Neural Networks and Applications

Computation or Weight Adaptation? Rethinking the Role of Plasticity in Learning

This paper challenges the traditional view that neural plasticity is essential for adaptive learning by showing that sufficiently trained large language models can achieve adaptability through computation alone.

Neural networks as decision trees: an analytical solution for learning and neural selectivity

This paper develops an analytical framework showing how nonlinear neural networks decompose tasks into local linear regressions and how this structure relates to neural selectivity and activation regions.

Neuromodulated Equilibrium Learning

This paper introduces Neuromodulated Equilibrium Learning (NEL), a single-equilibrium reward-based framework, and evaluates its performance on image datasets compared to multi-phase baselines.

NeuralRNN: a unified framework for recurrent neural network methods in cognitive neuroscience

NeuralRNN unifies RNN methods for cognitive neuroscience tasks by providing a single framework for model implementation, comparison, and extension.

Neural representational geometry of a joint code for stimulus category and category-independent features

The paper investigates how neural representations can jointly encode stimulus categories and category-independent features, showing that regression-relevant geometry can be optimized without compromising classification geometry.

Weight initialization shapes task organization in recurrent neural networks

This study shows that low initial weight variance in recurrent neural networks leads to overlapping weight subspaces for subtasks, while high variance results in more segregated subspaces.

Reward-Based Equilibrium Propagation

This paper introduces a reward-based variant of equilibrium propagation that uses sparse reward signals instead of full target vectors, achieving performance close to conventional EP on multiple datasets.

The Role of Training Paradigms in Shaping Connectivity and Dynamics: A Comparison of Gradient-Based and Evolutionary Methods in Recurrent Neural Network Models

This paper compares gradient-based and evolutionary training methods in RNNs, showing they produce distinct connectivity and dynamics, especially in working memory tasks.

Dendritic Wave Recurrent Neural Networks

This paper introduces the Dendritic Wave Recurrent Neural Network (DW-RNN), which improves upon wave recurrent neural networks by adding dendritic input integration, achieving better performance and stability on sequential tasks.

Signed-XOR Error and Sparse Coding in a Dale-Complaint Substrate for Sequence Memorization

The paper demonstrates a biologically plausible model that memorizes and recalls sequences using a sparse coding scheme and Dale's law constraints, achieving perfect recall with a modest latent size.

Interpretable compositional computation with recurrent neural networks

This paper develops a theory of interpretable compositional computation using low-rank recurrent neural networks, identifying shared latent components and their task-dependent usage.

Incorporation of single-neuron projectome-based connectivity motifs enhances the cortex-specific performance of artificial neural networks

This paper develops a connectome-informed neural network algorithm that improves RNN and LLM performance by incorporating natural cortical connectivity motifs.

Rank dependency of rescaled pruning in recurrent neural networks

The paper shows that synaptic rescaling preserves low-dimensional dynamics in low-rank RNNs but fails in high-rank networks during pruning.

Local gated-Hebbian learning of deep cerebellar networks with quadratic classification capacity

The paper shows that a two-hidden-layer network with local Hebbian learning can achieve quadratic classification capacity, supporting fast and efficient learning in cerebellum-like circuits.

Linking macroscale structure and function in brain-like recurrent neural networks

This paper introduces BrainRNN, a brain-like recurrent neural network architecture, and shows that structural constraints lead to functional organization similar to the human cortex.

Toward defining loss functions in neuroscience: an XOR-based neuronal mechanism for local learning

This paper investigates whether a signed error derived from an XOR-based neuronal mechanism can serve as a learning signal in a binary autoencoder, showing improved reconstruction accuracy compared to traditional methods.

Evolutionarily Optimized Network Topology as a Structural Prior for Data-Efficient Sparse Neural Classification

The paper tests if evolutionarily optimized network topologies can improve data efficiency in neural classifiers, finding that biologically derived sparse networks outperform synthetic alternatives with similar sparsity.

Individual differences in artificial neural networks capture individual differences in human behavior

This study shows that individual differences in artificial neural networks mirror human behavioral differences in accuracy, confidence, and response time across tasks.

Synaptic pruning facilitates online Bayesian model selection

This paper shows synaptic pruning enables efficient Bayesian model selection by reducing neural network structures through connectivity parameters.

Predictive Coding algorithms induce brain-like responses in Artificial Neural Networks

This paper shows that predictive coding-inspired neural networks exhibit brain-like behaviors such as mismatch responses and prior formation, suggesting they can model biological neural networks.

Human-like individual differences emerge from random weight initializations in neural networks

This study shows that different random initializations of neural networks produce individual differences similar to humans, and using these differences improves predictions of human responses.

Single-unit activations confer inductive biases for emergent circuit solutions to cognitive tasks

This study shows that single-unit activation functions in RNNs influence circuit solutions and generalization behavior on cognitive tasks.

Non-Negative Connectivity Causes Bow-Tie Architecture in Neural Circuits

This study shows that non-negative connections in neural networks lead to bow-tie architecture, which is efficient, robust, and generalizable.

Partial recurrence enables robust and efficient computation

This paper introduces partially recurrent neural networks (pRNNs) and shows they can match or exceed fully recurrent networks in performance while using fewer parameters.

Taming the chaos gently: a Predictive Alignment learning rule in recurrent neural networks

This paper introduces a novel learning rule called Predictive Alignment for recurrent neural networks that suppresses chaos to generate patterned activities using biologically plausible plasticity.

Impact of dendritic non-linearities on the computational capabilities of neurons

This paper shows that dendritic non-linearities improve a neuron's computational power by increasing learned associations, learning speed, and generalization performance.

Parallel Synapses with Transmission Nonlinearities Enhance Neuronal Classification Capacity

This paper shows that neurons with multiple nonlinear synapses can classify data better than a Perceptron and improve test accuracy in MNIST classification.

Mapping the Learning Curves of Deep Learning Networks

This paper introduces a method to map learning curves of deep learning networks, identifying insights like nonlinearity and domain distinctions across tasks.

A biological model of nonlinear dimensionality reduction

This paper develops a biologically plausible nonlinear dimensionality reduction algorithm inspired by Drosophila olfactory circuits and demonstrates its effectiveness on datasets like entangled rings and MNIST.

Diversity in the impact of heterogeneities on recurrent networks performing a cognitive task

This study assesses how different levels of heterogeneity affect the performance and dynamics of recurrent neural networks trained on a cognitive task.