Create Sparse Matrix Based Liquid State Machine

This text explains the process of creating a sparse matrix-based liquid state machine using tensorly and torch libraries. It focuses on efficient data processing for intricate datasets such as brain neurons.

use tensorly to create random sparse tensor and eye sparse tensor with ease, which could be numpy only, and requires the sparse package.

scipy, pytorch, tensorflow, jax all support sparse tensor construction but advanced apis are not.

use eye to create bias and input matrix, extract node values. use random sparse tensor to initialize weight matrix. use self matrix multiplication to perform propagation.

the human brain has roughly 87 billion neurons, and every one of them has thousands of synapses.

1
2
3
4
5
6
7
8
import torch
large_number = 1_000_000
torch.arange(large_number).unsqueeze(0).repeat(2, 1)
index_arr = torch.arange(large_number).unsqueeze(0).repeat(2, 1)
val_arr = torch.ones(large_number)
sparse_eye = torch.sparse_coo_tensor(index_arr, val_arr, (large_number, large_number))
# sparse_eye.to('cuda')

alternatively:

1
2
3
4
5
6
import torch
import tensorly.contrib.sparse as tsl_sp
large_number = 1_000_000
numpy_eye = tsl_sp.eye(large_number)
torch_eye = torch.sparse_coo_tensor(numpy_eye.coords, numpy_eye.data, numpy_eye.shape)

Comments

▲