Sleep cycling for Probabilistic Generative Models

The media have lately been full of excitement about progress in Artificial Intelligence. Not only can computers now beat humans at Go, and detect cats in Youtube videos, but soon we will have robots in the house, self-driving cars, and many jobs might become automated. Apart from the societal challenges that this revolution will bring, many hurdles are still to be overcome before Artificial Intelligence will obtain truly human-like capabilities.

In particular, current artificial systems might be very good at specific tasks, they cannot easily apply their processing power to other problems. Moreover, in order to become experts in a certain problem these machines often need millions of training examples. Current AI systems follow a strategy very different from humans and obtain their strength from brute compute power and massive amounts of data rather than by cleverness. This is also the reason why it is hard to communicate with these machines, understand their decisions and instruct them. The fact that computers use an approach that is so different from that used by humans seriously hinders application of AI to real world applications. The research community is well aware of these issues, and it generally believed that the problem arises because machines don't construct higher level understanding of the problems that they are solving. How this should be addressed is however not clear.

In humans and animals sleep plays an important role in creating high level representations. During sleep, the brain consolidates information, rearranges it, finds links between different types of knowledge, reformulates problems, and comes up with creative solutions. Most people have experienced this at some point - as they feel better able to solve a problem after a good night of sleep. It is only very recently that researchers have become able to manipulate the processes that go on during sleep, and thereby pick apart the roles of the various sleep phases play and the reason why the sleep phases are ordered in a particular way.

Here we propose to research how to processes occurring during sleep can be mimicked in computational models, and thereby open the possibility to build more human-like artificial systems.

By examining whether it is feasible to mimic the cognitive processes that occur during sleep in computational models, the research is in the first instance designed to benefit both computational and cognitive researchers.

For computational researchers, there are benefits for those who seek to build better artificial systems, including academic researchers such as those grouped in the Alan Turing Institute, and the rapidly growing community of machine learners and data scientists employed in government, such as GCHQ, and industry, such as Google Deepmind in London. These researchers will profit by having insight to algorithms that process in a more human-like fashion, systems that use higher level and hence more compact ways to reach decisions, and systems that need less training data and can transfer knowledge between domains, and thereby increase energy efficiency, widen application domains, and reduce development cycles.

For cognitive researchers, the main benefits will stem from the increased understanding which stems from creation of a good computational model. The importance of the SWS and REM sleep stages, and the interleaving of these, for memory consolidation is too complex to understand without a computational model. To date, no such model exists, and we are unaware of any group working to develop one. Thus, the model which we propose to develop will be hugely beneficial to studies in this area, as it will make explicit testable predictions about how these sleep stages interact with memory. Beneficiaries of this will include academic researchers interested in memory consolidation, sleep, and the combination of these fields, as well as researchers interested in human development, and cognitive reasoning.
Our access to these communities, including the iCub project, is detailed in the Pathways to Impact attachment.

There is an emerging market for technology that can enhance our memory, and a quickly growing crop of start-up companies aiming to do this through manipulation of sleep. Because our research will lead to a better understanding of how memory replay in each sleep stage, and in fact at each time in the night, impacts on consolidation, it will directly benefit these companies. For example, PZIZZ Ltd. is enhancing sleep through delivery of sounds on an mobile phone app, Rythm Inc. and Deep Wave Technologies Inc. are companies seeking specifically to enhance slow wave sleep. We are already working directly with these companies (for details see Pathways to Impact).

On a longer time-scale, medical researchers are also interested in sleep, as many cognitive problems are linked to sleep problems. These communities will profit from both the insights gained by our experiments, and the computational framework that will act as a source of hypotheses and subsequent experiments.

Finally, the training of the RAs and the exposure to each other's complementary research will impact the research landscape in the UK and help to address the current lack of computationally trained cognitive scientists. The cognitive researchers involved in this project will learn computational tools which they could implement in the future in other laboratories, including those that develop therapeutic devices to modulate various sleep stages. The computational researchers in the project will be exposed to applications of their research to cognitive paradigms. The training delivered in the context of this grant will thus contribute to the UK's competitiveness and further emphasize the UK's prime position for integrating computer and cognitive sciences.