Towards training robust computer vision models for inference on neuromrophic hardware

Gregor Lenz

TU Delft, 10.3.2023

  • BSc/MSc (2014) in biomedical engineering at UAS Technikum Vienna, Austria
     
  • PhD (2021) in neuromorphic engineering at Sorbonne University Paris, France
     
  • Worked for Prophesee, on Intel's Loihi
     
  • Now machine learning engineer at SynSense

About me

Neuromorphic computing

  • Takes inspiration from biological systems to process information much more efficiently
     
  • Has bottom-up approach to imitation
     
  • Hardware a key factor

Event-based datasets

  • Neuromorphic cameras not easy to come by currently
     
  • Rely on curated datasets to start exploring new data type
  • Data rate from event-based cameras can easily reach 100 MB/s
     
  • Prophesee's automotive object detection dataset is 3.5 TB in size for <40h recording

Efficient compression for event data

https://open-neuromorphic.org

Efficient compression for event data

Efficient compression for event data

https://github.com/open-neuromorphic

Where are we so far?

  • Publicly available dataset
  • Reading from an efficient format that saves disk space
  • Apply transformations that are suitable for PyTorch
import tonic

transform = tonic.transforms.ToFrame(sensor_size=(128, 128, 2), time_window=3000)

dataset = tonic.datasets.DVSGesture(save_to="data", transform=transform)

from torch.utils.data import DataLoader

testloader = DataLoader(
    dataset,
    batch_size=8,
    collate_fn=tonic.collation.PadTensors(batch_first=True),
)

frames, targets = next(iter(testloader))

Training models on event data

  • Could bin events to frames and train CNN / visual transformer on it --> high latency, no memory
     
  • Want to make use of spatial and temporal sparsity --> spiking neural networks
  • Anti-features at SynSense: intricate neuron models, biologically-plausible learning rules
     
  • Focus on training speed and robustness

Leaky integrate and fire neurons

animation from https://norse.github.io/norse/pages/working.html

Training SNNs

  • Focus on gradient-based optimization
    good results
    complexity of BPTT scales badly
     
  • Simple architectures: feed-forward, integrate and fire neurons
    power-efficient execution on hardware
    infinite memory not ideal
     
  • Needs to match hardware constraints
    event-based vs clocked processing

Event-based vs clocked computation

Dealing with event-based data on a time-discretized system

  • Fine-grained discretization
     
  • Reducing weight/threshold ratios
     
  • Spread out spiking activity
     
  • Penalize high amount of spikes in the same receptive field

Noisy event camera output

  • 50/60/100 Hz flicker generated from artificial light sources
     
  • Hot pixels: pixels that fire with >50Hz even though no change in illumination
     
  • Mitigation: augmentation with noise recordings, low-pass filters

Monitoring

  • Firing rates, synaptic operations, temporal density.
    Ideally per neuron!

Monitoring

  • Weight distributions to account for quantization

EXODUS: Stable and Efficient Training of Spiking Neural Networks

Bauer, Lenz, Haghighatshoar, Sheik, 2022

Speed up SNN training

https://lenzgregor.com/posts/train-snns-fast/

Speed up SNN training

Practical steps to SNN training

  • Time investment in setting up the right tools pays off
     
  • Understand your data
     
  • Start with simple model, increase complexity gradually
     
  • SNN models typically have very high variance
     
  • Monitor intermediate layers automatically, also on-chip!
     
  • Identify training bottlenecks (data, backprop, logging?) to iterate faster

Model definition and deployment to Speck

import torch.nn as nn
import sinabs.layers as sl

model = nn.Sequential(
    nn.Conv2d(2, 8, 3),
    sl.IAF(),
    nn.AvgPool2d(2),
    nn.Conv2d(8, 16, 3),
    sl.IAF(),
    nn.AvgPool2d(2),
    nn.Flatten(),
    nn.Linear(128, 10),
    sl.IAF(),
)

# training...

from sinabs.backend.dynapcnn import DynapcnnNetwork

dynapcnn_net = DynapcnnNetwork(
    snn=model,
    input_shape=(2, 30, 30)
)

dynapcnn_net.to("speck2b")

Demo videos

Conclusion

  • Events are a new data format different from images, audio, text
     
  • Real-life challenges include noisy data, resource constraints, async computing paradigm
     
  • We need the full pipeline: sensor, algorithm, hardware

https://lenzgregor.com

Gregor Lenz

Next talks

  • April 4th (online): ONM Hands-on session with Sinabs and Speck
  • Telluride 2023 (in person): Open-Source Neuromorphic Hardware, Software and Wetware

ONM Delft

By Gregor Lenz

ONM Delft

  • 179