458 AFM_LIF
458 : AFM_LIF
- Author: Haihang Xia
- Description: Hardware implementation of approximate computing based LIF neuron model
- GitHub repository
- Open in 3D viewer
- Clock: 1000000000 Hz
Approximate computing based Leaky-Integrate-Fire Neuron Model
How it works
This AFM-based LIF neuron model implementation features:
- 16-bit membrane potential
- Fully customizable parameters
- 8 pipline stages
- 8-bit current input
- 8 neurons in one structure
The equition of AFM based LIF neuron model is: $ V[t+1] = AFM!\left(e^{-\frac{1}{\tau}}, V[t]-E_{rest}\right) + E_{rest} + I \ \text{if } V[t] \leq V_{th}, \text{ then } V[t+1] = E_{rest} $
The execution of this architecture consists of three phases:
- Parameter initialization
- Neuron initialization
- Normal operation
How to test
1. Reset (Low to reset, rst_n = 0)
- All the parameter and membrane potential reset to 0.
2. Parameter initialization
- E_REST requires two cycles of input.
- e_tau requires one cycle of input.
- v_th requires two cycles of input.
3. Neuron initialization
- The initial membrane potentials of 8 neurons are input in sequence.
- After waiting for 2.5 clock cycles, the initial currents are input sequentially.
- Each neuron requires 1 clock cycle for initialization.
4. Normal operation
- Input currents sequentially
- Each neuron computation takes 1 clock cycle
- 8 neurons require 8 clock cycles in tot
External hardware
None required - all the operation of this design can be done through digital I/O
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | ui_in[0] | uo_out[0] | uio_in[0],uio_out[0] |
| 1 | ui_in[1] | uo_out[1] | uio_in[1],uio_out[1] |
| 2 | ui_in[2] | uo_out[2] | uio_in[2],uio_out[2] |
| 3 | ui_in[3] | uo_out[3] | uio_in[3],uio_out[3] |
| 4 | ui_in[4] | uo_out[4] | uio_in[4],uio_out[4] |
| 5 | ui_in[5] | uo_out[5] | uio_in[5],uio_out[5] |
| 6 | ui_in[6] | uo_out[6] | uio_in[6],uio_out[6] |
| 7 | ui_in[7] | uo_out[7] | uio_in[7],uio_out[7] |