Research ideas

A sound's sketch

We want to recognize the words that different sound sources are drawing, the different variations these can take, how they are combined with other words, and so on, learning the vocabulary and grammar of an imaginary secrete language they use. They: in general, and in particular.

The original form is a superposition of analog signals. The form we have to work with is digital waveform. We reduced this waveform further into featurized, then quantized, segments.

It seems self evident that any pattern that would be useful to include in our language should be a pattern that occurs frequently enough. Let's imagine the patterns we're looking for as particular patterns of the curve made by a moving point in our feature space. Pushing the analogy further, we can imagine that the language we're trying to reconstruct is one generated by several unknown authors, in unknown languages, by (possibly simultaneously) "swipe typing" on a multi-dimensional keyboard (whose contents and shape is also unknown).

Piece of cake.

Frequent patterns

Let's start with the obvious impulse to anyone who is old enough to have ever been called a "data miner": Sequential pattern mining or more precisely, with n-gram. This impulse is in the right direction, but we need to be clear on what we'd want out of those techniques, and how these may be limited.

... to be continued

Should discuss:

• n-grams
• n-sets
• regular-expressions (equivalently, finite-state-machines)
• probabilistic regular-expressions

In light of:

• sensitivity to variations
• combinatorial explosion when we count order
• how order may not matter, and if it does, should be captured by n-sets
• advantage of n-sets over n-grams to be able to use snip expansions (form snip networks (e.g. similarity networks))

Also:

• clustering snip_expansion aggregates (adding them is just one way)
• going back to features and clustering sequences or sets of those instead?
• how do we solve the window size choice here (same problem as with tiles/chunks, but with different contexts, contents, and therefore possible solutions)

https://en.wikipedia.org/wiki/Sequential_pattern_mining https://en.wikipedia.org/wiki/Structure_mining https://en.wikipedia.org/wiki/Graph_kernel