Scaling-efficient in-situ training of CMOL CrossNet classifiers

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Nanotechnology

Background:

• CMOL CrossNets, a hybrid CMOS/nanoelectronic approach, offer potential for brain-like AI.
• Limited nanodevice functionality challenges conventional training algorithms for CrossNets.
• Existing methods struggle with large-scale datasets like MNIST.

Purpose of the Study:

• To develop effective supervised training methods for CMOL CrossNet-based pattern classifiers.
• To address the scalability issues of current training algorithms for complex AI tasks.
• To enable the practical application of neuromorphic circuits in AI.

Main Methods:

• An in-situ error backpropagation variant was developed for supervised training.
• The method was tested on benchmark classification tasks (Proben1).
• A novel in-situ approach combining training with hidden layer construction was proposed.

Main Results:

• The initial in-situ error backpropagation method successfully trained CrossNets for simple tasks.
• This method did not scale effectively to larger datasets like MNIST.
• Simulated results indicate the new combined in-situ method is suitable for practical classification problems.

Conclusions:

• Novel in-situ training methods are crucial for advancing CMOL CrossNet capabilities.
• The proposed combined in-situ approach shows promise for training neuromorphic circuits on complex AI tasks.
• This research paves the way for more powerful and practical brain-inspired AI systems.