Friday, 31 May 2013

Wavelets and Expectations

Pinpointing something interesting in a stream of information is a big challenge. Recently, wavelets have provided a means by which temporal changes in structure can be identified. This provides a way of making distinctions in continuous data automatically using a fairly simple algorithm. With image data, video and sound, the techniques promise remarkable data processing capabilities, making things like automatic video description a reality - YouTube's recently acquired ability to pattern-match video content from computer games indicates where things are going (and how new areas of copyright infringement will emerge that we hadn't even thought of!)

Video is essentially a continuous function, and wavelet analysis requires this. Sound waves are ideal, and the data of pixels in a picture will also do quite nicely. But what about text? And, what about music represented in the form of notation? This kind of data is not continuous. There is, however, a continuous ebb and flow of expectation, but that ebb and flow is dependent on redundancies and entropies in the text. Text presents different problems because it is the result of decisions, which deep down in our consciousness may be a continuous function (although we don't know), but even if it is, it's a continuous function that we do not have direct access to. Indeed, musical decisions might be the closest we can get to the continuous function of consciousness.

We can speculate what happens in the continuous function with things like text. Mostly, with text it is redundancy. Mostly I think it serves to prolong expectations. The continuous function serves to structure expectations in some way. Sometimes expectations are transformed by some new event. The transformation of expectations is accompanied by new redundancies which in turn feeds new expectations. An expectation maintains itself over shifting redundancies. This is the essence of the art of variation as practiced by artists and composers.

The central question is whether there a way of charting the continuity. Is there a way of mapping the relationship between tension and release? I wonder if there might be a correlation between the length of the path of recursive meta-game tree and the degree of tension that is experienced. In essence that might  map onto the length of a branch of a structure which is maintained by the continuous process. This might equate to entropy levels in choosing a next possible act. High entropy means tension. Low entropy means release. Similarly high tension would mean stimulating the creation of redundancy, and low tension would mean easing-off the creation of redundancy (after all, with the lowest entropy, there is little scope for redundancy).

Wavelets are good for removing redundancy. In compression algorithms like JPEG, wavelet analysis is used to remove the repetitions of bits in a process whereby those repetitions are coded as transformations of the wavelet. What it leaves behind are the 'edges' where distrinctions can be made. There are remakable advances currently underway in the analysis of sound waves and the recognition of patterns. And wavelet analysis is an important feature of more advanced data mining techniques like granular computing.

Because a wavelets repeat throughout the structure, its distortions indicate where the redundancy has really changed. That is an indication of where expectations may similarly have changed. At those moments where it does change, we can expect information transfer and meaningfulness to occur. Maybe.

The question is "which way round to look at this?". We could, for example, go down the route of an artificial psychology (actually, I would prefer an artificial sociology!), and model candidates for the continuous function and compare the utterances that it makes. If the morphogenetic process map onto things we see in reality, then it might be worth pursuing. But the most  pressing issue is how we understand the information around us, and how we organise ourselves. The most pressing issues are political not epistemological!

That means sorting out a model of the continuous process which discounts those models of the process which lead to social pathologies. I think that many of our social pathologies stem from bad theory about cognition. A better theory demonstrably shows the nature of cognition as inherently convivial.

If such a model could be established, then we would run our institutions and governments differently...

No comments: