Correction: (max 500 charact.)
The complaint will not be published.
I 128 ff - 145
Neural Networks/Pinker: Learning/Problem: there are incorrect reinforcements with "XOR" (exclusive or; Sheffer stroke).
Solution: we have to interpose internal >
representation .
I 142
Neural nets/Rumelhart: neural nets return all errors.
"Hidden levels": several statements that can be true or wrong can be assembled into a complex logical function, the values then vary continuously. The system can place the correct emphasis itself if input and output are given - as long as similar inputs lead to similar outputs, no additional training is required. >
Homunculi .
I 144f
Connectionism/Rumelhart: the mind is a large neural network. - Rats have only fewer nets.
PinkerVsConnectionism: networks alone are not sufficient for handling symbols - the networks have to be structured in programs. - Even past tense overstretches a network.
Precursors: "association of ideas": Locke/Hume/Berkeley/Hartley/Mill >
Association/Hume .
1) contiguity (context): frequently experienced ideas are associated in the mind
2) Similarity: similar ideas activate each other.
>
Similarity/Locke .
I 146
Computer variant: is a statistical calculation with multiple levels.
I 147
VsConnectionism: units with the same representations are indistinguishable. - The individual should not be construed as the smallest subclass.
I 151
Connectionism cannot explain compositionality of representation.
>
Compositionality .
I 158ff
Recursion/Recursive/Neural Networks/Memory/Pinker: recursion solution for the problem of an infinite number of possible thoughts: Separation of short/long-term memory. The whole sentence is not comprehended at once, but words are processed individually in loops.
>
Recursion/Pinker .
I 159
Networks themselves have to been as recursive processor: for thoughts to be well-formed.
I 166
Neural Networks/Pinker: the networks do not reach down to the rules - they only interpolate between examples that have been put in.
>
VsConnectionism .