HVOR SIKKER ER GATEIO

A novel and general framework for dimensionality-reducing visualization, hypersphere2sphere takes high-dimensional distributions of labeled data as objects and shows their relationships.

Convolutional neural networks can achieve amazing performance on many computer vision tasks, but are brittle in the face of carefully selected, imperceptibly small changes to images ("adversarial attacks"). Networks could be made more robust with inductive biases that come from the statistics of simple dynamic worlds.

A novel, brain-inspuired neural network architecture based on biological feature map organization.

Previous theories of visual motion computation by neurons in area MT could not capture the wide temporal frequency tuning they commonly exhibit. A series of targeted experiments fit with theoretical models revealed the importance of tuned normalization.

Time for running experiments is limited. By prioritizing regions of spatiotemporal frequency space, the types of motion that excite and suppress MT neurons could be measured in unprecedented detail.

How motion tuning changes in different spatial locations of MT receptive fields was poorly understood. A novel filtered-noise stimulus revealed this spatial organization.

Autobiography of an electrode.