Human brains can perform extremely complex tasks with very low energy budgets. In contrast, digital computers consume a lot of energy when running cognitive-type tasks. AI and deep learning models seem to conform better to neuromorphic and analogue platforms. Neuromorphic Analog Architecture (NAA) is emulating human brain operation and neurons, so that the deep learning model can be modelled by mathematical simulation of neurons and implemented in analog hardware. This combination of neuromorphic computing and analog implementation greatly improve the power efficiency of the AI chip.
Deep learning models have millions of parameters, which are encoded in bits. Binary neural network (BNN) targets to minimize the required memory and complexity of the deep learning model by quantizing the original multi-bit (e.g. 32-bit) parameters to 1-bit parameters. This low bit neural network reduces both storage and processing resource, which can greatly accelerate the inference process of the deep learning model. The performance of the BNN can be further maximized by implementing the model with the Neuromorphic Analog Architecture. This combination can provide quick inference response with low power operation.
