330 ChatGPT designed Recurrent Spiking Neural Network :: Quicker, easier and cheaper to make your own chip!

How it works

This project implements 9 programmable digital recurrent LIF neurons. The neurons are arranged in 3 layers (3 in each). Spikes_in directly maps to the inputs of the first layer neurons. When an input spike is received, it is first multiplied by an 8 bit weight, programmable from a custom interface, 1 per input neuron. This 8 bit value is then added to the membrane potential of the respective neuron. When the first layer neurons activate, its pulse is routed to each of the 3 neurons in the next layer. There are 9x3 programmable weights describing the connectivity between the input spikes and the first layer (9 weights=3x3), the first and second layers (9 weights=3x3), and second and third layers (9 weights=3x3). Output spikes from the 3nd layer drive spikes_out.

How to test

After reset, program the neuron threshold, leak rate, feedback_scale and refractory period. Additionally program the first, 2nd, 3rd layer weights. Once programmed activate spikes_in to represent input data, track spikes_out synchronously (1 clock cycle pulses).

External hardware

List external hardware used in your project (e.g. PMOD, LED display, etc), if any

Memory

The memory block stores 39 words of 8 bits. The first 12 words represent Feedback scale, Refractory Period, Decay rate, Membrane Threshold of each layer ( 4 parameters x 3 layers ). The last 27 words represent the network weights ( 3 ins x 3 neurons x 3 layers ).

Address	Address	Data Index	Data Index	Data Description	Data Description
dec	hex	MSB	LSB	Description	LYR #
0	0	7	0	Feedback scale	1
1	1	15	8	Refractory Period	1
2	2	23	16	Decay rate	1
3	3	31	24	Membrane Threshold	1
4	4	39	32	Feedback scale	2
5	5	47	40	Refractory Period	2
6	6	55	48	Decay rate	2
7	7	63	56	Membrane Threshold	2
8	8	71	64	Feedback scale	3
9	9	79	72	Refractory Period	3
10	A	87	80	Decay rate	3
11	B	95	88	Membrane Threshold	3
12	C	103	96	weight1_0	3
13	D	111	104	weight1_1	3
14	E	119	112	weight1_2	3
15	F	127	120	weight2_0	3
16	10	135	128	weight2_1	3
17	11	143	136	weight2_2	3
18	12	151	144	weight3_0	3
19	13	159	152	weight3_1	3
20	14	167	160	weight3_2	3
21	15	175	168	weight1_0	2
22	16	183	176	weight1_1	2
23	17	191	184	weight1_2	2
24	18	199	192	weight2_0	2
25	19	207	200	weight2_1	2
26	1A	215	208	weight2_2	2
27	1B	223	216	weight3_0	2
28	1C	231	224	weight3_1	2
29	1D	239	232	weight3_2	2
30	1E	247	240	weight1_0	1
31	1F	255	248	weight1_1	1
32	20	263	256	weight1_2	1
33	21	271	264	weight2_0	1
34	22	279	272	weight2_1	1
35	23	287	280	weight2_2	1
36	24	295	288	weight3_0	1
37	25	303	296	weight3_1	1
38	26	311	304	weight3_2	1

#	Input	Output	Bidirectional
0	Input Spike 0	Output Spike 0	out_test 0
1	Input Spike 1	Output Spike 1	out_test 1
2	Input Spike 2	Output Spike 2	out_test 2
3	Spike Input Register Enable	End of Writing Parameters into Memory	out_test 3
4	RSNN enable	Parameter Data Written	out_test 4
5	Serial Data IN		out_test 5
6	Parameter Load		out_test 6
7	Test selection		out_test 7

330 ChatGPT designed Recurrent Spiking Neural Network