Neural networks and characteristics of human cognition

• Cognitive architecture
  • Sensory store

    Activity states of neurons are organised with the help of exitatory and inhibitory signals exchanged between them. Sensory information along with stored data result in convergence to a stationary state which lasts a fraction of a second. (Sensory store of time-scale tuned to typical progression of events in real life)

  • Short-term memory = firing of neurons (? time scale?)

  • Long-term memory = synaptic strength

    LTM is reorganised with the help of synaptic plasticity, eg. Hebb's rule - connections are strengthened when both neurons are active together. The effect is to increase the likelihood of those activity states that occured in the past.

  • Associative Memory (found in drosophila and snails) - an optimisation problem. Morelli example of Jets and Sharks - any one node can activate the other connected nodes.

  • Content-Addressability of Memory (not clear - is this same as associative memory? See Morelli

  • Constraint Satisfaction Processing (different from search-based techniques of classical AI)

• Not so good at deductive logic

• Performance improves with practice

• Possibility of unsupervised and supervised learning; ability to use external feedback on the desirability of present activity state

• Dealing with uncertainty, imprecision, noise - imitates human flexibility.

• Spontaneous generalisation - classification by similarity rather than by property lists. Use of prototypes is seen in humans in categorisation tasks calling for say: a list of examples of birds, response times to verify ___ is a bird and checking defaults in comprehension - I saw a ``bird'' on the grass.

• Might offer new explanations for some puzzling discontinuities in cognitive development.³

• Some philosophers are even hopeful that neural networks may exhibit qualia (= patterns of firings of neurons)

Supervised learning: A desired output signal is provided to the circuit along with the input signal. Over several cycles the system adjusts its parameters to bring its response closer to the desired output. (Example: the back-propagation algorithm).

Most human intelligence (eg. perception - recognising objects and events in the environment) is aquired by mere exposure, without a teacher. Tasks like sorting by appearance can be done by very young children. The mechanism behind such ``unsupervised learning'' is assumed to be ``self organisation''.

Unsupervised learning in a neural network does in fact involve target values: most often the targets are the same as the inputs. With this target, unsupervised learning performs the task of dimensionality reduction, compressing the information from the inputs. Unsupervised learning (I think under the name of ``cluster analysis'') is used in data visualization. (Interpreted from ftp://ftp.sas.com/pub/neural/FAQ.html)

Though sl was the popular form in the '80s, usl now predominates (DeLiang Wang. review of ``Unsupervised Learning: Foundations of Neural Computation'' Eds. Geoffrey Hinton and Terrence Sejnowski - reviewed in AI Magazine, Summer 2001).

Objective of usl is to find hidden structures/ correlations in input data: discover clusters, extract features to characterise data compactly, uncover non-accidental coincidences.

Although most connectionist models of high-level cognition are still over-simplistic and biologically implausible, the underlying (distributed/ bottom-up) principle remains an important feature of cognitive science.

Current NN applications range trom home appliances to predicting the stock market, extracting image data from radar information, controlling cars, robots ...