AI summary · may contain errors
Gene co-expression–initialized sparse multilayer perceptrons classify handwritten digits with 65% accuracy using 100 training examples, far above a dense baseline at 11%.
- Open access
- 1 cites
Moving in bioRxiv, Nature Communications, Nature Computational Science, Nature Neuroscience, eLife and Nature Reviews Neuroscience.
Rebuilt
Nothing in this topic has moved yet this week.
Newest first · last 60 days
AI summary · may contain errors
Gene co-expression–initialized sparse multilayer perceptrons classify handwritten digits with 65% accuracy using 100 training examples, far above a dense baseline at 11%.
AI summary · may contain errors
A variational principle in recurrent neural networks trained on time-interval reproduction predicts three regimes of learned neural geometry, including boundary distortions and affine readout behavior.
AI summary · may contain errors
Feature-based readout adaptation in a simulated torque-driven planar two-link arm shows retained compensation after updates stop, with tracking error reduced by 40–97% versus pre-adaptation.
AI summary · may contain errors
Below Hopfield memory capacity, a neuron’s activity is dominated by high redundancy between external pattern input and internal recurrent input, while synergy and unique information stay near zero until capacity is exceeded.
AI summary · may contain errors
A physical deep-learning framework on an optoelectronic platform uses state-skipping direct feedback alignment and a single-map 1D dynamics architecture to cut computational resources by ~40% while matching performance.
AI summary · may contain errors
Optimizing controversial stimulus sets can make neural network models disagree on predictions, improving discrimination among brain information-processing hypotheses despite high model capacity.
AI summary · may contain errors
A binary autoencoder trained for five epochs on binarized MNIST using a signed XOR mismatch error reaches 97.7% bit-wise and 94.8% balanced reconstruction accuracy without backpropagation.
AI summary · may contain errors
Heterogeneous self-interactions in a recurrent neural network generate scale-invariant memory in subcritical regimes, without assuming criticality, and promote spontaneous hierarchical organization.
AI summary · may contain errors
Sixty instances of the same ANN architectures show substantial individual differences in digit recognition, with mapping to human differences approaching the human-to-human benchmark across accuracy, confidence, and RT.
AI summary · may contain errors
YamAdam, an adaptive optimizer with no user-specified optimizer hyperparameters, reaches 66 targets that comparators miss and matches 2,634 of 2,700 target observations.
AI summary · may contain errors
A systematic comparison shows when a dopamine (DA)-modulated neural mass mean-field reduction accurately matches its fully known spiking network across multiple dynamical regimes.
AI summary · may contain errors
Appellian Neural Network (ANNet) unifies inverse and forward dynamics for a double pendulum by learning Appell’s acceleration energy, enabling accurate real-time simulations for unseen trajectories.
AI summary · may contain errors
A dendritic-inspired hybrid-attention foundation model achieves a 50M context window (about 40 volumes of Shakespeare), improving long-term memory while lowering computational demands.
AI summary · may contain errors
Heterogeneous rodent head-direction cell tuning is compatible with quasicontinuous attractors, and disordered recurrent networks built from data exhibit circular symmetry linked to continuous-attractor organization.
AI summary · may contain errors
Optimally designed perturbation inputs improve neural system identification accuracy by substantially outperforming conventional, intuition-based inputs in simulations.
AI summary · may contain errors
Accuracy increased across four covertly manipulated evidence durations in noiseless, long-duration perceptual judgments, supporting protracted sampling but not uniquely identifying integration or a decision bound.
AI summary · may contain errors
INSPIRE enables on-chip in situ physical gradient descent for photonic integrated circuits, achieving trained matrices with 0.26% relative error and supporting in situ training beyond native tunable elements.
AI summary · may contain errors
PC and FCI can recover meaningful causal information from single-time-point observations in ecological dynamical systems, identifying clustered super-unshielded colliders that encode partial macro-level causal ordering.
AI summary · may contain errors
People in preregistered behavioral experiments prefer the simpler of two alternative explanations for uncertain data, with preferences matching model-selection theories that penalize excessive flexibility.
Journals this month
Topics moving, week to 17 Nov 2026
By field