MemBrain: A Memristor based, 40 Watts, Biologically Inspired, Human Scale Brain

In MemBrain, we will develop computation in memristor (CIM) structures to implement a biologically plausible model of cortex called Bayesian Confidence Propagation Neural Network (BCPNN).

MemBrain aims at implementing a human-scale BCPNN, in a form-factor and power-consumption level, that is comparable to human brain.

BCPNN has the benefit of being an abstract model grounded in and constrained by neurobiology, implementation friendliness and comes with unsupervised and reinforcement-based learning.

MemBrain would enable human-like intelligent, compact and power-efficient machines with un-supervised learning capabilities and the potential for higher-order cognition.

Previous research on using CIM structures to implement the single-neuron, leaky-integrator and fire models with Synaptic Time Dependent Plasticity rules are not suitable for mapping the complex inference and learning equations of BCPNN. In MemBrain, we will manufacture a small test-chip to implement a scaled down version of BCPNN.

The test-chip will functionally validate the MemBrain design and also serve as the basis for validating the quantitative  objectives of MemBrain: human-scale BCPNN, in the form-factor and power consumption level, comparable to human brain, has been achieved or not. Success of MemBrain will enable design of exa-scale custom super-computers to liberate and accelerate science.

Finally, the energy efficiency of MemBrain will lead to a sustainable path for high-performance custom computing.