Brain-inspired computing and learning

The last decade has seen a boost of artificial intelligence techniques, mostly driven by progress in the field of artificial neural networks. However, there are also some fundamental concerns. Despite huge leaps forward, true human-like intelligence remains out of reach. In order to become very good at single tasks, deep learning techniques are relying on ever larger networks, which are extremely data-hungry and even more power-hungry. In contrast, the human brain is capable of solving many different and much more difficult tasks, all at the same time, and consuming about as much power as an old-fashioned light bulb. In addition, although biological learning takes much longer, the number of examples we need for learning is much smaller than in artificial neural networks. In short, the efficiency gap between digitally implemented artificial neural networks, which was in place from the early days, has only grown larger.

At IDLab-AIRO, we try to address these issues from a number of perspectives. On the one hand, we are delving deeper into the (possible) mechanisms behind biological learning.

Biologically inspired learning For a long time, the study of biological learning systems (biological neuroscience) and the modelling of the computational processes that may be going on (computational neuroscience) remained separate research fields from Artificial Intelligence. However, recently, new bridges between both worlds are being explored.

However, biological brains and digital computers are completely different computational substrates. What is (energy) efficient in one can be horribly inefficient in the other. In close collaboration with biological and computational neuroscientists, IDLab-AIRO tries to identify abstractions of biological mechanisms that can be efficiently translated into digital hardware. This research was initialised through our participation in the European Human Brain project and is now continued in the European Marie Curie SmartNets project .

Post-CMOS computing Although digital computing is ubiquitous, it is in fact horribly inefficient for sensor processing applications. Most sensors capture (often noisy) analogue signals. These Analogue quantities are then translates into rows of bits (often with a precision that is much higher than the original signal) and processed by millions of tiny devices that operate on binary signals. These devices themselves are in fact built out of analogue devices (transistors) which are driven to operate only at two voltage levels. When all this is done, the digitally represented results are often converted back into analogue signals, e.g. to serve as control signals or to drive loudspeakers.

The main reason for omnipresence of digital computing is the robustness of the digital abstraction. Once a signal is translated into its digital representation, logic gates and flipflops can be made such that they make (almost) no mistakes, even in the presence of noise. This makes the automated design of hugely complex systems relatively easy. And thanks to the CMOS technology, power consumption of digital systems remained within acceptable bounds.

However, the continued downscaling of digital technology is shaking these strengths. At extremely small scales, transistors become “leaky”, which increases power consumption, and less robust. For this reason, research into alternatives, in the form of analogue computing, has recently gained in relevance. The main bottlenecks are the areas where digital computing was excelling: robustness and design methodology. Again we look at the brain for inspiration. IDLab-AIRO has a long standing tradition in the field of physical reservoir computing, which exploits the natural dynamics of physical systems for computation, thanks to machine learning. However, the complexity of tasks that can be solved with reservoir computing is limited. In the European Marie Curie project Post-Digital we are exploring new computational paradigms that maximally exploit the natural dynamics of physical anallogue systems. The approaches we envisage build upon and extend reservoir computing, while allowing for a systematic design methodology that can be scaled up to solve more complex tasks.