Pathways towards truly brain-like computing primitives

Abstract

Taking inspiration from biological information processing in neural assemblies, deep learning and artificial intelligence made solutions to complex problems more feasible. While at first realized by silicon-based hardware, memristive devices for vector-matrix multiplication are attracting huge focus due to their potential integration in large, parallel cross-bar arrays, mimicking the aspect of highly parallel processing (and in-memory processing) in biological neural networks. In real neuron assemblies, however, the connections are not strictly formed in a rigid 2D crossbar array, but form and degrade dynamically in a 3D environment with features of hierarchy, modularity and reconfigurability. To fully explore the capabilities of truly brain-like hardware computing, the transition towards a platform with dynamically reconfigurable connections is mandatory. This work showcases different approaches to address this biological motivation, covering the fields of Topology & Structure as well as Dynamics of biological systems, and classifies them with respect to seven fundamental principles of brain-like computing. The approaches are ranging from highly interconnected nanogranular networks with dynamically reconfigurable connections over liquid-solid composites rearranging connections via dielectrophoresis and guided redox-wiring to the mimicking of neural action potentials by relaxation-type oscillators that are used as input stimuli.