A spiking neural network module, tempotron for classification
tempotron learning rule inspired by this thesis: The tempotron: a neuron that learns spike timing–based decisions
And explained detailed by this:
Learning real-world stimuli by single-spike coding and tempotron rule
Tempotron concept:
A type of spiking neural network, an implementation to simulate the work of brain neurons. Comparing to traditional neural network, SNN is more closer to the operation of human brain neurons.
Work in spiking neuron:
When a neuron accept a stimulate from pre-neuron, it's store in neuron. As long as the stimulate accessed exceed a threshold of neuron, the neuron generate a spike, that passed to the post-neuron. After the spike, the neuron "sleep" for a while and it's storage value return to rest potential(may be zero). The neuron stay silent for a period of time, within the interval don't accept any input stimulate.
A neuron accepts all the input stimulate from pre-synaptic neuron, store the stimulates in it's potential, called 模電位。
V(t) = sigma( W_i * sigma(K(t - t_i)) ) + V_rest
K: kernel function, value locate in [0, 1]. Output of K represents the contribution of spiking happend at t_i.
The latest spiking occur in t_i, then the contribution of that spiking during current timestep t is K(t - t_i). Kernel function
K(t - t_i) is calculated by:
![image]
在此式中，spiking發生後，Kernel function 隨時間的前進而變小，spiking t_i的影響力隨時間遞減。tou_m 和 tou_s為超參數。
V0為normalization function, 將kernel function值限制在 0 和 1 內。此式中 t_i 必小於 t，因為只有過去產生的spiking會對膜電位產生貢獻。
在膜電位觸發spiking後，neuron會進入一段時間的「不應期」，而potential會返回復位電壓V_rest。
Spiking neuron的輸出只有「觸發」和「不觸發」兩種。若欲使用SNN輸出多類別: 1. 使用binary編碼，在output neuron上做binary decode產生十進位的output value. 舉例來說，可用五個output neuron表示二進位輸出。若輸出為00000代表1, 00010代表2.
在這個簡單的實現中，使用三個神經元代表三個類別。代表類別的神經元產生脈衝，則表示樣本為該類別。其它類別的輸出神經元則抑制。我們用這種方式訓練數量同等於類別個數的二元分類器，使SNN可分辨三種類別。
Model training: 若樣本為類別B，而輸出神經元A發出脈衝，連接此神經元A的synaptic weight需更新權重。training目標為仰制輸出神經元A的膜電位。反之則目標為增強輸出神經元的膜電位。
Input encoding: SNN將輸入資訊視為神經脈衝，因此我們必須將numerical value編碼為時間單位的輸入刺激。這裡使用Gaussian function做為encoding function. 對於每一個輸入值x, 用N個gaussian neuron來編碼，輸出即為此input value所造成的input neuron spiking time. 假設一個input對上12個gaussian神經元，每個gaussian神經元皆為相隔一小段距離的gaussian函數。
![image]