2019/03/11 by Johannes Thiele, Thiele, Johannes C., Olivier Bichler +7
Engineering · Neuroscience · Computer Science · #Advanced Memory and Neural Computing #Neural dynamics and brain function #Neural Networks and Reservoir Computing
paper · pdf · doi:10.48550/arxiv.1903.04341
The increasing need for intelligent sensors in a wide range of everyday\nobjects requires the existence of low power information processing systems\nwhich can operate autonomously in their environment. In particular, merging and\nprocessing the outputs of different sensors efficiently is a necessary\nrequirement for mobile agents with cognitive abilities. In this work, we\npresent a multi-layer spiking neural network for inference of relations between\nstimuli patterns in dedicated neuromorphic systems. The system is trained with\na new version of the backpropagation algorithm adapted to on-chip learning in\nneuromorphic hardware: Error gradients are encoded as spike signals which are\npropagated through symmetric synapses, using the same integrate-and-fire\nhardware infrastructure as used during forward propagation. We demonstrate the\nstrength of the approach on an arithmetic relation inference task and on visual\nXOR on the MNIST dataset. Compared to previous, biologically-inspired\nimplementations of networks for learning and inference of relations, our\napproach is able to achieve better performance with less neurons. Our\narchitecture is the first spiking neural network architecture with on-chip\nlearning capabilities, which is able to perform relational inference on complex\nvisual stimuli. These features make our system interesting for sensor fusion\napplications and embedded learning in autonomous neuromorphic agents.\n