The Neural Marketplace

Modern computers are more powerful than many ever dared expect. So it is remarkable how much today's computers still can't do. Strangely, some of the hardest tasks for computers are effortless to humans. Problems like vision, natural language comprehension, and walking control will undoubtedly require massive computing power. But the real difficulty is our inability to write down sets of rules that a computer can follow to perform these tasks. The only solution may be to develop computer systems that, like us, learn by example and by trial and error, without needing explicit instructions. The brain contains roughly as many neurons as there are transistors in a modern supercomputer. These cells are computationally more sophisticated than was once believed. But what is most amazing is their ability to organize into large, functionally coherent networks, that constantly learn and adapt to an animal's changing circumstances. This happens with no central point of control, suggesting that something about neurons causes them to automatically assemble into information-processing systems. This fellowship proposal is based on a new hypothesis, derived from neurobiological research, for how this self-organization occurs through competitive processes analogous to those of a market economy. A typical neuron in the cerebral cortex receives about 10,000 inputs, which it integrates to produce a single output, broadcast in turn to about 10,000 targets. Our new hypothesis is for a mechanism by which a neuron receives feedback from its targets, signalling how useful the information it carries is to the rest of the network. Several lines of evidence suggest that in the brain, molecules called neurotrophins can act as carriers of this feedback signal. According to the hypothesis, neurons throughout the brain constantly experiment with new information processing strategies. In most cases, the new information will not be required by the neuron's targets, no feedback will be received, and the neuron will return to its prior state. A few neurons, however, will happen upon information that is useful to the larger network, and will receive feedback causing the recent changes to be retained. In a market economy, interactions like this allow autonomous agents (people and firms) to organize into networks. A firm that makes cars buys parts from suppliers, who buy components from their own suppliers, and so on. At each stage of the supply chain, multiple firms compete to produce the best products, experimenting with new designs that, if successful, will increase market share. The decisions required to build a good car are thus distributed over a large number of agents. No one person has to understand every part of the manufacturing process; instead, decisions made by multiple individual agents cause the system to organize itself. Improvements and adaptations occur by experiments with new approaches at all levels. Scale this picture up, and you have a global economy encompassing billions of individuals. Could similar interactions organize the billions of cells in the brain into a single coherent system? And could they allow us to build scalable learning machines to solve currently intractable problems in computing?The current proposal will answer these questions by constructing a series of increasingly large market-based neural network systems, to solve a series of increasingly challenging tasks from speech recognition and robot control. This research will have impact far beyond these domains, informing the construction of learning systems for applications as diverse as vision and medical diagnosis, as well as to domains such as internet routing that require scalable self-organization of multiple computing devices. Confirming the computational validation of the hypothesis would also provide a step-change in our understanding of how the brain processes information, potentially yielding new approaches to disorders of brain organization.

Computer systems that learn from data, without explicit programming, were once just a dream, but are now an everyday reality. Machine learning has seen an incredible number of industrial applications including: internet search; personalization (e.g. collaborative filtering); targeted advertising and sales; financial market analysis and automated trading; credit scoring; automated customer service; voice recognition; machine vision; quality management; robotics; bioinformatics; and homeland security. The brain has long served as a model for machine learning algorithms. Of course, for industrial applications, whether an algorithm functions similarly to the brain is not important - all that matters is that it be powerful, flexible and easy to use. By these criteria, neural networks have to date had partial success. The backprop neural network has been one of the most successful machine learning systems, as it can deal flexibly with many kinds of data, and is simple, intuitive, and easy to understand. However, there are problems with backprop algorithm that severely limit its industrial applicability: the inability to efficiently train recurrent networks and simulated neurons with intrinsic dynamics; and difficulty scaling to large networks. The consequence of these drawbacks is that backprop cannot deal elegantly with dynamic applications (speech recognition and robot control being two prime examples), and that it is not scalable to high-dimensional complex problems. The current proposal will introduce a completely new idea into artificial neural network engineering, with the potential to correct backprop's two major shortcomings, the inability to train recurrent networks that perform dynamic processing and the inability to make scalable systems. This would therefore open a whole new range of application domains for neural networks including: nonlinear control; robotics; image, sound, and movie recognition; automated diagnosis of biomedical signals; speech and natural language processing; security (e.g. automated cctv analysis); econometrics and finance (analysis and prediction of multivariate time series). While the research of the current proposal is aimed at artificial neural networks, the idea of market-based interactions organizing autonomous agents may have applications to organizing a much wider class of scalable multi-agent systems. The idea of social computing : structuring interactions between autonomous computing agents along similar lines as the organization of humans into companies, societies, and economies, is an exciting recent trend in computer science. The current research will investigate how market-based mechanisms can organize networks of simulated neurons; however the principles we will learn from this may have much wider applicability. This promises applications to a wide number of domains including mobile robotics, computer animation, game programming, manufacturing, wireless networking, internet and telecommunications routing, road and air traffic control, power grid management, scheduling, and sensor fusion. This work also has impacts for public health. One in four Britons will experience some kind of mental health disorder in their lifetime, at a cost to the economy of 77.4bn each year. Rational development of treatments for mental illness cannot occur until we understand the way information is processed in the brain. The neural marketplace hypothesis has the potential to revolutionize our understanding of cortical information processing. Experimental work performed in parallel with this research will characterize the role of retroaxonal signals in vivo, and the underlying molecular pathways. Understanding how neurons organize into information-processing networks will provide key insight into the pathology of mental illness. Discovering the underlying molecular pathways would have a revolutionary impact on drug discovery and other therapies for mental illness.