Modeling Transformers in Natural Language Processing Based on Spiking Neural Networks
AmirHossein Ebrahimi
September 2022
1
School of Mathematics, Statistics and Computer Science
Supervisors
Dr. Mohammad Ganjtabesh
Dr. Morteza Mohammadnori
brain of fruit flies
It all starts here.
Image credit: The FlyEM team, Janelia Research Campus, HHMI (CC BY 4.0)
2
natural language processing
Transformers
Chemical Reaction Prediction
Image / VIDEO processing
3
TSNN
Transformer inspired
architecture
based on
Spiking Neural Network
4
Reading Background
5500 years of evolution.
5.1
Delay
Find the optimal delay for each connection
6.1
Delay
Find the optimal delay for each connection
Weight
Find the optimal weight for each connection
6.2
〞
un·break·able
– Morpheme Detection
7.1
Biological Model
Adapted from Cognitive Neuroscience: The Biology of the Mind-W. W. Norton & Company (2019)
7.2
Architecture
8.1
Beahviours
8.2
1
LIF as building block for our biological model.
2
Delay makes it possible to accumulate firing effect.
3
Homeostasis makes the stabilization process easier
5
STDP optimize weights via LTD and LTP
mechanism
4
winner take all simplify the model and its
output
6
Dopamine plays an important role
in reinforcement learning
8.3
9.1
Data Specification
p = 0.9
tau = 7
noise = 0.8
1000 words
Experiments
The longer it takes
for the model to converge,
the darker it grows.
9.2
〞
This show was an amazing, fresh & innovative idea in the 70s when it first aired...
– Sentiment Analysis
10
Architecture
11.1
Behaviours
11.2
1
Population Base Activity regulate long term effect on neuron decision.
2
Stimulus merge and accumulate other population currents.
3
Trace make attention mechanism possible.
5
Condition Reseter separate the simulation in the inference & learning phase
4
Population Winner
is a strategy for output detection in neuron populations.
6
Decision Maker select the worthy neuron population.
11.3
Benchmarks
Adapted from Sentiment Analysis on IMDb benchmark by paperswithcode.com
12
FlyVec
2021
13
Images credit: https://flyvec.vizhub.ai
Images credit: https://flyvec.vizhub.ai
Memristor
2022
Image credit: "Text Classification in Memristor-based Spiking Neural Networks" paper
14
Memristor
sentiment analysis model
Partially bio-plausible
Delay Learning
Generality
Scalability
tsnn
transformer-inspired architecture
NEW
Fully bio-plausible
Delay Learning
Generality
Scalability
FLYVEC
word embedding model
Bio-plausible
Delay Learning
Generality
Scalability
MODELS & COMPARISONS
15
Future Works
Model expansion
Increase the number of neurons in the neural population by increasing the model size layer by layer.
Multi domain
Benchmark the model's performance in additional data domains, such as image using gabor filters or other nlp tasks.
Mapping
Achieve optimal convergence by reordering the mapping function in the first seen of the desired outputs.
16
Thanks
For you attention
STACK & SOURCE CODE
17
tsnn
By Amirhossein Ebrahimi
