The Neuro SICS Laboratory, with its NEAR project, is the second RWCP laboratory at the Swedish Institute of Computer Science, located in Stockholm. SICS is a cooperative effort of Swedish government and industry, and has close ties to universities. SICS has been an overseas partner of RWCP since 1993 when the PIRAIA project was established at the Novel Functions SICS Laboratory.

The research of the NEAR project began with the work of Kanerva on Sparse Distributed Memory (MIT Press, 1988) and is outlined in the book's last chapter. It was first conducted at Stanford University and at the NASA Ames Research Center in the United States. RWCP funding of the Neuro SICS Laboratory in 1994 made it possible to continue the research at the Swedish Institute of Computer Science.

When we think of flexible computing, we naturally think of what people can do, and how that differs from what we can program computers to do. People's ability to learn and to use language is a striking example of this difference. Other examples include the learning and exercise of skills; coping in nature; memory and recognition of people, things, places, and events; dealing with ambiguity; and acting on intuition. These abilities represent the working of the brain, and so brains are our ultimate model of flexible computing. The fundamental research questions of the NEAR project are:

1. What gives brains their power? 2. How to build that into artificial systems?

The project's titleﾅgNeural Encoding And Representationﾅhreflects the main premise of our research: To understand the brain's powers, we must understand how brains encode and represent information. Discovering the principles that govern the brain's representations is a deeply mathematical problem.

We share the popular view that a representation is a pattern of activity over many neurons. Neural-net research is done mostly with small nets, and reducing the dimension of the pattern space is a standard goal. Such research is an extension of traditional statistics and logic, and it can produce useful applications, just as statistics and logic can. However, it runs the risk of overlooking the real power of neural nets.

The main mathematical insight in our research is that large systems have nonobvious ﾅgemergentﾅh properties on which new computing algorithms can be based. Apparently brains use such algorithms. Brains have vast numbers of neurons, and the details of the brain's construction are overwhelming. However, the vast numbers are much more important than the infinite detail, as suggested also by the statistical law of large numbers, which smooths out individual differences. Consequently, algorithms that are imprecise and unreliable in low dimensions (i.e., with small patterns in small neural nets) converge to precise and reliable algorithms in high dimensions, resulting in computational behavior that is not possible on small deterministic machines or on small neural nets. Statistical mechanics is an example of this idea in physics. The NEAR project's focus is on methods that depend on large pattern size in large neural nets.

Another major mathematical insight is that the pattern components can be simple. The important computational properties of large systems are present even if the system's components--theﾅgneuronsﾅh--are binary. This simplifies both the analysis and the engineering of the systems without sacrificing their power. The mathematical methods of studying such systems include the geometry of high-dimensional spaces, coding theory, information theory, probability and statistics, signal processing, control theory, and computer simulation.

Recent research results include:

1. Fast activation algorithm for Sparse Distributed Memory (SDM). The original SDM algorithm finds active locations by massive computation, and large simulations require parallel hardware (e.g., the Connection Machine). The new algorithm removes this bottleneck to large simulations.

2. Boosting the signal of SDM, and recognizing and separating highly correlated patterns stored in SDM. The methods are statistical and work only with large patterns. They are therefore well suited for the SDM, and they improve its storage capacity, noise tolerance, and resolution.

3. Spatter Coding of sets in terms of their elements. Recursive encoding of sets is the first step in representing compositional structures, such as language, in neural nets. Hinton, Plate, Pollack,and others have encoded compositional structures into neural nets. The spatter code is a new and particularly natural (ﾅgneuralﾅh) way of doing it.

Current and future research and development include:

4. Statistical properties of networks of spiking neurons. Most neural-net research abstracts away the spiking of neurons. Our preliminary results indicate that, in addition to spiking frequency, the statistics of spiking can encode other useful information. This view is current also in neurobiology.

5. Encoding and combining sensory data (data fusion). Normal people experience the world as an integrated whole, with the senses complementing each other. How the nervous system accomplishes this, is an open question.

6. Extending the spatter code to ordered sets and sequences. Such coding is necessary for understanding higher mental functions and intelligence in physical (neural) terms.

7. Paper design of Pattern Computer. The practical goal of our research is to help design computers that employ the brain's operating principles. We want our research to make a contribution to how computers are built and used in the future.

The RWCP Neuro SICS Laboratory includes Dr. Gunnar Sj in as project leader, Dr. Pentti Kanerva as principal investigator, Dr. Roland Karlsson, Dr. Jan Kristoferson, Dr. Anders Lansner (part time), and Seppo Pohja.