The Darshan Lab: Theoretical and Computational Neuroscience Lab
Publications
Pereira-Obilinovic, U., Daie, K., Chen, S., Svoboda, K., Darshan, R. (2025). Neural dynamics outside task-coding dimensions drive decision trajectories through transient amplification.
Biorxiv
​​​This study challenges the assumption that low dimensional task aligned coding subspaces drive decisions. Using a new recurrent neural network model fit directly to neural population recordings, we reveal that activity in high dimensional residual subspaces, typically dismissed as noise, can causally alter behavior. These residual dimensions transiently amplify neural trajectories before decisions settle into discrete attractor states. Perturbing them flips choices, while perturbing classic choice dimensions does not. Our findings redefine weakly task encoding neural activity as structured, computationally essential dynamics driving flexible decision-making.
Schlisselberg, O. & Darshan, R. (2025).
The impact of allocation strategies in subset learning on the expressive power of neural networks.
The Thirteenth International Conference on Learning Representations (ICLR 2025)
We introduce a theoretical framework for analyzing how the allocation of a limited number of learnable weights affects neural network expressivity. Using a teacher-student setup, we establish conditions for allocations yielding maximal and minimal expressivity in linear recurrent and deep feedforward networks, showing that broader distribution of learnable weights across the network enhances expressive power. Insights from linear networks extend to shallow ReLU networks. Our study emphasize the critical role of strategically distributing learnable weights across the network, showing that a more widespread allocation generally enhances the network’s expressive power.
Manoim-Wolkovitz, J.E., Camchy, T., Rozenfeld, E., Chou, YH.,
Darshan, R., Parnas, M. (2025).
Nonlinear high-activity neuronal excitation enhances odor discrimination.
Current Biology
By combining modeling with experiments, we show that in the fly's olfactory system, nonlinear intraglomerular excitation improves odor classification. This excitation, mediated by muscarinic type B receptors (mAChR-B), emerges only at high olfactory receptor neuron (ORN) firing rates and enhances discrimination between odor signals. Behavioral experiments confirm the model’s predictions, showing that knocking down mAChR-B increases correlations across ORNs and impairs odor discrimination. Our study unravels a novel mechanism for neuronal pattern decorrelation.
Finkelstein, A., Daie, K., Rozsa, M., Darshan, R. & Svoboda, K. (2025). Connectivity underlying motor cortex activity during naturalistic goal-directed behavior.
Nature
We study neural activity in the motor cortex during naturalistic behavior in which mice gathered rewards with multidirectional tongue reaching. We used an all-optical method for to reveal the causal functional connectivity mapping in the mouse motor cortex. Neurons were organized into a fine-scale columnar architecture, with local like-to-like connectivity according to target-location tuning, and inhibition over longer spatial scales. Connectivity patterns comprised a continuum, with abundant weakly connected neurons and sparse strongly connected neurons that function as network hubs. This network of neurons, encoding location and outcome of movements to different motor goals, may be a general substrate for rapid learning of complex, goal-directed behaviors.
Amsalem, O., Inagaki, H., Yu, J., Svoboda, K., & Darshan, R. (2024).
Sub-threshold neuronal activity and the dynamical regime of cerebral cortex.
Nature Communications 15 (1), 7958
We assessed whether the cortex operates in a fluctuation-driven regime by combining theories of network and single-neuron dynamics with analysis of spiking and sub-threshold membrane potentials of neurons in the sensory and frontal cortex during a decision-making task. While the frontal cortex seems to operate in a fluctuation-driven regime, excitatory neurons in the layer 4 of the barrel cortex respond to occasional synchronous inputs, indicating fundamental dynamical differences between cortical areas.
Kim, C. M., Finkelstein, A., Chow, C. C., Svoboda, K., & Darshan, R. (2023). Distributing task-related neural activity across a cortical network through task-independent connections.
Nature Communications, 14(1), 2851.
During goal-directed behaviors, neural activity in populations of neurons changes, but little is known about the underlying synaptic reorganization. In this study, a spiking network with strong synaptic interactions was trained, and task-related activity emerged even in untrained neurons, suggesting a mechanism where task-independent strong synapses mediate the spread of task-related activity across cortical networks.
Arthur, B. J., Kim, C. M., Chen, S., Preibisch, S., & Darshan, R. (2023).
A scalable implementation of the recursive least-squares algorithm for training spiking neural networks.
Frontiers in Neuroinformatics, 17.
We present optimized CPU and GPU implementations of the recursive least-squares algorithm in spiking neural networks, enabling faster training of large models. The fast implementation facilitates interactive in-silico studies of multi-area computations and closing the loop between modeling and experiments.
Rajagopalan, A. E., Darshan, R., Hibbard, K. L., Fitzgerald, J. E., & Turner, G. C. (2023).
Reward expectations direct learning and drive operant matching in Drosophila
Proceedings of the National Academy of Sciences 120 (39), e2221415120
We discovered operant matching in Drosophila and utilized computational tools to identify synaptic plasticity mechanisms in the mushroom body that incorporate reward expectations. By optogenetically bypassing the reward expectation representation, we abolished matching behavior, demonstrating the essential role of reward expectation signals in the fly brain.
Darshan, R., & Rivkind, A. (2022).
Learning to represent continuous variables in heterogeneous neural networks.
Cell Reports, 39(1).
We developed a theory for manifold attractors in neural networks that approximates a continuum of persistent states without relying on unrealistic symmetry. Our work suggests that the brain's functional properties regarding monitoring continuous variables can be inferred from asymmetries in connectivity and the low-dimensional representation of encoded variables, offering insights into the formation and stability of manifolds in trained neural networks.
Lebovich, L., Darshan, R., Lavi, Y., Hansel, D., & Loewenstein, Y. (2019). Idiosyncratic choice bias naturally emerges from intrinsic stochasticity in neuronal dynamics.
Nature Human Behaviour, 3(11), 1190-1202.
Idiosyncratic tendency to choose one alternative over others in the absence of an identified reason is a common observation in two-alternative forced-choice experiments. We found substantial and significant biases in perceptual and motor tasks, which theoretical evidence suggests arise from the dynamics of competing neuronal networks, making idiosyncratic choice bias virtually inevitable in any comparison or decision task, even under idealized symmetric settings. In addition, we develop a novel decision-making model that aligns with the observed spiking statistics of neruons in the cortex.
Engelhard, B., Darshan, R., Ozeri-Engelhard, N., Israel, Z., Werner-Reiss, U., Hansel, D., Hagai Bergman & Vaadia, E. (2019).
Neuronal activity and learning in local cortical networks are modulated by the action-perception state.
bioRxiv, 537613.
In this study, we explore how the brain utilizes neuronal patterns during sensorimotor learning, focusing on familiar motor tasks and brain-machine interface (BMI) learning. The neuronal patterns during BMI learning are distinct from movement-state patterns, evolving to increase the firing rate of the target neuron while maintaining excitatory-inhibitory balance. Our novel neural-level reinforcement-learning network model predicts these results and indicates that the brain adapts patterns to the current action-perception state to gain rewards. This finding may have significant implications for clinical brain-machine interface applications to modify impaired brain activity
Darshan, R., Van Vreeswijk, C., & Hansel, D. (2018).
Strength of correlations in strongly recurrent neuronal networks.
Physical Review X, 8(3), 031072.
We develop a general theory linking pairwise correlations' strength to anatomical features in strongly interacting neuronal networks (balanced networks). Despite strong interactions, activity remains irregular in a large parameter space. We identify architectural features that allow macroscopically correlated activity to arise when the architecture embeds a group of neurons connected to other groups in a unidirectional manner without reverberations. This feed-forward structure can be explicit or hidden.
Darshan, R., Wood, W. E., Peters, S., Leblois, A., & Hansel, D. (2017).
A canonical neural mechanism for behavioral variability.
Nature communications, 8(1), 15415.
The ability of humans and animals to learn motor tasks, acquire language or make decisions, often relies on the ability to explore the outcomes of their actions. Using a combination of modeling, electrophysiological recordings, and behavioral studies, we show that a model circuit with topographically organized and strongly recurrent neural networks can generate irregular motor behaviors. Neural correlations increase across the circuit in agreement with experimental findings in singing birds, and the model accounts for similar behavioral statistics observed in 5-6-month-old human infants and juveniles from three songbird species.
Darshan, R., Leblois, A., & Hansel, D. (2014).
Interference and shaping in sensorimotor adaptations with rewards.
PLoS computational biology, 10(1), e1003377.
The brain's ability to adapt to sensorimotor changes can be influenced by the shape of reward signals. Using a mathematical model, we found that adaptation dynamics depend on motor variability and the shape of the reward function. When adapting to multiple sensory stimuli, interference effects arise from overlapping neural representations and physical distance, impacting the speed of adaptation. Remarkably, increasing the number of sensory stimuli during training accelerates adaptation to widespread sensorimotor perturbations, demonstrating that learning is faster with a greater number of stimuli.
Leblois, A., & Darshan, R. (2014).
Basal Ganglia: Songbird Models.
We review recent advances in songbird models, focusing on their learned vocalizations called songs. Songbirds have specialized basal ganglia thalamocortical circuit that play a crucial role in song learning and plasticity, offering a unique opportunity to study the function of the basal ganglia in skill learning and motor execution.