An Interview With Professor Tobias Gemmeke

  Professor Tobias Gemmeke Copyright: © Iris Schümmer

"neuroAIx" promises to find a new, more efficient way of computing

Over the past 60 years, semiconductor technology has evolved at an unprecedented scale, benefiting all of us. This development has laid the groundwork for impressive innovations in many areas – including machine learning, often referred to simply as AI. However, according to Moore's Law, the underlying technological advancement is reaching physical limits threatening further progress. RWTH sees this challenge as a chance to drive innovation – with the help of neuroAIx: The name is derived from a combination of elements – or more specifically, the domain of AI is framed by "neuro” on one side and some additional ingredients, the "x," on the other. While "neuro" refers to information processing akin to biological neuronal networks, the "x" alludes to hardware innovations that complement existing manufacturing processes in the chip industry with novel material systems and, thus, unique component properties. In summary, RWTH envisions a convergence of the different domains, where the teams encourage and inspire each other, exchange methods, and establish new common approaches in order to solve the major challenges humanity is facing. "neuroAIx" thus links the trifecta of neuromorphic computing, machine learning, and neuroscience to bring about disruptive developments in hardware, software, and wetware. Opportunities and challenges were recently discussed at a Frontier Workshop. In the following interview, Professor Tobias Gemmeke, head of the Chair of Integrated Digital Systems and Circuit Design (IDS), explains what the current state of research is and what perspectives were addressed at the workshop.

Professor Gemmeke, will the computer as we know it soon become obsolete?

Gemmeke: No, the conventional computers we know today will not disappear any time soon, and the industry will continue to innovate in this area because we need conventional computer technology for many applications that will continue to exist.

But current computer technology has its limitations, and you are taking a new approach with neuroAIx. What will that be?

Gemmeke: "neuroAIx" targets a different area of application that existing computer technology cannot sufficiently address. We are talking about Big Data – vast amounts of uncertain and fuzzy data. This is where conventional computers reach their limits. We are working on an efficient solution for working with Big Data – efficient in terms of energy used and speed in processing. My vision is to draw inspiration from how the human brain works. We don't want to copy it, but we want to understand the underlying processing principles. Armed with this knowledge, we then want to work on processing the flood of data more efficiently. Kind of in the same way that a jet aircraft is not a copy of a bird but still utilizes the basic principles of flight.

What would such a computer look like?

Gemmeke: The cornerstone in the efficient development of new applications – so-called apps – is the clear separation into software and hardware development based on the fixed description of the hardware functions, namely the Instruction Set Architecture, or ISA for short. We initially want to circumvent these limitations to break new ground in application and algorithm development that will enable us to find significantly more efficient solutions – not least thanks to new components at the lowest technological level. Essential to this, of course, is the inspiration we hope to draw for new algorithms from analyzing and synthesizing the cognitive processes of the human brain. The knowledge accumulated in this area is impressive. However, we still lack a clear understanding of the "middleware" – the connection between cell-level processes, such as individual ion channels, and the resulting behavior, our human ability to recognize and explain relationships. With our multidisciplinary team, we want to boost the research precisely in this area. We will pool our expertise from the fields of hardware, materials science, algorithm development, and neuroscience, as well as from medicine and biology, to help us solve crucial problems. In this way, we will create a powerful formation of outstanding research flagships that are united in their purpose instead of individual flagships in separate disciplines on their own.

What is it that connects you together?

Gemmeke: We will be united in our interest to look beyond our own noses. At RWTH, we have ideal conditions for solving major challenges as a team. By networking, we also overcome existing problems – such as the very different intervals at which technological advancements occur in the various disciplines. In information technology, we experience disruptive developments every three months. But it often takes years to design a new chip – and even longer for new components. In Aachen, we are creating a new type of vertical integration of all those involved, which still allows everyone to act at their own pace but ensures that we regularly connect and thus work towards our common goal.

How will neuroAIx computer technology affect our everyday lives?

Gemmeke: Obviously, the goal is not to come up with the umpteenth banking app. We aim to develop applications in which we can process large amounts of data, even those that are noisy (roughly linear). Autonomous driving is a very popular application field where neuroscience approaches are already being pursued. Human drivers manage to focus their attention on all the relevant data. Similarly, modern systems must filter the data stream early on and not send everything to the cloud. That's a given for data protection reasons. But let's take another example: Telephone communication.

What opportunity do you see here?

Gemmeke: The progression from the rotary dial telephone to the smartphone could only happen thanks to electrical and information technology. What’s more, if our field continues to develop at such a rapid pace – which is to be expected – then according to some studies, electronic data processing will consume as much energy in 2040 as we now generate worldwide. This inevitably means that technological innovations can no longer be made available to everyone due to a lack of energy reserves. Is that what we want? Of course not. So we need more energy-efficient ways to enable future transformations such as the one we saw from rotary dial telephones to smartphones.

We need this energy efficiency in many places...

Gemmeke: Yes, we have now all realized that something has to happen. But the advantages are far greater: More skillful calculation requires fewer steps and thus not only less energy, it gives us answers faster – helping avert dangerous situations – and it buys us more time to address other questions. It also eliminates the need for caching or inefficient communication with the cloud – I say inefficient because it is slow and energy intensive. "neuroAIx" ultimately offers us the promise of a new, more efficient way of computing.