Professional Documents
Culture Documents
01/24/2022
Arizona State University
Motivation of Project
Fewer hardware implementations of SNN
[1] Diehl, Peter U., and Matthew Cook. "Unsupervised learning of digit recognition using spike-timing-
dependent plasticity." Frontiers in computational neuroscience 9 (2015): 99.
[1] https://github.com/peter-u-diehl/stdp-mnist
Diehl, Peter U., and Matthew Cook. "Unsupervised learning of digit recognition using spike-timing-dependent
plasticity." Frontiers in computational neuroscience 9 (2015): 99.
.
Venkataramanaiah, Rai, Nayyar EEE 598 Neuromorphic Computing Hardware Design
Arizona State University
Synapse weights:
Hardware Architecture
Clk
rst
Membrane Weights
potential Top
threshold Memory Neurons
Ex. weights module
In. weights
Spike
Neuron Model:
Fire
Spike_in
Neuron_sel
Ex_wt
Σ +
0
- Spike
In_wt en Vth
Integrate
Weight selection
Counter
Refractory Counter
Membrane_V
Spike_train
Ex Ex
In In
Ex Ex
In In
Ex Ex
In In
Memory
Membrane_V
Spike_train
Ex Ex
In In
Ex Ex
In In
Ex Ex
In In
Memory
Membrane_V
Spike_train
Ex Ex SPIKE!
In In
Ex Ex
In In
Ex Ex
In In
SPIKE!
Memory
Membrane_V
Spike_train
Ex Ex
In In
Ex Ex
In In
Ex Ex
In In
Memory
700
600
500
Spike Count
400
300
200
100
0
0 23 46 69 92 115138161184207230253276299322345368391
Neurons
Handwritten digit 1
99.89
7 1
Handwritten digit 28
Obtained image from spike rate
Percentage of weighted spike counts
9 6 4
Handwritten digit 4
Obtained image from spike rate Percentage of weighted spike counts
9 4
Handwritten digit 2
Obtained image from spike rate Percentage of weighted spike counts
2.21
34.02
52.24
11.39
1 2 3 5 6
0.25 500000
Power in watts
0.2 400000
Gate count
0.15
300000
0.1
200000
0.05
0 100000
8 10 12 16 20 32
0
number of bits 8 10 12 16 20 32
Number of bits
Power Dynamic Power Leakage
Summary
Key highlights
– Implementation of the entire network architecture in hardware
– Coherence between hardware and software classification
Challenges:
– Analyzing the entirety of the MNIST dataset (60000)
– Processing 1184 (784 +400) inputs to the neuron
– Test-bench generation
Future work
– Exploiting the sparsity of input image to reduce area and power
– Good Calibration of hardware
References