Below is an interview with Dr. Hava Siegelmann conducted by Sander Olson. Dr. Siegelmann is the head of the Biologically Inspired Neurological and Dynamical Systems (BINDS) lab at the University of Massachusetts. More information on her research can be found at the BINDS lab.
Question: You oversee the Biologically Inspired Neural and Dynamical Systems Lab (BINDS) lab at the University of Massachusetts. What is the goal of that lab?
I head the BINDS lab here in Amherst; we have two primary goals: The first is to obtain a computational understanding of natural systems, memory and learning in health and disease. The second is to create computational paradigms similar to the brain’s, to produce increasingly functional simulations of intelligence.
Question: IBM’s Watson recently won the Jeopardy contest. Did that constitute a major AI advance?
IBM’s Watson pushes Turing computation toward its limits using enormous computer clusters and clever programming. But, Watson, like other similar AI demonstrations needs to operate in very orchestrated, specific environments. No matter how resources are increased, a Turing computer’s program is incapable of updating itself to accomplish other tasks or work in an environment other than what it was programmed for. AI needs a new paradigm for a machine capable of intelligent learning.
Question: You coined the term “Super Turing” in 1993. How do you define “Super-Turing”?
In the 1930s, Turing suggested the concept of a universal computing machine that could perform any algorithm it is given. “Super-Turing ” refers to going beyond standard Turing machines. It can be thought of as a series of Turing computers that change at each computational step – it does not need to have theses machines described – it adapts to them to suit input from the environment. Equivalently, unlike Turing machines, Super-Turing’s memory and processing take place together, much more like the brain, so when memory change so does the program. . Super-Turing is sometimes confused with “Hypercomputation,” which describes any abstract/philosophical computational model that differs in any way from the Turing paradigm; Hypercomputation models are not necessarily stronger and are often unrealizable. Super-Turing derives from nature; its existence attests to its viability as a potentially realizable form of computation that may lead to innovations in computation and AI.
Question: How has our understanding of neurons changed over the past twenty years?
Our understanding of neurons and related cells has changed tremendously over the past two decades. In particular, the new theories of reconsolidation have transformed the way we view memory., which used to be seen as a static form of storage, like the one used in Turing type computers.
Question: So conventional beliefs regarding memory creation are inaccurate?
Yes, we now know that every time that we recall something, the process of remembering actually changes memories; these modifications go on constantly; they’re part of the brain’s program.This new understanding of memory parallels a basic Super-Turing attribute; adaption to its environment based on input. This suggests that Super-Turing is appropriate in achieving machine intelligence.
Question: What is your assessment of Henry Markram’s Blue Brain project?
The ultimate aim of the Blue Brain project is to use supercomputers to perform simulations of the human brain. I support this goal strongly as a necessity to increasing our understanding of ourselves and building a knowledge base that can aid our efforts in AI. Blue Brain is by far the most advanced example of detailed brain simulation that currently exists; Markram’s group is at the cutting edge of what we know about the brain.
Question: How accurate are simulations of the brain? Will we ever be able to achieve an accurate and fine-grained simulation of the brain using digital computers?
There are several problems inherent in digital computer brain simulations. Our knowledge of neurons, axons, and dendrites is incomplete; we know, however, that the brain continuously adapts, changes, and evolves based on a continuously changing series of inputs. So a purely physical snapshot of the brain isn’t sufficient to explain the brain’s complex functions; the principles of adaptivity based on continuous environmental input must be part of any useful model. Another problem is that the brain itself operates under analogue principles, which cannot be efficiently simulated using fixed-register, digital machines.
Question: So is the only feasible solution to custom design analogue computers for the task?
There are two potential solutions. You can employ huge numbers of CPUs to digitally emulate the adaptive analogue behavior of neurons; this is what’s being done currently. Or you can use a modest evolving analogue computer to perform the task better. Digital machines can approximate the analogue world only as long as the environment is fixed; but to capture the richness, creativity and dynamic nature of the real, analogue world – a more sophisticated machine is required. Turing was sure that such systems could be developed. It is possible that Super-Turing computation is the missing puzzle piece.
Question: Has an analogue computer capable of simulating the brain been built?
No, the machine would need to be custom designed. The architecture and construction of such a computer would be relatively straightforward in principle, though a monumental task in actuality.
Question: How much would you need to build such a computer?
Consider the billions of dollars and colossal number of hours we have invested in developing digital computers from Turing’s original concept in 1936; an analogue machine could be built for a fraction of that cost. It would also be much better suited to performing intelligent tasks. I am not against digital computing; there is need for reliable, high performance machines, but regarding simulating human brains, an evolving analogue system is clearly superior.
Question: Given sufficient funding, could a computer capable of simulating the brain be designed and built within five years?
No, but with sufficient funding we could make considerable progress within five years in developing the system itself. It will take us a few more years, funding, and collaboration to build a machine that really begins to simulate the human brain.
Question: You were awarded a grant to build the first analogue hardware for a super-Turing machine. How is that progressing?
Yes, we just received an NSF grant to build the first analog Super-Turing computer. The purpose now is to work out the hardware and software aspects; from there, we will move toward more specific applications and studies. We don’t expect to simulate the brain initially, but hope the project will become the basis for just such an effort.
Question: The brain researcher Hugo De Garis is also trying to evolve brains. Have you ever collaborated with him?
I have known Hugo since he was a grad student in the early nineties. We haven’t collaborated, though his work in evolvable hardware has some parallels in Super-Turing’s ability to adapt to its environment.
Question: Given adequate funding, how much progress could you make within the next ten years?
With enough funding, we could make enormous progress within the next decade. We could make much better robots and decision systems; we could expect significant breakthroughs in developing Super-Turing computing, understanding the nature of intelligence and making real advances in AI.