Unsupervised Learning of Overlapping Image Components Using Divisive Input Modulation
Node activations generated in response to
input images containing overlapping pixel square components. (a) Test
images. Numbers indicate which components are present in each test image.
(b)–(f) Images reconstructed from the response of each network. Numbers
indicate which components are represented by strongly active nodes in each
network (nodes that have an output activation greater than the mean of all node
activations). Each network was given predefined weights so that distinct nodes
represented pixel squares at all possible locations.
(a) Test images
Article of the Year Award: Outstanding research contributions of 2020, as selected by our Chief Editors. Read the winning articles.