Dharmendra S. Modha

At Supercomputing 06 conference, the 28th TOP 500 Supercomputing Sites List was released.

Please see the original article at: http://top500.org/lists/2006/11

"IBM BlueGene/L system, installed at DOE’s Lawrence Livermore National Laboratory (LLNL), retains the No. 1 spot with a Linpack performance of 280.6 teraflops (trillions of calculations per second, or Tflop/s)".

The BlueGene/L at IBM’s Almaden Research Center ranked 64th on the list with a Linpack performance of 9.43 teraflops, for details, please see: http://top500.org/list/2006/11/100 

Today, we had a wonderful time with Dr. Srini Narayanan who leads the AI Group at ICSI and is an Adjunct Associate Professor at UC Berkeley.

Here is abstract of the talk:

The UCB/ICSI NTL project has been developing an explicitly neural theory of language. The core premise is that language is largely determined by the computational character of neural networks, the structure of our brains, and our interactions with the physical and social environment. Work within the NTL project coupled with a variety of converging evidence suggests that understanding involves embodied enactment or "simulation semantics".

Simulation semantics hypothesizes the mind as "simulating" the external world while functioning in it. The "simulation" takes sensory input about the state of the world (whether linguistic or perceptual) together with general knowledge and makes new inferences. Monitoring the state of the external world, drawing inferences, and acting jointly constitute a dynamic ongoing interactive process.

We report on a neurally plausible, computational realization of the simulation semantics hypothesis, and on preliminary results from behavioral and fMRI imaging experiments testing its biological predictions.

The core ideas of NTL are captured in a recent book, From Molecule to Metaphor: A Neural Theory of Language, by Professor Jerome Feldman.

At SFN 2006, I learned about an interesting blog Action Potential by editors of Nature Neuroscience.

I am currently at the SFN 2006 Conference in Atlanta, where over 25,000+ other scientists are participating. The conference has featured talks (which are held in a ballroom that can hold 5,000 people), regular presentations (in a number of parallel sessions), posters (which are held in a room which is the size of perhaps 3-4 football stadiums), symposia, mini-symposia, exhibit booths, satellite events, and a variety of scientific socials. The breadth of the topics pursued and the scale of the conference is, to say the least, mind-boggling.

I have met a number of old friends and colleagues, made new conections, and obtained a number of valuable insights for future work.

The conference is definitely dominated by scientists with a bottoms-up approach and is lacking in theorists who put together the whole picture.

In the highlight event for the conference, noted architect Frank Gehry advised us to "search for your own character" but "within the grounded constraints of reality." He explained how he started his career by going back 300 million years to the structure of primitive fish!

In his beautifully articulated acceptance speech for the Peter Gruber Prize, Professor Masao Ito (who is famous for explaining the cerebellar functioning and is credited with the understanding long-term synaptic depression) said that his dream is to eventually explain "unconscious domain of mind."

In a featured lecture, Professor J. A. Movshon described how he and his colleagues uncovered the properties of MT (V4).

In a satellite event, namely, Advances in Computational Motor Control V, Professor Chris Atkeson from CMU showed a number of very impressive robotic demos. His work is based entirely on instance-learning, one-shot learning, or motor tape theory. He openly admitted to having diverged from neuroscience. A question was raised whether the approach is "just a bag of tricks". At the same event Professor Emo Todorov explained his theory of compositionality whereby simple control systems such as linear quadratic regulators and eigen-controllers can be composed together while maintaining mathematical and analytical tractability as long as they share the underlying dynamics. Dr. Jeff McKinstry showed off a demo of a brain-based device that can learn to avoid obstacles by using a model of cerebellum that learns to replace reflexes with a predictive controller, namely, preflexes.

In a symposium on Intenal Models for Sensorimotor Integration, Professor Reza Shadmehr described how to model "living with a changing body", for example, fatigue tires us but vanishes relatively quickly whereas disease changes us for a relatively longer period. He explained several models that use two learning rules: a rule that learns and forgets fast and a rule that learns and forgets slowly. In a fascinating talk, Professor Andrea M. Green described how a monkey (whose body is made rigid) can distinguish between tilts and translations.

It is impossible to succinctly summarize the entire technical experience, but I met people who study saccades, grasping and reaching, neocortical circuits, cerebellar microcomplex, spiking neurons, firing rate neurons, neuro-anatomy, neuro-informatics, robotics, rats in various mazes, cognitive maps of memory, reinforcement learning, hippcampus, striatum, neuro-prosthetics, place cells in hippocampus and related grid cells in the cortex, thalamus, and many others.

A tasty tidbit: umami has now been recognized as the fifth basic taste.

My prediction is that SFN 2007 and beyond will start attracting a lot more (cognitive) computer scientists…at least it should!

My brain and body are tired (the distances are too large, there are too many presentations/posters, the food is terrible, the temperature is too cold, ventilation is not very good, it is hard to navigate through crowds, there is no California sunshine), but my mind is all fired up and intensely happy 🙂 So far, the meeting has been exceptionally energizing and productive.