Transcript Slide 1
Dense-Near/Sparse-Far Hybrid Reconfigurable Neural Network Chip Robin Emery Alex Yakovlev Graeme Chester Overview • Motivation • System Elements & Structure • Current Work • Future Work 2 Previous Work • Artificial neural network • Xilinx Virtex-II FPGA • Variable precision • Generated using mark-up • Controlled via PC Threshold Inputs Fire Latch Weight Table External Stimulus Sum Decay Module Reserved For Decay 3 Fired Previous Work • Exhausted area before routing resource • Synchronous, Low neuron count • No autonomous learning • FPGA routing resources occupy 70-90% • Real-time learning awkward 4 A Neuron 5 A Network of Neurons • Billions of neurons in the brain • 100 to 3000 connections per neuron • Majority of connections are proximal • Spikes are generally the same 6 Clusters • Axons of neocortical neurons form connections in clusters 7 Learning • In the synapse • Plastic connection • Use learning rule • Autonomous in synapse • Wider mechanism may exist 8 Motivation • A FPGA-like neural network device would be of interest to neuroscience • Connectivity is also of interest • Observations support a hybrid of local and distal connectivity • More useful with real-time learning 9 System Elements • Neuron • Synapse • AER Router • AER/Spike Bridge • Routing Resource • Protocol 10 AER • Address Event Representation • Asynchronous digital multiplexing • Stereotyped digital amplitude events • Nodes share frame of reference • Information is encoded in the time and number of events 11 Dense-Near Connectivity 12 Sparse-Far Connectivity 13 Network Structure 14 Current Work • Neuron – Configurable threshold – Asynchronous – 7-bit count – Decay – Spike generator 15 Current Work • Neuron & Spike Generator • 130nm UMC CMOS Area 1145.6μm2 (90nm: 700μm2) Gates 390 Density 873 p. mm2 (90nm: 1429 p. mm2) Spike Period 4.5ns Generated Clock Frequency 160MHz Max. Spike Rate (theshold=100) 2.35 million p. second 16 Current Work • Software model & protocol refinement • Ongoing work: – Autonomous Synapses – AER Router/Bridges 17 Evaluation • Topographic map • Compare to popular software modelling tool such as NEURON 18 Future Work • Long-term learning process • Improve capacity of AER link by grouping spikes • Aggregation of pulse-widths could improve range of dendritic input • Multiplexing of some direct links 19 Conclusions • Reconfigurable, adaptive neural network system • Real qualities of interest to neuroscientists • Neuron and spike generator manufactured • Interesting avenues for further work 20 Thank you [email protected]