I'm working on making the jump from READING about recent machine learning techniques to APPLYING recent machine learning techniques. I have some nice datasets at work, so I am doing a bit of double-dipping by playing with those.
I previously struggled with getting started, getting bogged down in picking technologies and such. I'm specifically going after some Neural Network models, so also get overwhelmed with the Python ecosystem of tools and libraries. The best way to get started is to get started .. so I turned the corner with a little help.
; Fluent Python (book) : Great advanced-introduction to Python, letting me better understand/recognize meta-programming and DSLs. Things like the array-slice syntax turning into an object which can be passed to any class that implements the right methods.
; Keras : I think this is the second-most popular python NN library, after Tensorflow itself, and it has Tensorflow as a backend. Seems nice and usable. Then I pull in whatever numpy stuff as needed.
; Convolutional Neural Networks : I knew about image and audio convolution filters, so had to spend some effort in transferring that knowledge into how convolutional layers work in a CNN. I read a bunch of articles like Visualize CNN With Keras to get my mind around this.
; Dog vs Cat with Augmentation : I had been through random tutorials before, but this time I went through Building Powerful Image Classification Models Using Very Little Data in detail, step by step, and made sure that I roughly understood each step. To help, I switched from Dog and Cat images into my own dataset -- so I had to go through the usual data cleaning process in addition to the learning.
; Living Code : Meanwhile I've also been sharing my excitement with friends and coworkers, and started a new Machine Learning club at work. We're taking even the simple examples and building them into services that our live systems can use - code that isn't running is dead code, and dead code is soon abandoned code. A big part of this learning is getting LIVING examples, and the double-dipping of making something useful for work is a great way to get that.