Eigenanalysis for Lossy Compression

Eigenanalysis is a method for reducing a set of data to the principle dimensions along which that data varies. In the context of imaging data, it has been applied very successfully to Eigenfaces:

Where a set of faces is broken down into a smaller set of face "prototypes" that can recombined in varying portions to recreate the original data set with a limited accuracy.

In the context of music, I can imagine that the spectral characteristics of songs have some self-similarity: portions repeat, chords are repeated in different voices and different octaves, rhythms repeat, etc. I can imagine a lossy compression algorithm that takes the frequency domain representation of a song, does Eigenanalysis on these vectors, and stores the song simply as the collection of N eigenvectors and the reduced representation of each frequency-domain chunk.

Quantization methods may be employed for further reducing bit usage due to similarity between adjacent chunks. Or different portions of the spectrum can be analyzed separately, which allows for better representation of lower frequencies and less information dedicated to higher frequencies. This unfortunately does not account for the obvious relationship between the lower and higher frequencies.

A more advanced implementation may involve doing eigenanalysis on mutiple chunks simultaneously in a moving window, or at different scales, which will help with rhythmic repetition.

The octave or overtone relationship is a little more complicated, and would require something like a constant-Q transforms to get a logarithmic frequency domain.

Jesus Glitch

The holy is often found in unexpected places. Jesus in naan, Mary in a Chicago underpass. Dan Paluska has immortalized this concept with his Holy Toaster.

Why don't we ever see Jesus in corrupted image files?

Image compression algorithms are generally rated on their ability to convincingly ignore non-perceptually-relevant features. I propose a new metric for these algorithms: how likely they are, when corrupted, to produce an image of a holy figure.

Precision CD Glitching

"Wounded" CDs have been prepared by artists like Yasunao Tone and Oval, encompassing the experimental and pop domains of music, respectively. In both cases, the music has to be re-recorded from the glitched CD to be heard (and in Oval's case, it is subject to further production). Why not use a laser cutter to make precision glitched CDs, allowing them to be distributed directly?

Crowdsourcing Glitches

I'm interested in the possibility of destroying data through natural processes. I've already explored this a bit with Future Fragments, where quotes from friends were carried in their pockets and destroyed over the period of a summer.

Another interesting system that naturally destroys information over time is the human mind (memory). Consider the possibility of crowdsourcing distortion: for example, with Mechanical Turk, create a task that plays a subset of a song that the worker is then asked to repeat. These responses are then averaged to create an approximation of the original tune.

This task takes advantage of our ability to hear melodies. We can also treat the mind as a more generic digital signal processor: ask each worker to recall the midi pitches of each note. Or better: portions of the mp3 encoded audio in hex. Besides the auditory system, we can also take advantage of the visual system. Ask each worker to recall and draw the audio signal (at a sufficient scale).

These tasks would be especially interesting in the case of people like Ben Pridmore, who is able to quickly memorize large amounts of data (e.g.: 364 playing cards in 10 minutes). I imagine his memory slowly degrades. It'd be great to see what a compressed image looks like that is memorized slightly incorrectly, and watch it degrade over time.

Naked Mail Distortion

A method for destroying data: send optical discs in the mail, with no protection, back and forth between two locations. The data may be collected each time and noted for future reference. Potential subjects include compressed and uncompressed media (audio, video) and written texts. Consider also a iterative/recursive algorithm where the distortion provides a decision for a chance operation: the CD contains a list of addresses, and every time an address is lost the CD is sent to that address.