Memristive self-organizing dendrite networks for brain-inspired computing

Introduction

Artificial Intelligence needs a hardware revolution to sustain the ever-growing demand for computing power in our society, where the huge energy consumption and environmental impact of computation with current technologies is unsustainable.

Bioinspired Technologies

In the race toward future computing, bioinspired technologies have been shown as promising hardware solutions for computing beyond the Turing model and the classical von Neumann architectures. Going beyond transistor-centered hardware solutions, the research community is exploring new device concepts and architectures that leverage physical phenomena for computing “in materia” with physical laws to emulate the effectiveness of information processing capabilities of our brain.

Memristive Devices

While arrays of memristive devices realized with a top-down approach represent emerging solutions for the hardware realization of artificial neural networks, these systems do not emulate the topology and emergent behavior of biological neuronal circuits. In biological systems, the principle of self-assembly and self-organization regulates both structure and functions, providing adaptability, efficiency, and robustness.

MEMBRAIN Project Goals

Tackling the main challenges of neuromorphic computing, the MEMBRAIN project aims to develop a radically new concept of physically grounded computing nanoarchitecture based on self-organizing memristive nanonetworks of dendrites. This architecture is intended to efficiently process information and store knowledge on the same physical substrate at the matter level through physical laws.

Vision for Future Computing

Overcoming the concept of nanotechnology as a simple advancement of microtechnology, the ambition is to compute like nature – thermodynamically – to push computation near fundamental limits of efficiency.

Hardware-Software Codesign Framework

By establishing a hardware-software codesign framework at the crossroads of material science, machine learning, and neuroscience, the aim is to retarget the original goal of neuromorphic computing: creating general-purpose truly intelligent systems that endow dynamic learning and multitasking capability.