Loading...

Implementation of Spiking Neural Networks on Memristive Crossbar Structure

Bavandpour, Mohammad | 2012

824 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 43552 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Bagheri Shouraki, Saeed
  7. Abstract:
  8. Hardware implementation of Spiking Neural Networks (SNN) can bring about a fast and low cost neural network which has more biological support. Memristor nanodevices are most proposed devices for use as synapses to add dynamic learning to SNNs because of their nano-scale dimensions, low power consumption and memory property. One of the most important bottlenecks in the memristor crossbar based SNNs is system-level simulation of learning process due to its huge memristor equations that should be solved for each sample of time. Due to parallel computation capability of our simulation work, we simulate the circuit on single core CPU and then proposed high performance parallel platforms as Message Passing Interface (MPI) and POSIX® Thread (Pthread) multi core CPU libraries and GPU CUDA platforms and compare performance of these platforms. Some significant agents that determine application robustness of circuit and implementation difficulty and complexity are coding scheme, learning algorithm and spike shape. Spikeprop is a most applied SNN which is so difficult to implement because of its variable delay blocks and its learning algorithm. In the other works usually just digital inputs are considered. It is in the condition that most of the applications need to use neuron in a mixture of analog and digital spaces. In this study, we propose a coding scheme and learning algorithm compatible with a specific spike shape. the proposed coding scheme can be used in digital, analog and combined spaces. Furthermore, we introduce memristor crossbar based circuit for this scheme. the proposed analog coding scheme bring about a powerful spiky neuro-fuzzy clustering algorithm based on proposed digital data coding and its compatible learning algorithm. This neuro-fuzzy system is fully implementable which contains a decoder block to decode analog data to spikes and a memristor crossbar circuit. This spiky neuro-fuzzy clustering algorithm creates a membership function of the cluster on each input feature axis and modify it during the learning phase. This scheme can create various cluster shapes in input space due to its variable membership function shape and this capability brings this scheme to very powerful clustering algorithm. This scheme is based on a proposed analog to digital and digital to spike coding scheme which satisfy algorithm conditions and perform the algorithm properly. Also competitive nature of network makes it more robust to implementation parameters changes. The digital scheme is tested by a clustering task on standard handwritten alphadigit database named Simon Lucas which provide digital input bits. Also the spiky neuro-fuzzy clustering algorithm is tested on Fisher’s Iris standard database which provide analog input data and results show that this circuit can obtain the higher clustering rate in comparison with other similar neural networks with considering that proposed scheme is implementable and have a low computation cost.
  9. Keywords:
  10. Memristor ; Spiking Neural Network ; Graphic Processing ; Spike Coding ; Spiky Neuro-Fuzzy Clustering Algorithm

 Digital Object List

 Bookmark

...see more