NM500 Neuromorphic Classification of Ships in Satellite Imagery

This Kaggle project demonstrates an application of General Vision's NM500 neuromorphic chip to image classification. The chip implements a Restricted Coulomb Energy / Radial Basis Function network in hardware using 576 artificial neurons. A pattern of up to 256 bytes in length can be presented to the neurons in parallel, essentially realizing a constant-time classification relative to the network size. Furthermore, multiple NM500 chips can be modularly connected into arbitrarily large networks while continuing to incur essentially no increase in computation time. General Vision provides an evaluation board, the NeuroShield, which includes a single NM500 chip. The NeuroShield also accepts stackable NeuroBrick boards, each providing an additional two NM500 chips.

The Kaggle project classifies the ship images with a NeuroShield, using a MicroPython PyBoard to mediate the SPI interface between the computer and the NeuroShield. 97% classification accuracy was achieved, even after reducing the original 19,200-byte images to 256 bytes to fit into the NM500's neural patterns.

Please head over to Kaggle to read the details of this project.

Here are some positive and negative examples from the dataset: