Is This Connectionism?

"Intelligence seeks connection"

Connectionism, in computing terms, is the understanding that connection plays a dominant part in the architecture of a computing mechanism such as the brain. Where connectionism goes astray is its assumption that the connections of an Artificial Neural Network (ANN) are a good model for the connections that exist in brains.

When we seek to understand the connection between A and B, we are much more subtle than to expect an implication such as


or that some weighting on A will provide an output B. We don't require a directed connection, or can wait for any directionality to appear dynamically.

But the hardware of brains, the neurons, are directed.

At the lowest level, yes. But that doesn't mean our thinking is constrained to only operate on this level. By back-connection, by connecting the output of a system to its own input, an apparent undirectedness can arise in the system.

And we don't have to use an exact analogue. Just as our transport machines use wheels instead of feet to move around, we can use undirected wires instead of back-connected directed components to avoid the limitations of directionality.

And sometimes, humans demonstrate their intelligence by the seeming lack of connection - what might be termed the "octopus effect". An example is marketing analysis of instant coffee, where a cause for a particular effect is sought by searching in many directions - up, down and sideways. The sales of a particular product have risen - is it because all such products have risen (up), or has this product's sales risen in particular stores (down), or have the sales of a competing product fallen (sideways). Here, the human is pivoting on some point in a virtual web of connections, where the connections in their head can be broken and remade at will.

Another example where simplistic connectionism has a problem is the associative web of words. A person may have a vocabulary of 50,000 words, with their associations. And yet, when confronted with a new utterance of say a dozen words, the words and their meanings can be brought together and analysed in at most a few seconds. Even if a local structure specific to the utterance were created in this time, it still suggests that we are capable of transferring reasonably complex messages through our connections, messages that themselves can stand in the place of remote connections.

In linking connectionism with ANNs, usually only the operational state of the ANN is considered, the whole process of back propagation and layer creation being ignored, as lying outside the theory. A successful theory needs to encompass the whole process, not a small static part once the connections have  been established.

Connectionism seems rather like a religious sect, where some aspect of the truth is taken to be all of the truth. Connections are obviously important, but we should not assume that connections are necessarily directed, and we should not assume the static weightings found in an ANN to be of any relevance to understanding the workings of an animal's cognitive system, nor of machine analogues.

Back Connection

Knowledge Representation