Skip to content

bars_stripes.md

Latest commit

 

History

History
134 lines (113 loc) · 4.88 KB

bars_stripes.md

File metadata and controls

134 lines (113 loc) · 4.88 KB

The Bars and Stripes problem, and a RISP solution

This problem comes from the paper "An FPGA-Based Neuromorphic Processor with All-to-All Connectivity" [Maheshwari23]. You are given an input of r x c black or white pixels, and you wnat to determine whether the input has at least one vertical bar or horizontal stripe. Here are four examples:

img/bs_example.jpg

In [Maheshwari23], they give a simple spiking neural network that performs the calculation. There are r x c input neurons, one for each pixel, and you spike in all of the black pixels at time zero. There are two output neurons -- one for bars (vertical lines) and one for stripes (horizontal lines). You run the network for three timesteps, and at the last timestep, each output neuron spikes if there is a bar or stripe.

Script to create the network

There is a shell script, scripts/bars_stripes.sh, that creates a RISP network for the bars and stripes problem. You give it the number of rows and columns as command line arguments, and it outputs the RISP network (which is a RISP-1+ network modified so that thresholds are higher, and leak is on) that solves the problem.

Here's an example of creating a 2x3 network:

UNIX> sh scripts/bars_stripes.sh
usage: sh scripts/bars_stripes.sh rows cols
UNIX> sh scripts/bars_stripes.sh 2 3 > tmp_network.txt
UNIX> ( echo FJ tmp_network.txt ; echo INFO ) | bin/network_tool
Nodes:         13
Edges:         17
Inputs:         6
Outputs:        2

Input nodes:  7(I[0][0]) 8(I[0][1]) 9(I[0][2]) 10(I[1][0]) 11(I[1][1]) 12(I[1][2]) 
Hidden nodes: 6(H-B2) 2(H-S0) 3(H-S1) 4(H-B0) 5(H-B1) 
Output nodes: 0(O-Bar) 1(O-Stripe) 
UNIX>

As you can see, the input neurons start at neuron 7, and outputs are neurons 0 and 1 (for bar and stripe). Let's see it in action. Suppose the input contains a bar in the last column:

001
001

That corresponds to spiking neurons 9 and 12. Let's do that and then run the network for three timesteps. If all is working, neuron 0 will spike at timestep 2:

UNIX> ( echo ML tmp_network.txt
        echo AS 9 0 1
        echo AS 12 0 1
        echo RUN 3
        echo OT ) | bin/processor_tool_risp
node 0(O-Bar) spike times: 2.0
node 1(O-Stripe) spike times:
UNIX>

Nice -- let's do a stripe too -- we'll spike neurons 10, 11 and 12:

UNIX> ( echo ML tmp_network.txt
>       echo AS 10 0 1    11 0 1    12 0 1
>       echo RUN 3
>       echo OT ) | bin/processor_tool_risp
node 0(O-Bar) spike times:
node 1(O-Stripe) spike times: 2.0
UNIX>

A script to test bigger networks

The script in scripts/bars_inputs.sh takes four parameters -- a bars-and-stripes network, the number of rows and columns, and the starting input neuron. It reads a rectangle of rows x cols pixels, and then it creates the processor_tool commands to run the network on the input.

I have the four examples from above in the files txt/7x7-bar.txt, txt/7x7-stripe.txt, txt/7x7-both.txt and txt/7x7-none.txt. Let's run them:

UNIX> sh !$ sh scripts/bars_input.sh
usage: sh scripts/bars_input.sh network rows cols starting_neuron < pixels
UNIX> sh scripts/bars_stripes.sh 7 7 > tmp_network.txt                          # Create the network
UNIX> ( echo FJ tmp_network.txt ; echo INFO ) | bin/network_tool
Nodes:         65
Edges:        112
Inputs:        49
Outputs:        2
                                   # We can see from the next line that neuron 16 is the first input.
Input nodes:  16(I[0][0]) 17(I[0][1]) 18(I[0][2]) 19(I[0][3]) 20(I[0][4]) ...
Hidden nodes: 6(H-S4) 11(H-B2) 7(H-S5) 15(H-B6) 3(H-S1) 14(H-B5) 2(H-S0) ...
Output nodes: 0(O-Bar) 1(O-Stripe) 
UNIX> sh scripts/bars_input.sh tmp_network.txt 7 7 16 < txt/7x7-bar.txt     # The script produces processor_tool commands
ML tmp_network.txt
AS 21 0 1
AS 25 0 1
AS 28 0 1
AS 35 0 1
AS 38 0 1
AS 42 0 1
AS 49 0 1
AS 55 0 1
AS 56 0 1
AS 58 0 1
AS 63 0 1
RUN 3
OT
UNIX>                 # We run it on the four inputs to see that it works properly.
UNIX> sh scripts/bars_input.sh tmp_network.txt 7 7 16 < txt/7x7-bar.txt | bin/processor_tool_risp 
node 0(O-Bar) spike times: 2.0
node 1(O-Stripe) spike times:
UNIX> sh scripts/bars_input.sh tmp_network.txt 7 7 16 < txt/7x7-stripe.txt | bin/processor_tool_risp 
node 0(O-Bar) spike times:
node 1(O-Stripe) spike times: 2.0
UNIX> sh scripts/bars_input.sh tmp_network.txt 7 7 16 < txt/7x7-both.txt | bin/processor_tool_risp 
node 0(O-Bar) spike times: 2.0
node 1(O-Stripe) spike times: 2.0
UNIX> sh scripts/bars_input.sh tmp_network.txt 7 7 16 < txt/7x7-none.txt | bin/processor_tool_risp 
node 0(O-Bar) spike times:
node 1(O-Stripe) spike times:
UNIX>

Reference

[Maheshwari23] D. Maheshwari, A. Young, P. Date, S. Kulkarni, B. Witherspoon and N. R. Miniskar, "An FPGA-Based Neuromorphic Processor with All-to-All Connectivity," IEEE International Conference on Rebooting Computing (ICRC), 10.1109/ICRC60800.2023.10386808, 2023, pp. 1-5.