more likely to stabilize inputs with weaker correlations to postsynaptic spiking. Using a visual input model, we show that compartment-specific reduction in STDP-mediated depression recapitulates in vivo experimental measurements of spine tuning. Furthermore, our experimental results show that reduced STDP-mediated depression is restricted to distal dendritic compartments with complex branching structure and not observed in other distal branches. Therefore, our model makes an untested prediction that complex branches will be hotspots for retinotopically-displaced inputs.

Memories are thought to be encoded in synaptic connections between assemblies of neurons that are reactivated during memory recall. However, in light of ongoing molecular turnover and synaptic decay, this widely accepted view cannot explain how individual ensembles are maintained over (life-)long timescales. Experimentally, learning has not only been associated with synaptic modifications among neurons, but also with epigenetic alterations of learning-related gene transcription within neurons. Although these epigenetic changes are involved in all stages of memory dynamics, they have been largely omitted in computational studies. In this update, we advocate for the integration of epigenetic mechanisms into computational models of memory. Using a recurrent neural network model that includes epigenetic plasticity as a variable, we explore the role of epigenetic priming in the maintenance of memories across long timescales; we then investigate the implications of epigenetic modifications for memory allocation and for reversing cognitive decline associated with neurodegeneration; and finally, we predict several computational advantages of including epigenetics over traditional models of synaptic memory. Overall, this paper stands as a first step towards the integration of epigenetics in computational models of memory and corroborates the experimentally derived notion that memory might not be encoded solely in synaptic weights, but rather co-encoded in epigenetic patterns within the nucleus.

Backpropagation (BP) is widely viewed as biologically implausible, in part because it requires feedback weights to be the transpose of forward weights for error propagation. Interestingly, when training a network with fixed random feedback weights to circumvent this issue, learning aligns the forward weights with the feedback weights, leading the backpropagated error signal to become an approximation of the standard gradient used by BP. This process, called Feedback Alignment (FA), occurs in MLPs and very shallow CNNs but does not scale well to deeper architectures. In this work, we first investigated differences between BP and FA models, trained on CIFAR10, specifically focusing on the effective rank of the signal. We found that the FA error has a considerably lower rank and hence is constrained to a lower-dimensional subspace compared to BP, limiting exploration of the parameter space. Motivated by this observation, we evaluated two mechanisms for increasing the effective dimensionality of FA: Muon, an optimiser that orthogonalises weight updates; and hidden activity normalisation, which promotes activation orthogonality. Across larger architectures and benchmarks, we find that these methods consistently improve over FA baselines, for example, on CIFAR100 with a Resnet-18, accuracy increases by 9 percentage points. Our results identify low-dimensional gradient dynamics as a key obstacle to scaling FA and suggest that inducing higher-dimensional update geometry is a promising route toward scaling alternatives to backpropagation.

The ability to reconstruct images represented by the brain has the potential to give us an intuitive understanding of what the brain sees. Reconstruction of visual input from human fMRI data has garnered significant attention in recent years. Comparatively less focus has been directed towards vision reconstruction from single-cell recordings, despite its potential to provide a more direct measure of the information represented by the brain. Here, we achieve high-quality reconstructions of natural movies presented to mice, from the activity of neurons in their visual cortex for the first time. Using our method of video optimization via backpropagation through a state-of-the-art dynamic neural encoding model, we reliably reconstruct 10 s movies at 30 Hz from two-photon calcium imaging data. We achieve a pixel-level correlation of 0.57 between ground-truth movies and single-trial reconstructions. Previous reconstructions based on awake mouse V1 neuronal responses to static images achieved a pixel-level correlation of 0.24 over a similar retinotopic area. We find that critical for high-quality reconstructions are the number of neurons in the dataset and the use of model ensembling. This paves the way for movie reconstruction to be used as a tool to investigate a variety of visual processing phenomena.

The cerebellum and cerebral cortex form tightly coupled circuits thought to support flexible and efficient temporal processing. How this interaction shapes cortical learning dynamics, and whether such heterogeneous modularity can benefit artificial systems, remains unclear. Here, we augment a recurrent neural network (RNN) with a cerebellar-inspired feedforward module and evaluate the resulting architecture on temporal tasks of varying difficulty. The cortico-cerebellar RNN (CB-RNN) learns faster and reaches higher maximum performance than parameter-matched fully recurrent baselines across a variety of regimes. Crucially, freezing the recurrent core after minimal training and delegating subsequent learning to the cerebellar module preserves superior learning efficiency, suggesting the cerebellar module is a primary driver of efficiency and that the cortical network can largely function as a fixed reservoir. Our results suggest that heterogeneous modular architectures can act as a powerful structural inductive bias in neural systems.

Abstract Spontaneous neural activity reflects sensory experience yet occupies a lower-dimensional subspace than evoked responses. However, the principle governing which components of experience are incorporated into this intrinsic activity has remained unknown. Here, we propose that temporal predictability is one of the organizing principles that selects which evoked dimensions enter spontaneous neural manifolds. We show that local predictive synaptic plasticity provides a circuit mechanism that selectively embeds components of evoked activity that are predictable on the intrinsic timescale of recurrent dynamics, while excluding unpredictable fluctuations. As a consequence, the dimensionality of spontaneous activity is not fixed but depends on environmental timescales: rapidly fluctuating inputs are excluded, whereas slowly varying components are retained as contextual dimensions. This framework reproduces on- and off-manifold coding observed in visual cortex, and reconciles apparently conflicting developmental observations that spontaneous and evoked activity become both more similar and more geometrically distinct over maturation. Together, these results identify temporal predictability as a key principle linking environmental statistics to the geometry of intrinsic population dynamics.

Learning new tasks requires the brain to reshape the flow of neural activity, but how these changes arise from dynamic neural connectivity remains unclear. Here, we used two-photon photostimulation and calcium imaging to map learning-related changes in connectivity in layer 2/3 of mouse motor cortex, induced by learning of an optical brain-computer interface (BCI) task. Mice rapidly (within minutes) learned to change activity in a conditioned neuron to earn rewards. Activity changes were sparse; the conditioned neuron increased activity more than surrounding neurons. Mapping connectivity before and after learning revealed changes in motor cortex connectivity, enriched in neurons that were active before trial initiation, analogous to motor cortex populations that are active preceding movement. Motor cortex plasticity reroutes preparatory activity to neurons that are active later and control the conditioned neuron. Our findings show how rapid learning can be achieved through structured changes in motor cortex connectivity.

Abstract When an animal enters a new environment, neurons in the hippocampus begin to map out the space. They become selectively responsive to features like the animal’s location, the position of rewards, and the presence of stimuli relevant to navigating or performing a specific task. With experience, hippocampal neurons also develop behaviour-related biases. It is thus believed that the hippocampus encodes multi-sensory, behaviourally-relevant cognitive maps of environments that are critical to navigation, learning and guiding behaviour. The predictive learning hypothesis proposes that these complex maps emerge because a core goal of the brain is to learn to predict the features of its environment. In sensory cortex, predictive learning provides a compelling explanation of anticipatory and error-like sensory responses. Pyramidal neurons receive top-down and bottom-up inputs to their proximal basal and distal apical dendrites, respectively. These complementary inputs streams are thought to enable them to act as comparison units, signaling discrepancies between predictive and sensory inputs. In the hippocampus, however, the potential link between pyramidal neurons and predictive learning is still underexplored. Here, we investigate the possibility that pyramidal neurons perform a similar comparator function in the hippocampus. In our model, two-compartment pyramidal neurons receive sensory information about salient features of the environment at their distal apical dendrites which is compared to tuned spatial inputs received more proximally to their cell body. We demonstrate how predictive learning implemented in this circuit using behavioural timescale synaptic plasticity and distal apical inhibition can explain a variety of spatial and behaviourally-relevant features encoded in the hippocampus. We also lay out key predictions for validating our model experimentally. As such, our work helps bridge an important gap in the literature on predictive learning in the hippocampus and set the stage for more robust experimental validation of this prominent hypothesis.

Neurons have rich input-output functions for processing and combining their inputs. Although many experiments characterize these functions by directly activating synaptic inputs on dendrites in vitro, the integration of spatiotemporal inputs representing real-world stimuli is less well studied. Using ethologically relevant stimuli, we study neuronal integration in relation to Boolean AND and OR operations thought to be important for pattern recognition. We recorded single-unit responses in the mouse auditory cortex to pairs of ultrasonic mouse vocalization (USV) syllables. We observed a range of integration responses, spanning the sublinear to supralinear regimes, with many responses resembling the MAX-like function, an instantiation of the OR operation. Integration was more MAX-like for strongly activating features and more AND-like for spectrally distinct inputs. Importantly, single neurons could implement more than one integration function, in contrast to artificial networks, which typically fix activation functions across all units and inputs. To understand the mechanism underlying the flexibility and heterogeneity in neuronal integration, we modeled how dendritic properties could influence the integration of inputs with complex spectrotemporal structure. Our results link nonlinear integration in dendrites to single-neuron computations for pattern recognition. NEW & NOTEWORTHY Sensory neurons compute over their inputs, combining stimuli to achieve selectivity and invariance for pattern recognition. Using real-world stimuli, we show that cortical neurons are flexible, capable of implementing more than one computation. We investigate this flexibility by modeling how synaptic activation patterns of real-world stimuli affect dendritic integration and resultant neuronal computation. Our work bridges the gap between biophysical mechanisms and computation, linking neuronal input integration to pattern recognition.

Abstract Neural representations drift gradually over time even under stable environmental conditions, but the synaptic mechanisms driving this drift remain unclear. Here we show that representational drift can arise intrinsically from predictive synaptic plasticity in spiking excitatory-inhibitory networks. During repeated exposure to unchanged input patterns, individual neurons gradually changed their preferred pattern while ensemble-level coding remained stable. These changes in preference were preceded by a weakening of the net excitatory-inhibitory drive supporting the current preference relative to competing patterns, and this relative drive predicted changes on the following trial. Extending the model to hippocampal place coding reproduced experience-dependent tuning-curve drift in CA1, including the dissociation between elapsed time and intervening exposure. At the population level, drift was expressed as coordinated rotation and translation of neural state space. Thus, representational drift can emerge as an intrinsic consequence of predictive E/I plasticity that maintains balanced, selective population codes.

Across psychology, economics, and decision science, it is well established that choosing a good internal representation can make complex problems easier to solve [1][1],[2][2]. This idea, central to theories of bounded rationality [3][3] and heuristics [4][4],[5][5], suggests that intelligent be-havior depends not only on what is stored in working memory, but on the format in which it is stored. Whether non-human animals flexibly select internal formats, and how such choices are implemented in neural circuits, has remained unclear. Here we show that rats adjust the format of working memory to reduce cognitive load. When a task encourages an action-based code, rats maintain the relevant information as a stable motor plan supported by frontal cortex. When the same task is reformulated to make this code costly, they in-stead store a sensory trace in auditory cortex. These strategy shifts are rapid and reliable, revealing that rats can switch between distinct neural circuits to store information in the most economical internal format. These results demonstrate that flexible representational selection is not unique to humans [6][6] but is present in rodents and reflected in circuit-level reallocation of working memory. This establishes a neural basis for classic theories linking problem representation to computational efficiency and provides a path toward a neural circuit-level description of selecting internal formats to reduce cognitive costs. ### Competing Interest Statement The authors have declared no competing interest. Wellcome Trust, https://ror.org/029chgv08, 219627/Z/19/Z, 318818/Z/24/Z Gatsby Charitable Foundation, GAT3755, GAT4057 111 Project, Base B16018 NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai National Natural Science Foundation of China, https://ror.org/01h0zpd94, 31970962 Science and Technology Commission of Shanghai Municipality, 15JC1400104, 15XD1503000 [1]: #ref-1 [2]: #ref-2 [3]: #ref-3 [4]: #ref-4 [5]: #ref-5 [6]: #ref-6

Recurrent neural circuits often face inherent complexities in learning and generating their desired outputs, especially when they initially exhibit chaotic spontaneous activity. While the celebrated FORCE learning rule can train chaotic recurrent networks to produce coherent patterns by suppressing chaos, it requires non-local plasticity rules and quick plasticity, raising the question of how synapses adapt on local, biologically plausible timescales to handle potential chaotic dynamics. We propose a novel framework called "predictive alignment", which tames the chaotic recurrent dynamics to generate a variety of patterned activities via a biologically plausible plasticity rule. Unlike most recurrent learning rules, predictive alignment does not aim to directly minimize output error to train recurrent connections, but rather it tries to efficiently suppress chaos by aligning recurrent prediction with chaotic activity. We show that the proposed learning rule can perform supervised learning of multiple target signals, including complex low-dimensional attractors, delay matching tasks that require short-term temporal memory, and finally even dynamic movie clips with high-dimensional pixels. Our findings shed light on how predictions in recurrent circuits can support learning.

Behavioral Timescale Synaptic Plasticity (BTSP) is a form of synaptic plasticity in which dendritic Ca2+ plateau potentials in hippocampal pyramidal neurons drive rapid place field formation. Unlike traditional learning rules, BTSP learns correlations on the timescales of seconds and rapidly changes single-unit activity in only a few trials. To explore how BTSP-like learning can be integrated into network models, we propose a generalized BTSP rule (gBTSP), which we apply to unsupervised and supervised learning tasks, in both feedforward and recurrent networks. Unsupervised gBTSP mirrors classical frameworks of competitive learning, learning place field maps (in the feed-forward case), and attractive memory networks (in the recurrent case). For supervised learning, we show that plateau events can reduce task error, enabling gBTSP to solve tasks such as trajectory matching and delayed non-match-to-sample. However, we find that credit assignment via gBTSP becomes harder to achieve with increased network depth or CA3-like recurrence. This suggests that additional features may be needed to support BTSP-mediated few-shot learning of complex tasks in the hippocampus. ### Competing Interest Statement The authors have declared no competing interest. Biotechnology and Biological Sciences Research Council, https://ror.org/00cwqg982, BB/N013956/1 Wellcome Trust, https://ror.org/029chgv08, 200790/Z/16/Z Simons Foundation, 564408 Engineering and Physical Sciences Research Council, EP/R035806/1, EP/X029336/1 European Research Council, https://ror.org/0472cxd90, EP/Y027841/1

Animals use feedback to rapidly correct ongoing movements in the presence of a perturbation. Repeated exposure to a predictable perturbation leads to behavioural adaptation that compensates for its effects. Here, we tested the hypothesis that all the processes necessary for motor adaptation may emerge as properties of a controller that adaptively updates its policy. We trained a recurrent neural network to control its own output through an error-based feedback signal, which allowed it to rapidly counteract external perturbations. Implementing a biologically plausible plasticity rule based on this same feedback signal enabled the network to learn to compensate for persistent perturbations through a trial-by-trial process. The network activity changes during learning matched those from populations of neurons from monkey primary motor cortex - known to mediate both movement correction and motor adaptation - during the same task. Furthermore, our model natively reproduced several key aspects of behavioural studies in humans and monkeys. Thus, key features of trial-by-trial motor adaptation can arise from the internal properties of a recurrent neural circuit that adaptively controls its output based on ongoing feedback.

Neuroscience has inspired artificial intelligence (AI) for decades but, in recent years, AI tools have begun to revolutionize neuroscience research. The emerging field of NeuroAI has the potential to transform large-scale neural modelling and data-driven neuroscience discovery. The field must balance exploiting AI’s power while maintaining interpretability and biological insight.

The brain learns an internal model of the environment through sensory experiences, which is essential for high-level cognitive processes. Recent studies show that spontaneous activity reflects such a learned internal model. Although computational studies have proposed that Hebbian plasticity can learn the switching dynamics of replayed activities, it is still challenging to learn dynamic spontaneous activity that obeys the statistical properties of sensory experience. Here, we propose a pair of biologically plausible plasticity rules for excitatory and inhibitory synapses in a recurrent spiking neural network model to embed stochastic dynamics in spontaneous activity. The proposed synaptic plasticity rule for excitatory synapses seeks to minimize the discrepancy between stimulus-evoked and internally predicted activity, while inhibitory plasticity maintains the excitatory-inhibitory balance. We show that the spontaneous reactivation of cell assemblies follows the transition statistics of the model’s evoked dynamics. We also demonstrate that simulations of our model can replicate recent experimental results of spontaneous activity in songbirds, suggesting that the proposed plasticity rule might underlie the mechanism by which animals learn internal models of the environment.

Flexibility and stability of neuronal ensembles are crucial features of brain function. Little is known about how these properties of local circuits are influenced by long-range inputs. We show, in mice, that lateral entorhinal cortex glutamatergic (LECGLU) and γ-aminobutyric acid (GABA)-ergic (LECGABA) projections to CA3 recruit specific microcircuits that conjunctively provide stability to neuronal ensembles, thereby supporting learning. LECGLU drives excitation in CA3 but also substantial feedforward inhibition that prevents somatic and dendritic spikes. Conversely, LECGABA suppresses this local inhibition to disinhibit CA3 activity with compartment and pathway specificity by selectively boosting somatic output to integrated LECGLU and CA3 recurrent inputs. This synergy allows the stabilization of spatial representations relevant to learning, as both LECGLU and LECGABA control the formation and maintenance of CA3 place cells across contexts and over time.

Episodic and semantic memory are classically thought to play distinct roles: episodic memory encodes unique experiences, while semantic memory generalizes across them. Current conceptualizations of episodic and semantic memory interactions emphasize a one-way consolidation from episodic traces in the medial temporal lobe (MTL) to semantic knowledge in neocortex (CTX). However, this tradition has left largely unexplored how semantic memory may affect episodic encoding of new memories. Here we introduce a cognitive model in which the code used in episodic memories shifts from purely sensorial to including explicit semantic representations of stored events. Simultaneously, we propose a computational circuit model of how such a cognitive strategy could be implemented in the brain using biologically-plausible learning rules. We show that increased sparsity during replay enables neocortex to extract compositional structure from overlapping episodes, which creates a dictionary of inter-connected concepts in semantic memory. Furthermore, we show that spontaneous activity in neocortical areas can imprint the abstracted representations into the medial temporal lobe, giving rise to concept-like cells. This bidirectional interaction improves episodic recall and replay fidelity, and facilitates the consolidation of higher-order representations based on previous semantic knowledge. The model accounts for behavioural advantages of schema-congruent learning, the emergence of concept neurons, and enhanced memory performance for semantically familiar stimuli. Together, our results provide a mechanistic account of how episodes and concepts reinforce each other, extending standard consolidation theories and suggesting a cooperative framework where semantic knowledge scaffolds episodic encoding, which in turn favours compositional abstraction. ### Competing Interest Statement The authors have declared no competing interest.

Prioritized experience replay, which improves sample efficiency by selecting relevant transitions to update parameter estimates, is a crucial component of contemporary value-based deep reinforcement learning models. Typically, transitions are prioritized based on their temporal difference error. However, this approach is prone to favoring noisy transitions, even when the value estimation closely approximates the target mean. This phenomenon resembles the noisy TV problem postulated in the exploration literature, in which exploration-guided agents get stuck by mistaking noise for novelty. To mitigate the disruptive effects of noise in value estimation, we propose using epistemic uncertainty estimation to guide the prioritization of transitions from the replay buffer. Epistemic uncertainty quantifies the uncertainty that can be reduced by learning, hence reducing transitions sampled from the buffer generated by unpredictable random processes. We first illustrate the benefits of epistemic uncertainty prioritized replay in two tabular toy models: a simple multi-arm bandit task, and a noisy gridworld. Subsequently, we evaluate our prioritization scheme on the Atari suite, outperforming quantile regression deep Q-learning benchmarks; thus forging a path for the use of uncertainty prioritized replay in reinforcement learning agents.

Flexible behaviour requires transforming abstract cognitive representations, such as value preferences, into concrete motor actions. During economic decision-making, individuals evaluate options to guide choices and then transform these choices into specific actions to obtain rewards, yet the neural circuit mechanisms linking abstract choices to spatial actions remain unclear. Here we show that a frontal motor network performs this transformation through dynamic circuit reconfiguration across decision stages. Using a mouse task that temporally dissociates value-guided decisions from spatial action planning, cortex-wide optogenetic perturbations revealed that a specific frontal motor circuit was causally required for both abstract and motor stages of choice. Electrophysiological recordings during abstract decision-making showed that frontal motor neurons encoded option values and economic choices independently of sensorimotor contingencies, and unilateral silencing impaired decisions without spatial bias. During spatial planning, value and spatial signals were integrated to guide action selection, and unilateral silencing produced an ipsilateral bias. A dynamical model captured this transformation, predicting a shift from distributed interhemispheric maintenance of economic choice to lateralized control of actions, which we validated with simultaneous bilateral recordings. These findings demonstrate how frontal motor circuits reconfigure their interactions to convert cognitive decisions into motor actions, providing a circuit-level mechanism for flexible, value-guided behaviour. ### Competing Interest Statement The authors have declared no competing interest. Gatsby Charitable Foundation, https://ror.org/0290hax27, GAT3755 Wellcome Trust, https://ror.org/029chgv08, 219627/Z/19/Z, 309104/Z/24/Z