Fun examples. Including figuring out nonlinear mappings of data (even into a jelly roll). Looks really difficult…
His mapping of distances to US cities (how did he do that??) is really similar to Gerritsen's “apply weighting by charge repelling and spring attracting” in her video.
They're really cool. Remember those bumps you see with correlation filters? It turns 'em into spikes! (the result is similar to doing a high pass filter, except a lot more official)
Time Series Analysis
Also known as real-time trading, identification, etc
Great overview set of slides: http://alumni.cs.ucr.edu/~mvlachos/ICDM06/. Talks about dynamic time warping, lcss algorithm, pca, all vs correlation and talks about benefits and disadvantages.
Anything you do other than correlation will be more expensive to compute, but could “see” your signal a lot better in the midst of noise / time dilation, etc.
However, “The more complicated the algorithm, the less it works” –Bhiksha Raj
To Read
Robi Polikar's Tutorial on Wavelet Analysis. Still not quite sure what it can do, will need to ask Matt if it will be useful. Seems like it might be useful for data compression. Also has really good tutorial PDF's on Pattern Recognition (hidden markov models, nearest neighbor analysis) and Ensemble Decision Making.
Periodic sampling of 250 Hz allows a 10 Hz signal to look like a -10 Hz signal and a 240 and 260 Hz signal in the frequency domain! Doesn't really matter in Matlab experimenting, but can screw you up when you are demodulating a signal or getting high frequencies corrupting your signal (because they too can be interpreted as a slow-moving sine wave).
Little difference between Hanning and Hamming windows, but you do want to use them! They reduce sidelobe leakage by a ton, but come at the cost of a 2X larger main lobe size.
FFT (Ch. 4)
Skipping for now…maybe I'll get more interested eventually. Basically, there are redundancies in computing the DFT (can halve spectrum computation w/ real input signals), and some other things I don't understand yet.