Sunday, November 6, 2011

ML-Class Insight into Neural Networks

Dr. Ng's lectures and exercises for the Stanford Machine Learning Class have opened my eyes to the essence of neural networks. They really are nothing new. A 3-layer network (with 1 hidden layer) is just a linear regression of a linear regression. But there's a subtly to switching outputs and regression coefficients during training, but other than that, the algorithms for neural netw are 100% identical those for logistic regression. What made it clear for me was the cut and paste we could do from the logistic regression code into the programming exercise for neural nets. I'll admit I was a bit disappointed when Dr. Ng started the class with lectures teaching basic linear regression and describing them using fancy terms like "logistic regression machine learning." However, I stand corrected. It was genius of Dr. Ng to set up the course this way. Using both logistic regression and a 3-layer neural net on the same, intuitive, problem--OCR (handwritten numerical digit recognition)--made this comparison of the two algorithms obvious. I'm looking forward to seeing how he deals with scale, translation, and ambiguous scribbles in upcoming image processing and OCR exercises. Now if only he could anticipate and share these insights at the beginning of a lecture series to get us excited about working through the math, that would turn this course into one hum-dinger of a class.

No comments:

Post a Comment