Ştefan Mihalaş

The paper presents a database of over 1,000 differentiable single-neuron models fitted to mouse and human patch-clamp recordings, with PyTorch code for integration into machine learning.

On 495 human Patch-seq neurons, cell type means predicted electrophysiology better than neural networks on scGPT embeddings or gene panels; fusing scGPT embeddings with cell type labels predicted best.

The paper introduces CoSyn-RNN, a continual-learning model inspired by biological neural systems, which maintains performance across tasks through synaptic state protection and context modulation.

The paper introduces a plasticity-switching recurrent neural network (psRNN) that models synaptic state-dependent plasticity and shows it improves learning efficiency in cognitive tasks.

The paper presents a database of spiking activity in the mouse visual system during an image change detection task, revealing task-specific coding patterns across brain regions.

This study shows how rapid learning in mice involves structured changes in motor cortex connectivity, rerouting preparatory activity to control a conditioned neuron.

The paper demonstrates that embedding connectome data into a geometric space improves prediction of neural activity compared to high-dimensional models.

This study integrates molecular and morphological data to classify cortical neurons and map their connectomes, revealing organizational principles of brain-wide networks at the single-cell level.

The paper identifies neural assemblies consistent with Hebb's theory using a novel dataset linking optical physiology to connectivity, finding they are structurally integrated but not defined by strong monosynaptic connections.

This paper studies RNNs with heavy-tailed distributions and finds a slow transition to chaos with tradeoffs in dynamical robustness and dimensionality.

The paper examines spatial convergence patterns of sensory inputs in the mouse isocortex, introducing the topography dimension (TD) and its relationship with neural activity and hierarchy.

This study examines if communication subspaces develop with training in ANNs and finds increased alignment between activations and weights in trained models.

This study investigates how training data properties affect the alignment of deep neural networks with mouse visual cortex activity, finding that naturalistic environments significantly improve similarity.

The paper shows that chaotic dynamics in recurrent neural networks separate learning regimes and correlate with a consciousness measure, suggesting a link between critical dynamics, learning, and consciousness.

This paper introduces a simple synaptic learning mechanism that enables a neural network to detect novelty in three distinct forms, demonstrating how basic plasticity can generate diverse novelty responses.

This study shows that backward masking in mice requires visual cortex, as optogenetic silencing of the cortex restores target discrimination.

The paper introduces a bio-inspired neural network architecture with switching units that improves generalization across different visual backgrounds.

The paper develops a neuronal model with complex dynamics and shows it can learn temporal tasks using gradient descent, exhibiting diversity and robustness.

This paper analyzes the neural population dynamics of a multi-plasticity network (MPN) that uses synaptic modulations for short-term computation, finding it outperforms RNNs on neuroscience tasks with simpler attractor structures.

The paper identifies diverse functional neuron types in the visual cortex that process novel and familiar stimuli in mice.

This paper fills gaps in cell-class specific connectivity data using a statistical model and kernel learning method, producing a mesoscale mouse brain connectivity matrix.

The paper maps brain-wide monosynaptic inputs to specific neuron classes in mouse visual cortex using an improved rabies virus tracing strategy and identifies connectivity patterns within and between visual areas.

The paper examines representational drift in mouse visual cortex neurons, finding it occurs along high-variance directions and is similar to dropout in artificial neural networks, suggesting a computational advantage.

The paper introduces MouseNet, a biologically constrained CNN model of the mouse visual cortex, and shows it achieves 2/3rds of VGG16's ImageNet performance while recapitulating neural representations.

This paper proposes a learning rule for recurrent spiking neural networks that incorporates cell-type-specific modulatory signals to improve temporal credit assignment.

This paper introduces a biologically plausible method for continual learning in ANNs that prevents catastrophic forgetting and promotes synaptic multi-modality.

This paper compares neural adaptation mechanisms to persistent activity in a visual change detection task using mice and artificial neural networks, finding that adaptation better matches neural and behavioral data.

The paper uses autoencoder networks to show that regularization can resolve ambiguity in connectivity weights, allowing latent variable extraction through nonlinear dimensionality reduction.

This paper shows that mixing diverse activation functions in CNNs improves image classification performance, suggesting a computational role for diverse cell types in biological circuits.

The paper studies how interactions between mouse visual cortical areas change over time using functional connectivity analysis of neural activity data.