February 1, 1999

Continuation of the last lecture, plasticity of the brain in adult animals.

There was a slide of a monkey's hand with indications of surgical procedures and the corresponding brain maps associated with the digits in the sensing area. The interesting point is that these maps are somewhat plastic in the adult. They change under certain circumstances in adults. For instance, if the monkey's third digit is amputated, the parts of the brain tuned to digits 2 and 4 will increase their territory. Digit 3 is not providing any input anymore, so the neurons previously responding to digit 3 start responding to digits 2 and 4. If you take a piece of skin and its nerves from one digit to another, then the receptive fields also reorganize, so that many of the neurons which were paying attention to cut segment, now pay attention to the transplanted piece of skin. A third example: digits 3 and 4 were sutured together. Some of the cells responding to digit 3 and 4 receptive fields begin to respond to the common area. This is true even after the digits are separated. These cases show that neurons which get no input, start to respond to other things. Also, if there is an enormous amount of stimulation to a particular area, the part of the brain responding to that area get bigger. This was shown with another experiment in which a small part of the monkey's finger was continuously vibrated.

In another set of experiments, part of a cat's retina is destroyed. When you destroy part of the retina, the coresponding part of the lateral geniculate becomes permanently silent. In visual cortex, the story is more complicated. The experiment concerns 12 cels mapped across a section of the cortex. Cells 5-10 had receptive fields in the retinal area which was destroyed. All but one of these found new input to respond to. The cells in the cortex are more plastic than the cells in the lateral geniculate. There was a slide showing diagrams of the receptive fields of the 12 cells in cortex. The fields have different orientations. Pre-lesion, these fields have different sizes and different orientations by angle. Immediately after the section of retina is destroyed the cells 5-10 get no input. Immediately, cells 5, 10, 11 get a huge receptive field. Some cells which were further away were unaffected. This happens instantly and the growth of neurons is much slower than that.

This increase in receptive fields comes from the fact that the cells 5-10 are no longer getting inhibition from neighboring cells. Cells with nearby receptive fields have collaterals which activate inhibitory neurons, which inhibits neighbors. If one of the cells stops functioning, it stops providing inhibition. In the last lecture, we saw that when the brain is wired up, it is over-connected. There are more connections than are useful at a given time. The way it tunes itself has a lot to do with this inhibition; cells which are focused on a particular thing inhibit cells which might provide a confusing messages. When there's damage, because of this redundancy that's built in, nearby cells can take over some of the function. So this recovery ability is pre-wired.

After some weeks in the retina story, the some cells go back to receptive fields similar to their original ones. The tuning process increases the inhibition of cell 5 by other things it's competing with, so the receptive field of 5 goes back to its original size. Cell 9, which had no receptive field immediately after the lesion, has a different receptive field at the same orientation as its original field. There's a degree of plasticity and a degree of pre-wiring in the neural system. These lectures haven't yet covered normal learning.

Cognitive Psychology *necker cube slide with conn. model here* A slide was put up with the Necker cube (after Swiss naturalist, Louis Necker), a drawing of a cube with labeled vertices which could be seen in two different perspectives, with the A vertex closer or with the G vertex closer. If you keep looking at the cube, the perspective will flip. Why does that happen? This course is concerned with learning what is it about the brain that makes our perception of this cube flip, among other things.

Mutual inhibition is part of this. If there is some unit which is active when A is closer than G and another that is active when G is closer than A, these units will be mutually inhibitory. In the slides, lines with circles on the ends are inhibitory connections; those without are mutually excitatory. The computational details of spreading activation and negative activation in this and other models in this lecture are not given. Those details will be covered later in the course. This model also has units which represent certain vertices being hidden, units which represent certain visual pieces of the cube (A vertex is a convex "Y" or G vertex is a convex "Y") and "grandmother" units which represent whether the perspective is the A closer or the G closer cube. These units may be distributed all over the brain, but there must be convergence of these unit activations at some point if we are to decide that we see one perspective or another. The system in the slide models that convergence.

Note: the term "unit" is used in this course to refer to either an neuron or an element of a model which could represent a group of neurons. At the level of discussion for this lecture, the distinction doesn't matter. It's the computational primitives that are important.

The basic idea is that some psychological phenomena, which are hard to describe in standard computational terms, can be done with parallel models. The brain is inherently parallel and uses spreading activation, so if you want to understand thought and language, you need to make models which explicitly take this into account. The rest of the lecture reviews more such models.

Examples:
Bob threw a ball.
Bob threw a ball for charity.
Bob threw a ball for charity but he missed the clown's nose.

For the basic sentence, there is one reading of "ball". If "for charity" is added there is another reading of "ball". If "but he missed . . . nose" is added, then the original reading of ball comes back. The slide is a model of how this might happen. It involves the lexical level, word senses and case roles.. Throw can mean 'propel' or 'sponsor' and "ball" can mean 'sphere' or 'dance'. Sentences have to also be analyzed at the case level or the role level. There must be an agent of propelling or sponsoring and an object of propelling or sponsoring. It turns out that sphere can be propelled but not sponsored and dances can be sponsored but not propelled, so you get mutual inhibition. The computational story is identical to the Necker cube: multiple levels, mutual inhibition of competing interpretation, mutual excitation of consistent interpretation, and rapid spontaneous flips, which are affected by context. This implies that one interpretation inhibits another, but the other interpretations are activated temporarily.

There are lots of ambiguities: Rose - flower or stood Straw - hay or something to sip through The chicken is ready to eat. - The chicken is well done, or it wants its dinner. Visiting relatives and be boring. - The relatives are boring or the active of visiting them is boring. If there is context for an ambiguous word, we don't find it ambiguous. In the sentence 'They all rose,' only 'stood' occurs to us as the meaning of "rose", but in the brain, the other meaning 'flower' is also activated for a while. Psychological experiments provide evidence for this. In such experiments, people are given a cue, such as a sentence "They all rose," and they are asked to decide if "flower" is a word of English. In such a case, if "flower" is shown immediately after the cue, the subject can decided faster that it is a word than is an unrelated word, such as "desk" is tested.

Prime
They all rose

Target with no delay Reaction Time
flower 685 ms
stood 677 ms
desk 711 ms

Target with 200ms delay Reaction Time
flower 659 ms
stood 623 ms
desk 652 ms

If the target is presented after a 200ms delay, the priming effect for 'stood' is about the same, but the priming effect for 'flower' is gone. The phonetic input activates all the words with the same sound. If the subject is tested in time, the activation of any word with the same phonetic representation will prime related words. After a certain amount of time, however, one interpretation is chosen and mutual inhibition suppresses the other interpretation. If words are tested with no context, there will be frequency effects; the subject will chose the interpretation with is more common to him/her.

You can also get priming at levels at which you wouldn't expect it. In English there are two dative constructions: double object (John gave Mary the book) and prepositional dative (John gave the book to Mary). We have to choose one or the other of these in speaking, so they mutually inhibit each other. With "give" you can choose either one, but with "donate" you can only choose the prepositional dative. This can be primed. A picture is shown in which a woman is showing a dress to a man and a prime is given which is unrelated to the picture, but which has either the double object or the prepositional dative form. An example of a prime would be "The governess poured a cup of tea for the princess" (prepositional dative) or "The governess poured the princess a cup of tea" (double object). A statistically significant number of subjects will use the same construction in their description of the picture as was given to them in the prime. If the node for prepositional dative, for instance, is activated, it may prime the next sentence produced to have the same construction. It is unclear, however, if the priming comes from the syntactic form itself, or the meaning of the grammatical construction, the pairing of grammar and roles, such as topic = agent or topic = patient.

One of the key results in connectionist modeling involves word superiority. The experiment involves an identification task in which an A or B is flashed quickly on a screen and the subject must decide which letter it was. Subject actually perform better on this task if the letter is in the context of a word. This is called the word superiority effect. It was an important finding for cognitive psychology because it shows that processing in the brain can't be serial. Otherwise, we would not be able to process more information faster than less. This important because it shows how your way of thinking about phenomena depends upon your underlying computational model. The people who solved this problem used spreading activation models, which were one of the converging things which led to work recent work in connectionist models. We think of these models as directly related to the spreading activation in the brain.

The detailed story of how word superiority works involves pretty much the same elements as previous models: multiple levels, mutual inhibition and mutual excitation. In the model, feature nodes send activation to letter units which would have those features and inhibition to units which could not have those features. A horizontal bar sends activation to the A, but not to the N. There is good reason to believe that such feature detection occurs in the visual system. The letter units also mutually inhibit each other; A inhibits N, etc. This is a kind of bottom-up processing. But there is also top-down evidence, which allows us to explain the word superiority effect. The extra information from the top level helps resolve the conflict about which letter is seen. The brain works in parallel, so it can process three letters as quickly as one. If a word with A, such as "apple" gets activated, then it provides evidence for the A which make the choice of A easier than the choice of B. This processing occurs unconsciously. This was a breakthrough in cognitive psychology.

Another model by Gary Dell, et al. explains speech errors. A slide of the model was put up which shows the different levels of producing speech and the connections between them. When trying to produce speech, you start with semantic features which will trigger words; syntax is chosen; word shapes are chosen; and particular phonemes (sound segments) are chosen. A word and a phoneme may mutually excite each other, while the phoneme and other phonemes mutually inhibit each other. For instance, "dog" will be mutually excitatory with "d", but "d" and "k" will inhibit each other. These principles help explain why speech errors are systematic. Speech errors are usually either repetition, anticipation, or exchange, and they tend to stay within the pronunciation system of the language.

Dell's model explains why pronouncing the words "deal beak" is much more likely to produce the error "beal" than pronouncing the words "deal back". In his model, "deal" and "beak" share the same vowel /i/. Because the phonemes and the words are mutually excitatory, sharing the vowel means the words have extra activation between them, and the slip "beal" is more likely. The key is that the activation moves up and down in the system.

Such models are useful for clinicians who are trying to find intervention methods for patients with brain damage or speech defects. Dell's models was useful to a team of clinicians with a patient who had seemingly unsystematic speech defects. They found that by varying the parameters in the model, they could use it to explain those speech defects. The spreading activation approach to such problems allows us to look as clinical problems differently and get new interventions to help patients. These models are also just beginning to influence thinking about language in education.

Selected Slides:

References: Reader: 1-4; Regier: ch.1