COGNITIVE SCIENCE ================= Introduction ------------ Cog. Sci. is a large subject, to which it is impossible to do justice in a single lecture. In addition to that, it is an area which I cannot claim a great deal of knowledge. I have emphasised the importance of performing experiments. Other important inputs to this science are introspection, and examination of the behaviour of selectively brain-damaged patients -- results of which are necessarily scanty. Neurophysiology is of doubtful importance at present -- the human brain is too complex for the functions of individual neurons to make sense, but experiments with much more primitive animals (e.g. molluscs) may yield some results of interest. I have omitted a number of important issues in this somewhat light treatment of Cog. Sci., such as the structure of the data held in STM, the mechanism of learning, problem solving and the nature of procedural knowledge. Many of the ideas presented here are based on research by John Anderson (ACT*). An account of this work is to be found in The Architecture of Cognition, by John Anderson (Harvard). ---------- Short Term Memory ----------------- Short term memory (STM) contains the items currently in consciousness. It is normally thought capable of holding approximately seven (plus or minus two) items. There are at least two theories as to why STM is limited to approximately seven items: (1) There are only seven locations to which our conscious minds have access. (2) The limitation arises because conscious thoughts decay without explicit rehearsal, which occurs as a result of productions being fired, depends on the rate of firing of those productions. Production firing is a sequential process. The first possibility still leaves open how the locations are ordered, and where in STM new items are added. The possibilities are (1) at the beginning, (2) at the end, or (3) in a location selected randomly among the seven. To discriminate between the two alternatives, an experiment could be used. Experiment 1 ------------ Try memorizing arbitrary sequences of digits. Whenever the sequence is longer than seven digits, note where in the sequence forgetting normally occurs. The first alternative predicts that either the first few digits are forgotten, or the last few, or digits are forgotten randomly, while the second predicts that the last item added should normally be remembered (because it has not had time to decay), and the first few should become progressively weaker (because the first one has been rehearsed the most, followed by the second, and so on). ---------- Declarative Long Term Memory ---------------------------- Declarative Long Term Memory (LTM) is often thought of as being held in some kind of semantic network, in which data retrieval occurs in parallel. If LTM is a semantic network (or frame system), concepts are related to other concepts, and it would be reasonable to expect that, given a concept, it should be possible to retrieve a concept which is closely related to a concept which is currently held in STM, by using some operation such as fget(elem, any slot). To test this, we could devise an experiment such as: Experiment 2 ------------ Here, two games are played: (1) Word Association Football. In this game, a group of people (two or more) take it in turns to say a word related to the word previously said, e.g. word .. association .. football .. match .. strike .. picket .. flying .. pigeon .. clay .. pot .. (2) Word Dissociation Elephant's-Foot-Umbrella-Stand In this game, a group of people (two or more) take it in turns to say a word completely unrelated to the word previously said, e.g. word .. dissociation .. elephant's-foot-umbrella-stand .. observatory .. accordion .. omelette .. eisteddfod .. It is forbidden to repeat a word which has already been said. Time the speed at which the words are said during the two games. Unless cheating takes place, Word Association Football should be faster. I found it harder to produce the second list. If data are retrieved in parallel, there should be no significant difference in times taken to retrieve facts at the same level, and a tree search take a time proportional to the depth of the tree which is searched rather than N ** depth, e.g. Rajiv Gandhi is an Indian, and Rajiv Gandhi is a politician should take roughly the same time to retrieve, whereas Rajiv Gandhi is an Indian should be retrieved faster than Rajiv Gandhi is a vertebrate, because it is unlikely that anyone would store such a fact explicitly. Note, though, that many people are likely to store facts such as Rajiv Gandhi is a person explicitly, as it is useful even though it is redundant (all Indians and all politicians are persons (though it would be better if some politicians were unpersons in the Orwellian sense)). A further issue related to LTM is how data in LTM are retrieved and added to STM. Given a particular node, other nodes can be reached by traversing the arcs leaving the node. We could say that "activation" spreads from that node along the arcs to adjacent nodes. Nodes with sufficiently high activation levels enter STM. We can observe a property of this spread of activation in the following experiment. Experiment 3 ------------ Memorize the following four sentences: (a) The doctor is in the bank. (b) The fireman is in the park. (c) The lawyer is in the church. (d) The lawyer is in the park. Which sentence was easiest to recall? Which sentence was hardest to recall? What you should have observed is a phenomenon called the fan effect. DOCTOR FIREMAN CHURCH | | | v 1 v 1 v 1 | | | (a) 2 (b) 1.5 (c) 1.5 | | | ^ 1 ^ .5 ^ .5 | | | BANK PARK --->--- (d) ---<--- LAWYER .5 1 .5 In recalling the sentences, activation spreads out from the six word nodes (DOCTOR, BANK, etc.). The more arcs leaving a node the weaker the activation on that arc, i.e. the activation "fans out". (a), with an activation level of 2 should be easiest to recall, and (d), with an activation level of 1, hardest. It is worth noting that likelihood of recall of data in LTM decays with time, but over days and even years, rather than the seconds over which STM decays. Different data are forgotten at different rates: it is much easier to forget the author of a book than how to ride a bicycle or how to swim. The first case is declarative knowledge (nodes and arcs in a semantic network), the last two procedural knowledge (production rules). Another cause of forgetting is repression -- if a memory has unpleasant associations, no matter how contrived, it is easier to forget it, even if you make a conscious effort to remember it. In addition to causing activation, memories can apparently inhibit it. Sigmund Freud (Psychopathology of Everyday Life) examined this phenomenon. ----------