Neuromorphic computing

Neuromorphic computing falls in one of my sweet spots. From a theoretical point of view, one of the key questions is to understand the types of computations that can be carried out in dynamical systems, and how the restrictions in architecture affect our ability to implement algorithms. Then there is the problem of how to best implement such systems, and the tradeoff between using new materials or taking advantage of conventional CMOS to design neuromorphic chips.

The insect brain as inspiration to design smart, dynamic sensors

While a lot of focus has been on the brain, my research is heavily inspired by the central neural system of insects, and the integration of sensing, computing, and execution using neural architectures. This can have a tremendous impact for instance in the area of advanced sensors and robotics.

One of the interesting aspects of insects is that, from an evolutionary standpoint, they have already developed a lot of the structural complexity and diversity found in vertebrates, yet the number of neurons and connections is more manageable, in some cases within the range of existing technology.

Finding the right level of complexity

If we want to design neuromorphic systems with capabilities similar to those of insects, one of the key questions is what is what is the level of complexity and fidelity that is required to reproduce the functionality of biological systems.

The approach that I have taken in this area is to step back from the materials and instead focus on understanding the dynamics of these systems from a mathematical point of view. This includes, for instance, exploring the difference between spiking and non-spiking systems in terms of their computational ability, the impact of the network structure on the performance of the system, and the exploration of ways in which systems can display dynamic learning capabilities.

The aim of this exploration is to identify the simplest possible architecture that can execute a given task.

Mapping bio-inspired architectures to hardware

A final aspect of my interest in neuro-inpired computing is how we can translate the dynamic models into hardware. Many of the recent advances in deep learning and artificial intelligence rely on high performance computing capabilities: typically problems are solved on a cluster and the solution then relayed to the user.

My focus is instead on the development of fully autonomous systems where capabilities are implemented locally (i.e. on a chip). This involves understanding the limitations of standard semiconductor manufacturing approaches, the type of devices that would be needed, and the requirements that novel materials must have in order to implement the desired functionality. Techniques such as atomic layer deposition can help us push the limits of 3D architectures and expand the range of materials while being compatible with semiconductor processing.

A fascinating challenge is that biological systems are open systems with a extremely capable biomolecular machinery able to self-heal and modify their performance and architecture over time. In contrast, the architectures that we can design through semiconductor manufacturing are frozen in time. Many of the challenges on making neuro-inpired chips come from the need of bridging this gap.

A 3D printed brain?

One of the advantages of focusing on insect-inspired systems is that the number of neurons is much smaller than that of the human brain: from the 2000 Kenyon cells in the fruit fly’s mushroom body to around 1,000,000 neurons for the whole central nervous system of a bee. If we consider system sizes of the order of 1,000 neurons, we suddenly open up the manufacturing possibilities to implement such systems.

Not only we can leverage foundries targeting 100nm nodes or higher, but we can also start thinking of alternative ways of implementing such a system beyond semiconductor processing. Printing technologies with resolution around 1 micron are now possible, as is the ability to manufacture complex electronic systems on flexible substrates. However, in order to make this possible we have to first understand the design principles behind such systems.

If you want to know more

A summary of our approach on this topic can be found here.

This DARPA program also highlights some of the challenges I am interested in: L2M program.