through Neuromorphic Circuits Hongyu An The Bradley Department of - - PowerPoint PPT Presentation

Realizing Associative Memory Learning through Neuromorphic Circuits

Hongyu An

The Bradley Department of Electrical and Computer Engineering Virginia Tech, Blacksburg, VA, USA May 07, 2019

Research Motivation

• H. An.

2

Associative Memory Humans Mammals Invertebrates Apple Aplysia

Tail Siphon

Sensory Neuron Sensory Neuron Response Neuron

Synapse

Sea slug Dogs

Design Methodology

• H. An.

3

Major Premises The Facts of Neural system The Facts of Associative Memory Designs Neurons Synapses Neural Network Results Biological Response Circuit Response Mutual Corroboration

Neurons and Synapses

4

• E. R. Kandel, Principles of neural science vol. 4: McGraw-hill New York, 2000.

Brain Neuron structure

Soma (cell body) Axon Dendrites Synapse Synapse

mV

Axon

mV

Neuron structure model

• H. An.

Neurotransmitter Postsynaptic neuron terminal Presynaptic neuron terminal synapse Presynaptic Spiking Signal 40 mV 0 mV

• 70 mV

Threshold Postsynaptic Spiking Signal

• 55mV
• 70mV

Threshold

Synapse functions:

• Transfer signals between neurons
• Attenuate the spiking signals
• Synaptic strength of transmission

can be modified

Neural network

Cellular Level Associative Memory in Sea Slugs

• H. An.

5

• E. R. Kandel, et al, Principles of neural science vol. 4: McGraw-hill New York, 2000. [2000 Nobel Prize]

Sea Slugs Sea Slugs Before associative memory learning

Unpaired Stimulation

Tail (US) Siphon (CS)

Gill motor received signal

Cell responses Before training After training Siphon

Tail

Conditional Stimulus (CS) Unconditional Stimulus (US) Sensory Neuron Synapse

Response Neuron Sensory Neuron

Gill Siphon sensory neuron

5 min

2 mV 50 ms 5 mV

Unconditional Stimulus (US)

Tail

Siphon Sensory Neuron

Sensory Neuron Response Neuron

Synapse Paired Stimulation

Siphon (CS) Tail (US)

Siphon sensory neuron Gill motor received signal

Before training After training Conditional Stimulus (CS)

2 mV 50 ms 5 mV

5 min

Gill Siphon Shelf Tactile Stimulus Tail

Experimental Setup After associative memory learning

Small Signal Larger Signal

Design Methodology

• H. An.

6

Major Premises The Facts of Neural system The Facts of Associative Memory Designs Neurons Synapses Neural Network Results Biological Response Circuit Response Mutual Corroboration

• Spiking Signals
• Signal Attenuation
• The synaptic strength becomes strong during

the associative memory learning

• Large signal at postsynaptic neuron indicates

a successful learning

Current Starved Ring Voltage Controlled Oscillator

The switches RC Oscillator 𝐽𝑒

Signal Intensity Encoding Neuron

• H. An, et al., "Monolithic 3D neuromorphic computing system with hybrid CMOS and

memristor-based synapses and neurons," Integration, the VLSI Journal, 2017.

• Spiking signal generation;
• Positive and negative outputs;
• Magnitude and frequency corresponding to the input.
• H. An.

7

Memristor as Synapse

8

TEM image: J.-Y. Chen, et al., "Dynamic evolution of conducting nanofilament in resistive switching memories," Nano letters, 2013.

1. The synapse should have the capability of attenuating signals; 2. The connecting strength of synapse is adjustable with a set voltage.

• H. An.

Set voltage

Design Methodology

• H. An.

9

Major Premises The Facts of Neural system The Facts of Associative Memory Designs Neurons Synapses Neural Network Results Biological Response Circuit Response Mutual Corroboration

• Spiking Signals
• Signal Attenuation
• The synaptic strength becomes strong during

the associative memory learning

• Large signal at postsynaptic neuron indicates

a successful learning

• Signal Intensity

Encoding Neuron

• Memristive Synapse

Reproducing Cellular Associative Memory Learning

b a

Neuron B1 (Tail) Neuron A1 (Siphon)

S1

Negative Spiking Signals Positive Spiking Signal

Unconditional Stimulus (US)

tail

Siphon Sensory Neuron

Sensory Neuron Response Neuron

Synapse Conditional Stimulus (CS) Paired Stimulation

Siphon (CS) Tail (US)

Siphon sensory neuron Gill motor received signal

Before training After training 5 min

2 mV 50 ms 5 mV

• H. An.

10

Associative memory behavior occurs Signals superpose together Set voltage Higher current after learning Lower current

• E. R. Kandel, Principles of neural science vol. 4: McGraw-hill New York, 2000.

High Level Associative Memory Learning System

• H. An.

11

The brain

Response

Central Nucleus

Tone Stimulus Shock Stimulus Tone signals Shock signals Auditory Neural Networks Somatosensory Neural Networks

Lateral Nucleus

• Distinct types of signals are preprocessed at the different

regions of brain

• The outputs signals after the preprocessing converged at

Lateral nucleus

• E. R. Kandel, Principles of neural science vol. 4: McGraw-hill New York, 2000.

Associative Memory Learning System

• H. An.

12

Audio signal of digit number “3” Visual signal of digit number “3”

Response Neurons

Memristive associative neural network layer

Signal Preprocessing Phase Association Phase Outputs Input

Spiking Signal Transformation Layer

Artificial Neural Network Artificial Neural Network

Unconditional Signal Pathway The brain

Response

Central Nucleus

Tone Stimulus Shock Stimulus Tone signals Shock signals Auditory Neural Networks Somatosensory Neural Networks

Lateral Nucleus

Input signals of SIENs and the response current with the auditory and visual signals of digit number 3

After Learning Learning Before Learning

Synaptic resistances with no learning (MΩ) Synaptic resistances after learning (MΩ)

Synaptic Weight Updating in Associative Memory

• H. An.

13

Conclusion and Significance

• Implement a brain-like associative memory learning that relates the pronunciation (auditory

signal) and image (visual signal) of digits together by associating two artificial neural networks

14

• H. An.

Brain Associative memory learning Model

Scientific Contributions:

• Potential explanations regarding the

human learning mechanism

• Potential interpretations of

memory/forgetting mechanism

• Diseases: Alzheimer’s disease and

visual agnosia

Explanation Engineering Contributions:

• Human-Like self-learning capability
• High adaptivity with dynamic surrounding

environment

• Novel Human-computer interaction system
• Spiking Signal based power efficient system

Reverse Engineering

Hongyu An

Virginia Tech hongyu51@vt.edu

Thank You