The figure shows the basic "information flow" diagram
for an organism that can interact with an environment and through that
interaction modify its synaptic connections so as learn about the environment.
The system shown is an example of a system with the capacity to form a
"two-fold tangled hierarchy". I use the term tangled hierarchy
here in the sense defined in Godel, Escher,
Bach (by Douglas Hofstadter) to describe how the brain gives
the appearance of containing functional units of neurons that produce concepts
in a recursive fashion. All the silly arrows that are inside the organism
in the diagram represent the flow of signals through the nervous system.
There is an outer loop from environment to sensory array to motor output
and back to the environment that represents the organism's ability to interact
with the environment. A response by the organism to the environment can
change the environment or at least the organism's interaction with the
environment. This feedback loop to the environment is the first tangle.
The second tangle is all internal to the organism. Information flows from
the sensory array to the motor output system, but there are also feedback
pathways internal to the brain's information processing system that allow
for complex regulation of the flow of information. It is a challenge to
determine the minimal set of network architectures and rules for experience-dependent
modification of synaptic connection strengths that can be embedded in a
biologically plausible model to allow for learning about the environment.
How are the strengths of connections at synapses (represented in the diagram
by all those arrows) altered by experience so as to build a human mind?
For fifty years attempts have been made to construct models of brain function
using computers. A dominant strategy has been to create formal model neurons
that capture the basic electrophysiological properties of biological neurons
and combine a group of the formal model neurons into a model neural network
in an attempt to create a model that will capture some aspect of brain
function such as the ability to track eyes over a visual scene (for example,
see the work of Christof Koch).
I propose that it will be useful if we modify this modeling approach in
two major ways.
First, I am interested in models that capture more of the realities of
the biological basis of brain development and learning and memory. We need
to expand our formal model neurons to reflect a greater complexity and
capacity for change through time. The reason for making this change is
that cellular and molecular biology are revealing the mechanisms of change
in the nervous system, and good modeling should always incorporate the
available experimental results. Secondly, this approach in no way contradicts
the dominant approach which uses simple model neurons. Any model containing
complex model neurons can be decomposed and reconstructed as a model with
only simple neurons in a slightly more complex network. An important point
is that these models with very simple model neurons will contain network
features that have no parallel in biological brains. As a biologist, I
am more interested in how biological brains work than in exploring "neural"
network models that just happen to be easy to put into computers. But these
two approaches are complementary.
Hypothesis: any neural network that does something truly interesting
will be so complex that we need to first model the developmental and learning
algorithms that will allow the mature system to form in response to the
interaction of the model with a complex environment rather than try to
construct a complete functional model of the mature system from scratch.
The "developmental and learning algorithms" that are used by
biological brains are now being revealed by molecular and cellular neurobiology
research. We need to incorporate these new experimental results into theoretical
neurobiology.
Second, I feel that it is time to move away from past attempts to model
individual sub-systems in the brain one at a time. I think that for the
features of human brain activity that are of real interest (like language
use or "consciousness"), most of the nervous system as well as
a complex environment should be included in the models. This need for including
more in our models is based on systems theory. Would anyone really expect
to be successful in modeling an economy by first constructing a model of
the producers and a separate model of the consumers with the hope of some
day combining the two into a coherent model of the economy? In particular,
I view the memory mechanisms of the brain as having to be modeled as integral
components of our network models. Crick said, "Consciousness Now",
I say, "don't forget the memory".
Hypothesis: The interaction of a human organism with its environment
is very complex. A functioning brain does not form in the absence of that
interaction. A model of a mature, functioning human brain can best be constructed
by allowing a plastic, changeable infant's brain model to interact with
a complex environment.
In practical terms, how can we hope to make models of whole brains when it is a major task for individual research groups to construct models of even simple brain systems? My hope is that the Internet will allow for a coordinated effort by many research groups scattered around the world. It is one thing to display systems like GENESIS (used to simulate large neuronal models and realistic neural networks) on the WWW, it is another to have a coordinated system for linking many such systems into a coherent whole brain model. In addition, we need a coordinated system that more efficiently channels the experimental results of cellular and molecular neuroscience into theoretical neurobiology. Systems like the Interactive Fly , the Whole Brain Atlas, and Fly Brain are examples of ways of putting neuroscience knowledge on the web, we need to adapt this approach to a more interdiciplinary goal. There is good evidence to suggest that hypertext networks on the internet are viable systems for organizing human knowledge. The Human Brain Project will hopefully grow into an open project, but for now, it is still just getting started and mainly an insiders game: "It is expected that researchers funded by different grants under the Human Brain Project will communicate; coordination and collaboration across different Human Brain Project grants is strongly encouraged." It is still up to individual research groups to select the degree to which they will coordinate their activities with eachother.
I am particularly interested in bringing cellular and molecular neurobiology research closer to the classical neural network modeling approach. As stated in the PROGRAM ANOUNCEMENT for THE HUMAN BRAIN PROJECT: PHASE I FEASIBILITY STUDIES:
"As scientific specialization increases, integration and synthesis of different types of data and about different aspects of brain structure and function become increasingly difficult. Models can serve as information spaces in which experimentally obtained data of different types and from different sources can be integrated and synthesized."
Funded projects like:
Neural Plasticity: Data and Computational Structures, headed by Michael Arbib at the University of Southern California: "The research supported by this grant will lead to an array of informatics tools to store, visualize, retrieve and model complex neuroscientific data obtained through a variety of methodological approaches and at multiple levels of detail. These tools will be developed in the context of research on neural plasticity, and will be available over electronic networks."
sound very nice. Only time will tell if they grow into a resource that can be used by all.
Does it make sense to avoid the complexities of the visual cortex and try to model the olfactory system?
Here is the type of : neural network model that I have been working with.
Go to John's Book Page.
Go to John's
Home Page.