Neural Networks and Applications

The paper introduces a model-free measure called effective number of shared dimensions (ENSD) to quantify shared structure between datasets, demonstrating its utility across neural networks and neuroscience data.

ElegansNet uses the C. elegans connectome to design deep learning systems that outperform random and artificial networks in image classification and digit reconstruction tasks.

The paper develops biologically plausible variants of RBMs with sparse and asymmetric connectivity and benchmarks their performance on MNIST.

This paper trains a biologically realistic neural network model, the stabilized supralinear network, on challenging tasks using a dynamics-neutral growth method during training.

This paper introduces a biologically realistic local learning rule that enables recurrent networks to achieve invariant object representation and classification without error backpropagation.

The paper explores how biological connectomes can inform artificial neural network design, showing that architectural statistics from biological circuits provide a useful prior for machine learning tasks.

This paper identifies low-dimensional 'operative dimensions' in RNNs that significantly influence dynamics, showing that much of the high-dimensional connectivity is functionally irrelevant.

This paper introduces APCNs, a neural network framework that learns reference frames and parses visual scenes into part-whole hierarchies, demonstrating compositional representation and transfer learning on standard datasets.

The paper introduces a Reset model composed of convolutional neural networks that self-organize into clusters and exhibit topographic organization similar to parietal cortex number representation.

This paper compares weight perturbation and node perturbation learning rules on temporally extended tasks, finding that weight perturbation can perform as well or better under certain conditions.

This paper introduces a developmental framework for machine learning that mimics neuroevolutionary principles to improve task performance and metalearning capabilities.

The paper explores how neural circuits can be compressed through a genomic bottleneck, showing that this compression can enhance transfer learning and suggest implications for AI design.

This paper introduces DGNs, a biologically plausible learning rule that outperforms backpropagation in efficiency and forgetting resistance, inspired by cerebellar structure and function.

This paper explores bio-instantiated recurrent neural networks, examining their performance in working memory tasks and demonstrating how empirical neural network data can be used to construct biologically realistic topologies.

Splitting network units into excitatory and inhibitory populations need not cost learning: initialize inhibition to centre and standardize excitatory activity, and scale inhibitory updates by Fisher information.

The paper introduces a novel unsupervised learning algorithm using modular sparse distributed codes that preserves similarity and enables fixed-time learning and retrieval.

This study examines how network structure and information content change during an artificial learning task, identifying three phases of network reconfiguration.

This paper studies how RNNs solve a hierarchical inference task with two latent variables and disparate timescales, finding they learn behavior similar to Bayesian baselines and develop interpretable models through knowledge distillation.

This paper introduces a Salience Affected Artificial Neural Network (SANN) that enables one-time learning through salience tagging, improving classification accuracy on MNIST and Fashion MNIST datasets.

This paper proposes a meta-learning approach to design biologically plausible synaptic update rules for semi-supervised learning, showing effectiveness on classification tasks.

The paper provides an overview of computations, optimization, and tuning in deep feedforward neural networks, using a small network to explain generalizations for larger networks.

This paper introduces a method using replicated autoencoders optimized for local entropy to infer compositional features from noisy data, showing improved performance on synthetic and protein sequence data.

This paper introduces an animal-like neural network (RRN) that mimics biological recognition mechanisms, showing it can maintain object invariance and learn without disrupting existing functions.

The paper introduces a Julia package for deep Boltzmann machines to enable unsupervised learning on biomedical data.

The paper shows that similarity-preserving neural networks develop localized receptive fields that tile manifolds, mirroring brain representations like place cells.

The paper proposes a network model of neocortical columns and layers that learns object recognition through temporal integration and lateral connections, showing improved performance with multi-column networks.