Hello all. So finally I'm taking this Machine Intelligence course (CMPE 677) this Fall 2016 at RIT. Prior to this class, I had some exposure to this discipline in the co-ops/internships/projects I did during my academic career so far. I must say, when you start thinking about how to make something learn something, you finally understand the value of why you are here in college, chugging through all the piles of course materials, in a given academic program (oh boy, I have 2 more years to go...God Speed).
To give a little more context to what I'm talking about, I used to rant about my college courses in my first 2.5 years at RIT. Then I went to the industry and was like, only if I took that Linear Algebra class more seriously, may be this task of writing this max unpool kernel might be easier. One of the things that bugged me about working in the industry was that, whenever I lacked understanding of something, I googled it and understood it to the point where I could get the code working. It bugged me because, often I found myself stopping when things started getting interesting (I was working under timed constraints. Meet the deadlines of clients. Gogogogo..). For instance, I understood how max unpool layer works by restoring the layer that got passed through during max pool in a deep neural network for image segmentation, however as I read more I became more interested into what kind of data loss happens between this operations, how the filters look if visualized, is parallelizing the matrix multiplication process in cuda the only way to make it faster, Scott Gray wrote his own assembly kernels for such multiplication - have we reached the limit to making this process fast?? Such questions bugged me and I realized I need a deeper understanding of these concepts and hence RESPECT (brotherly fist bumps), Machine learning pioneers in arxiv and other research avenues.
I hope you find these notes useful and feel free to share your understanding and develop the writings here as you see fit, by sending me a note at @tousifsays in twitter.
Tousif (like you 'toss' something and put an 'if' beside it)