The section discussion focused on the basic idea of neural nets - what are they and what they can and cannot do. The list below summarizes some of the key points that came up in the discussion. Please add your own comments or questions to this post:
- The term "neural net" is used to refer to several different things. Some people in cognitive science use the term to refer to artificial learning systems (software programs) that are capable of learning any kind of information by exposure to that kind of information. These computer programs can learn how to recognize human faces (as used by MI5 and other spy agencies) or how to recognize human hands. Problems arise when people assume that the human brain learns in the same way as these systems. Brains consist of a network of connections between neurons but there are more constraints on brain learning - for example, humans easily learn to recognize thousands of faces but it is unlikely that we would ever be able to learn to recognize thousands of hands. The key difference that I emphasize is that brains evolved and software programs did not. Natural selection guided the construction of the human brain to be able to process certain kinds of information very well. Brain neural networks are thus more specialized than articifical neural networks.
- Within the human brain, "neural networks" can refer to many different levels of brain function - this adds to confusion about the term. In this class, we refer to three very high level networks of the brain: recognition, affective and strategic systems. This description of three general areas of the brain comes from one of the grandfathers of neuropsychology: Alexander Luria. (Luria was a student\colleague of Lev Vygotsky in case that name is more familiar.) This high level division is very helpful in thinking about how the brain works and how educators need to attend to and support different aspects of learners and learning. As we always emphasize, these general systems always work in parallel but problems and differences can occur in any one of the areas which will affect overall learning.
- At a lower level, within these general networks there are highly specialized processing areas; sometimes these are referred to as neural nets and sometimes as modules. As we saw in class last night, several areas of the brain "light up" when we hear music, but these areas process different aspects of music - melody, harmony, rhythm, etc. The same is true for listening to speech or reading - many areas of the brain are involved, each performing a specific function. When we look at many individuals, these processing areas seem to appear in the same physical location in all brains, but this is a simplification. Brains develop differently in each person and different occurences during development (brain damage or different home environments) can cause the specialized functions to appear in a different position in the brain. It is best to think of these specialized neural nets in terms of their function - they are predisposed to process very specific kinds of information.
- One question that came up in section was whether these specialized neural networks would activate to other kinds of stimuli. So would areas of the face recognition network light up when you see patterns that are similar, like a doll's face or a dog's. The answer is yes, but what exactly will cause an area to activiate is not always obvious. Current neuroimaging work often finds that some unexpected "neural net" lights up during a task. Faced with this mystery, scientists then try to find out what it is about their stimuli that cause the activation.
- An interesting example of abstract pattern recognition comes from language - specifically speech production. Laura Ann Petitto, a graduate of HGSE, looked at infants who were either exposed to speech or to sign (children of deaf parents). She used a position tracking technology to measure the infant's hand movements when they were interacting in a linguistic" or a non-linguistic mode. By looking at the frequencies generated by the hand movements (measured by the technology), she could differentiation distinct patterns for each mode of hand movement. These patterns are very similar to vocal babbling by infants exposed to language. This finding led her to propose that the language processing networks in the brain are actually processing abstract patterns that are unique to language but not bound to sound alone - hand movements can create the patterns as well. The key articles were published in Science, 1991 and Cognition, 2004 - you can find these using e-resources (ask a librarian for help).
- We did not get down to the nuts and bolts of how learning occurs in the brain. Neurons are initially sensitive to certain kinds of input - a frequency, a dot or a line - but neurons are also massively interconnected to other neurons. Some connections become stronger or weaker as we encounter specific kinds of stimuli again and again. This allows us to learn to differentiate different kinds of things, like phonemes or letters. For a detailed treatment of how neurons learn in this way and how this learning is similar to artificial neural networks, I recommend The Mind Within the Net by Manfred Spitzer. He tends to believe that the brain learns in the same way as artifical nets, which I disagree with, but his nuts and bolts treatment is valuable.
