BitBrain and Sparse Binary Coincidence (SBC) memories: Fast, robust learning and inference for neuromorphic...

Area of Science:

• Computational neuroscience
• Information theory
• Sparse coding

Background:

• Current AI models require extensive training and computational resources.
• Neuromorphic and conventional architectures present challenges for efficient, adaptive learning.

Purpose of the Study:

• To present an innovative working mechanism (SBC memory) and infrastructure (BitBrain) for fast, adaptive learning and robust inference.
• To enable efficient implementation on neuromorphic and conventional hardware.
• To demonstrate high classification performance with minimal training.

Main Methods:

• Developed SBC memory storing feature coincidences from training data.
• Combined SBC memories into BitBrain for diverse feature coincidence analysis.
• Implemented on SpiNNaker neuromorphic platform and evaluated on MNIST/EMNIST benchmarks.

Main Results:

• Achieved high classification accuracy with single-pass learning, comparable to state-of-the-art deep networks.
• Demonstrated robustness to noisy and imperfect inputs.
• Showcased efficient training and inference on both neuromorphic and conventional architectures.

Conclusions:

• SBC memory and BitBrain offer a unique combination of single-pass, single-shot, and continuous supervised learning.
• The system is highly efficient for edge and IoT applications requiring fast, robust inference.
• This approach significantly reduces training costs and parameter tuning compared to deep networks.