Brain-inspired Computational Memory

ongoing

Deep neural networks (DNNs) have transformed the field of AI in recent years. However, a significant challenge persists in the form of inefficient hardware implementations of DNNs. Computation in memory (CIM) is an emerging approach that tackles the processor-memory divide in modern computing systems, enhancing their suitability for DNNs. CIM draws inspiration from certain computational principles found in the human brain, such as hard-wired neural networks and analogue processing. A key question is whether other attributes of information processing in the mammalian brain could help overcome conventional CIM challenges and lead to a radically enhanced variant of CIM. In biological brains, information is holographically encoded, without distinctions between significant bits, potentially allowing computing-in-superposition on the same hardware. Biological neural networks also use randomness constructively for information representation and computation. Time is a valuable computational resource in the brain, used for information representation and processes. Unlike current CIM-based DNNs, biological neural networks operate in three spatial dimensions. BRICMEM aims to develop a revolutionary brain-inspired computational memory by leveraging superposition, randomness, time, and space. For computing in superposition, the goal is to store multiple DNN models within the same computational memory. Randomness will be systematically used to enhance conventional DNN functionality and to develop new types of DNNs. Temporal encoding will be explored as an energy-efficient communication method between CIM cores, with novel mathematical functions operating on elapsed time. Additionally, we will utilize the third spatial dimension to improve weight capacity and to implement specific neural networks. Experimental validation will be conducted using current and future CIM prototypes developed by my team. If successful, BRICMEM could redefine the landscape of AI hardware.