Summary prepared by Selena Kang (additions and comments by Frawley)
[Prof. Chester looked at the nature of computing and struggled with the question: Are WE computers?" Suppose that by WE we mean NEURONS. Are we then computers? Probably so. But do WE scale up from THEM? Perhaps what neurons do well, WE also do well: so the 7+/- 2 limit on short-term memory may be related to the limit on oscillatory patterns in neurons. What do neurons not do well?]The initial focus of Professor Northmore's lecture is the nature of computation and creating models. How can neurons be computing devices? The fundamental theory of how the brain works comes from a strictly "bottom-up" approach. Here, we start from the biological basics and build up from them, creating levels of organization within the nervous system.
Marr has proposed that that a brain's job -- and the job of a brain explainer, a cognitive scientist -- is like a computer's job; it specifies a task, an algorithm, and an implementation. These levels can work independently of each other. This theory, however, is more suitable for a computer than the human nervous system because when a neuron takes in or puts out impulses, everything depends on how synapses decipher individual parts. There is an intimate relationship between levels in the nervous system and levels in the explanation.
Still computing models can help us because these models can help us bridge the gaps in our empirical investigations. Many models have been made in the past. One example is the AI approach, where algorithms of how the brain works are computed. Another example is cybernetic thinking, which is a control theory that serves as an error correction system. Still another theory is the connectionist approach that deals with units resembling neurons that start off simple but are then strung together to perform complex tasks.
An important modeling strategy is the bionic approach, which uses evolution's real solutions in artificial modeling. This approach is realistic. Units are hooked together for realistic use and implemented into realistic representations (how data is handled in the nervous system.)
To demonstrate our necessity of models, consider the fact that we are able to see neurons and study them from a microscope, but in many cases they are dead and non-functioning. (And single-neuron studies are just that: single!) This is far too selective a view. We are not able to understand how the cells function, and need to attempt to make realistic neuron models.
Note how neurons lend themselves to this kind of bionic modeling. There are direcitonal selective cells that respond to motion in a specific direction. Cell response can be converted into an electrical model and calculated as a funciton of capacitance and resistance. From this we can construct a model of how a cell should respond in cases where we cannot observe directly. Models thus provide insight. By building sufficiently realistic models, we can connect them into networks and hope they will function like the brain.
Neurons consist of soma, dendrites, and axons. Neurons function much like resistors, capacitors, and switches which may turn the cell on and cause it to fire an impulse, or turn it off. There is an excitory switch which bumps up the voltage that diffuses in time along the dendrite. There is also an inhibitory switch which turns off soma. There are many impulses that occur in different time and places along the dendrite tree. If potential exceeds a certain level, impulses fire onto dendrites of other trees.
This way of talking leads readily to silicon-chip modeling of neurons. Circuits of resistors, capacitos, and on-off switches can model dendrites. Chips allow scientists to pack large numbers of information into small components, and chips can be constructed down to the size of a neuron, but not to the synapse.
Experimentations on dendrites consist of exciting it with spikes, then recording it on graphs. Each excitation drives voltage to its maximum. Summation is less complete the closer in time the spikes get and nonlinearity occurs. Summation is a combination of the input into a neuron.
There are many things we can accomplish with such neural models. We can produce neurons tuned for a particular frequency, interval, coincidence, and correlation. That is, we can make neurons that directly reflect known properites of biological neurons.
[We can also engineer neurons that are not isomorphic to known biological neurons in organization, but still produce the proper output. For exmaple, directional selectivity can be gained by putting excitatory neurons before inhibitory ones, but this is not what the brain does. So, is Turing's criterion enough for intelligence here?]