Stochastic Assemblies in Spiking Neural Networks

Cognitive processes in the brain are based on activity patterns in a huge network of billions of neurons interconnected through trillions of synapses. How the interaction of neurons in this vast network gives rise to our mental capabilities is still mysterious. However, important insights have been obtained recently. First, it has become clear that an entity or concept such as a specific person or a well-known place is encoded in cerebral cortex through the activation of an assembly of millions of neurons distributed over large parts of cortex. This coding stands in stark contrast to the encoding of items in standard digital computers. Second, neural activity and synaptic communication is stochastic, i.e., neurons and synapses are seemingly unreliable, and in consequence, neuronal assembly codes of a given entity vary from trial to trial. In the project Stochastic Assemblies in Spiking Neural Networks, we will investigate stochastic assembly formation and computation in a highly integrated joint effort. We will perform experimental work on organic neural networks and theoretical work as well as simulation of computer models for biological neural networks. Our findings will have important implications for many application areas, from which we will intensively investigate two within the course of the project. First, our improved understanding will provide new concepts for the interpretation of brain activity that will be utilized to develop novel concepts for neuroprosthetics. Second, we will develop novel computing concepts based on stochastic assembly codes. It has become clear in recent years that this standard architecture (the von Neumann architecture) is severely limited in terms of its efficacy and scalability. In other words, it will become increasingly difficult if not impossible to continuously improve the performance of computers based on this architecture while keeping its power consumption at an acceptable level. Given the cognitive capabilities of the brain and its remarkable power efficacy, the distributed assembly organization of the brain seems to provide a solution to this problem at least for the cognitive functions that humans perform routinely.

This interdisciplinary research project involved a partner in Neurobiology (Unversity of Antwerp) and Neuroinformatics and Brain-inspired Computation (TU Graz). We started from two key discoveries: (1) neurons process information in concert with other cells, as assemblies, and (2) neurons and synapses are intrinsically unreliable, operating non-deterministically. Our goal was to rethink earlier theories on how computations are organized in the brain. Both biological and artificial neuronal networks were studied. Based on our findings, our goal was to develop concepts for computation and learning in novel brain-inspired computers. It is well-known that neurons in the brain show rhythms, exhibiting fast oscillations between low and high activity. We analyzed the role of such rhythms in neural networks. When we think, we can rapidly switch between different interpretations of our sensory input, and we can quickly check different alternatives when we try to solve a problem. We found that rhythms can improve such search for solutions. Without oscillations, networks tend to get stuck at specific interpretations and don't consider alternatives. This phenomenon disappears in oscillating networks. This finding can potentially explain the abundance of oscillatory activity in the brain from a computational perspective. Recent experimental findings show that concepts (such as a car, or a person) are not represented in the brain by single neurons but by rather large groups of neurons, so-called assemblies. Such assemblies were also found in the experiments of our project partner. Also, when we associate two concepts - e.g., when we remember that Joe Biden (concept 1) is the president of the USA (concept 2) - the assemblies for these concepts start to merge. We developed neural network models which can explain how such assemblies emerge from sensory experience and how they can merge when concepts are associated. Next, we investigated whether higher-level functions, such as the capability of humans to reason at an abstract level and to structure information into abstract categories can be implemented by such assemblies of neurons and by associations between them. We developed a neural network model where assemblies could represent either sensory content or more abstract structural information about that content. We found that the attachment of structural information to content emerges in this model through changes of synaptic connections between neurons. This provides a basis for the implementation of more demanding cognitive computations by biologically plausible neural network models. Finally, we investigated whether memory stored in assembly associations can enable neural networks to solve complex tasks. We showed - for the first time - that a biologically plausible network model can solve a demanding set of question-answering tasks on stories. These results open up a new research direction where artificial learning systems are extended with biologically plausible memory mechanisms.