1
$\begingroup$

What is the name of the problem to explain how the neural activation for "red circle and a blue triangle" is different from "blue circle and red triangle"? Just activating the neurons for "red", "blue", "circle", "triangle" is not enough.

$\endgroup$

1 Answer 1

4
$\begingroup$

Representing the relation between objects + their attributes in neurons is called The Binding Problem. The Wikipedia article I've linked to is a bit questionable at the moment, so you may want to read the paper "Biologically Plausible, Human-Scale Knowledge Representation" instead. In the paper, the authors describe different ways of encoding symbols and their relation in neurons, using the NEF and SPA.

The paper shows how populations of neurons can represent vectors and symbols can be defined as vectors. Relations between symbols can be defined as operations on these symbols carried out by connection weights between populations.

For example, the scenario in the original post could be represented as:

$V = BLUE*CIRCLE + RED*TRIANGLE$

Where each word is it's own N-dimensional vector, $+$ is vector addition and $*$ is circular convolution.

Further discussion of this model can be found in this StackExchange answer.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.