Code excerpt from David Rotermund, Mahbod Nouri, Alberto Garcia-Ortiz and Kaus R. Pawelzik trying to understand deep NNMF networks.

Origin of the algorithm

Refinement of the approach for deep NNMF networks shown in:

Competitive performance and superior noise robustness of a non-negative deep convolutional spiking network
David Rotermund, Alberto Garcia-Ortiz, Kaus R. Pawelzik
https://www.biorxiv.org/content/10.1101/2023.04.22.537923v1

Now a normal ADAM optimiser will work.

The BP learning rule is taken from here (it was derived for a spike-based SbS system, but it works exactly the same for NNMF):

Back-Propagation Learning in Deep Spike-By-Spike Networks
David Rotermund and Klaus R. Pawelzik
https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/fncom.2019.00055/full

CIFAR 10 Test run on RTX 4090

Last epoch

Epoch: 369
Actual lr:  1.0000e-10  1.0000e-09  1.0000e-10 
Training: Loss=0.01888 Correct=92.36%
Testing: Correct=85.25%
Time: Training=12.0sec, Testing=2.3sec

Note: This is just one run. Even though we have set the seed to a fixed value, the whole network suffers strongly under non-determinism during lerning.

Test performance

Network structure

Note: A block like

  (1): Unfold(kernel_size=(5, 5), dilation=(1, 1), padding=(0, 0), stride=(1, 1))
  (2): Fold(output_size=torch.Size([24, 24]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (3): L1NormLayer()
  (4): NNMF2d(75, 32, pfunctype=0, local_learning=False)

represents one(!) Conv2d NNMF Layer. We just see more of the innards that for a normal Conv2d.

Sequential(
  (0): ReLU()
  (1): Unfold(kernel_size=(5, 5), dilation=(1, 1), padding=(0, 0), stride=(1, 1))
  (2): Fold(output_size=torch.Size([24, 24]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (3): L1NormLayer()
  (4): NNMF2d(75, 32, pfunctype=0, local_learning=False)
  (5): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (6): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1))
  (7): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (8): ReLU()
  (9): Unfold(kernel_size=(2, 2), dilation=(1, 1), padding=(0, 0), stride=(2, 2))
  (10): Fold(output_size=torch.Size([12, 12]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (11): L1NormLayer()
  (12): NNMF2d(128, 32, pfunctype=0, local_learning=False)
  (13): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (14): ReLU()
  (15): Unfold(kernel_size=(5, 5), dilation=(1, 1), padding=(0, 0), stride=(1, 1))
  (16): Fold(output_size=torch.Size([8, 8]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (17): L1NormLayer()
  (18): NNMF2d(800, 64, pfunctype=0, local_learning=False)
  (19): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (20): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
  (21): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (22): ReLU()
  (23): Unfold(kernel_size=(2, 2), dilation=(1, 1), padding=(0, 0), stride=(2, 2))
  (24): Fold(output_size=torch.Size([4, 4]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (25): L1NormLayer()
  (26): NNMF2d(256, 64, pfunctype=0, local_learning=False)
  (27): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=False)
  (28): ReLU()
  (29): Unfold(kernel_size=(4, 4), dilation=(1, 1), padding=(0, 0), stride=(1, 1))
  (30): Fold(output_size=torch.Size([1, 1]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (31): L1NormLayer()
  (32): NNMF2d(1024, 96, pfunctype=0, local_learning=False)
  (33): Conv2d(96, 96, kernel_size=(1, 1), stride=(1, 1))
  (34): ReLU()
  (35): Unfold(kernel_size=(1, 1), dilation=(1, 1), padding=(0, 0), stride=(1, 1))
  (36): Fold(output_size=torch.Size([1, 1]), kernel_size=(1, 1), dilation=1, padding=0, stride=1)
  (37): L1NormLayer()
  (38): NNMF2d(96, 10, pfunctype=0, local_learning=False)
  (39): Conv2d(10, 10, kernel_size=(1, 1), stride=(1, 1))
  (40): Softmax(dim=1)
  (41): Flatten(start_dim=1, end_dim=-1)
)

Information about used parameters:
cnn_top: 14638
nnmf: 173344
batchnorm2d: 576
total number of parameter: 188558

We acknowledge support by the following grants: DFG: Efficient implementation of spike-by-spike neural networks using stochastic and approximate techniques (PA 569/6-1, GA 763/15-1), DFG SPP: Evolutionary optimisation of neural systems (SPP 2205) https://gepris.dfg.de/gepris/projekt/402741184 – Evolution of flexibility - optimisation of task-dependent information processing in the visual system (ER 324/5-1), Era-Net Neuron https://www.neuron-eranet.eu: I-See – Improved intra-cortical visual prostheses through complex coding and integration of spontaneous activity states (BMBF 01EW2104A), Stiftung Bremer Wertpapierboerse https://www.stiftung-bwb.de.