PCA object

Foot beat recognition using PCA and Svm in rubyk.

Finally, I found the courage to implement Principal Component Analysis using BLAS and LAPACK libraries. See this post for information on these libraries.

On the partition, the top line shows the recognition of a foot beat with three classes (nothing, leg up, leg down). As you can see, I am not really able to do something in the rhythm yet…

PCA object

This object implements Principal Component Analysis. If you are interested, I wrote a small PDF to help me understand this topic:

207 Kb

PCA Notes on Principal Component Analysis.

The object is really easy to use: it has an input with vectors of dimension M and it outputs vectors of dimension N < M. It has just one method “learn”. It reads the training data from files recorded with the ClassRecorder object. The PCA object also transposes all the training data into its new vector space after building the transformation basis. In our setup, we use this transformed data to train the Svm object.

We have finally come to the point where Vitalijs Butenko left the research on our first data set recorde with Nicholas in January.

From the noise my poor PowerBook G4 is making with its fans, it seems we will need the fastest mac minis out there to make all this work… It’s quite funny to see that the FFT object object is less CPU greedy the PCA using BLAS to compute the multiplication of a 288 vector of doubles with a transformation matrix of 288×32. You might think 288 is a large window, but it is not, it’s just 24 samples. This is 100ms with our current data rate (256) and 12 signals.

Again, many thanks to Vitalijs and the people at the LANOS laboratory (Mr Hasler, Ms Acimovic). Without their theoretical work on the initial data set, I would never have had the courage to go so deep into these (difficult) machine learning topics. Knowing that it worked in Matlab was really encouraging.

We can finally start the tuning phase of this setup (make it really work).