Spiking Neural Network Modeler (SpiNNMo)

What is a Spiking Neural Network?

In the machine learning community, spiking neural networks (SNNs) differ from traditional neural networks. Spiking neural network operates on spikes. These spikes are discrete events taking place at specific points of time. Therefore, it differs from Artificial Neural Networks that use continuous values. In the event of a spike, differential equations represent various biological processes.

The membrane capacity of the neuron is one of the most critical processes. A neuron spikes when it reaches a specific potential. Differential equations represent various biological processes in the event of a spike. After a neuron spike, the potential is again re-established for that neuron. It takes some time for a neuron to return to its stable state after firing an action potential. The time interval after reaching membrane potential is known as the refractory period.

Identification and Significance of Innovation:

Emerging neuromorphic technologies are expected to cause a transformative change in how edge computing is performed in low-power environments. However, there are currently no systems that allow neuromorphic engineers to build high-level abstractions of neuromorphic processors that can accurately predict the accuracy, latency and power use for a given network processing a given input data set. SpiNNMo could fill this important gap and allow engineers to test and profile their networks on a range of existing and hypothetical neuromorphic systems.

The Need:

It is increasingly difficult for non-experts to transition to spiking neural network development, which is limiting any type of hardware adoption. Device manufacturers need a way to determine which of server neural network evaluation chips meet their speed, accuracy, and power requirements.

  • Develop a generalizable neuromorphic hardware emulator.
  • Predict the accuracy, latency, and power usage performance of existing and new neuromorphic hardware platforms.
  • Flatten the learning curve for transitioning to neuromorphic hardware by allowing the user to develop network architectures in a familiar environment.
  • Speed the development of intelligent edge computing for drones, buoys, and IoT sensors.

Our Solution:

ChromoLogic has developed a generalizable Spiking Neural Network Modeler (SpiNNMo). SpiNNMO predicts the accuracy, latency, and power usage performance of existing and new neuromorphic hardware platforms. Our technology flattens the learning curve for transitioning to neuromorphic hardware by allowing the user to develop network architectures in a familiar environment.

SpiNNMo offers a cloud-based solution for developers to test spiking neural networks on a variety of low-power neuromorphic hardware platforms to quickly validate these emerging computing platforms for their specific application.

SpiNNMo Aids in Hardware-Specific Architecture Search

  • SpiNNMo allows users to examine tradeoffs between power, accuracy, and execution time.
  • SpiNNMo answers which architecture achieves >99% accuracy with the least power used.
  • SpiNNMo answers how much could our latency decrease with a larger power budget.
Help-Desk