Скачать или смотреть Unsupervised Feature Learning with Hebbian Plasticity and Adaptive Thresholding

In just over 100 time steps, a two-layer convolutional neural network learns to recognize objects utilizing only local Hebbian learning rules. The network initially focuses on the red and orange marker caps, but eventually converges on the marker bodies as the most common feature shared between the images. Layer 2 activations are represented by the magenta squares on the right screen.

CORTEX Systems has successfully tested a novel fast learning algorithm for deep convolutional neural nets that utilizes adaptive thresholds and the difference between average activation rates to modulate simulated spike frequency response windows.

Available for free download at the Play Store: https://play.google.com/store/apps/de...

Code available at http://cortexsystems.yolasite.com/

Note: in this video, a low threshold corresponds to low input variance acceptance (absolute difference of input and weight must be LESS than threshold). While the same function is used in the app, the threshold graph has been inverted to represent the conventional use of thresholding (activation function must be GREATER than threshold).