Sander M. Bohté

This paper proposes a recurrent neural network architecture that learns to trace curves and objects using multiscale, incremental grouping, showing generalization from short to longer curves and 2D-objects.

The paper shows that energy optimization in a spiking neural network can produce predictive-coding properties similar to experimental observations.

This paper trains a recurrent neural network with reinforcement learning to solve multi-step visual tasks, showing it can learn elemental operations and propagate enhanced activity between subroutines.

This paper develops a spiking neural network for predictive coding that generates hierarchical internal representations and reconstructs unseen data, suggesting biologically plausible mechanisms for perceptual inference.

This paper shows how a predictive coding network learns invariant object representations and generative models through local prediction error minimization, using dynamic inputs to simulate visual perception.

This paper introduces a biologically plausible recurrent neural network called RECOLLECT that can learn to remember and forget information over time, demonstrating its ability to perform tasks similar to animal learning.

This paper develops deep learning models to study dynamic object recognition, finding that sequential processing with lateral recurrence matches human performance and improves recognition with adaptation.

The paper presents a computational model showing how arousal state modulates task performance by altering sensory processing hierarchy, explaining the Yerkes-Dodson effect.

This study uses spiking deep neural networks to compare attention mechanisms, finding that modulating outgoing connections is more effective than precision or input gain in enhancing visual search performance.

This paper introduces adaptive spiking recurrent neural networks with a novel surrogate gradient, achieving state-of-the-art performance in time-domain tasks and improved computational efficiency.

The paper develops a deep predictive coding network trained with Hebbian learning that exhibits neural response properties similar to experimental findings in the visual cortex.

This paper introduces WorkMATe, a neural network model that learns to control working memory through reinforcement learning, solving both simple and complex tasks with flexible memory management.

This paper shows that deeper convolutional neural networks better distinguish objects from backgrounds in scenes, suggesting that scene segmentation can emerge from network depth without explicit mechanisms.

This paper develops a deep learning approach to predict ALS from genotype by leveraging genome structure, showing better performance than other methods.

This paper introduces a deep predictive coding network that learns latent representations from real-world images in an unsupervised manner, successfully inferring representations for novel objects.

The paper uses multi-task deep neural networks to analyze visual pathway organization, finding that unrelated tasks lead to less feature sharing in higher layers.