This summer I spent quite a bit of time speaking to various people about “intelligent” Edge computing. I put double quotes around the word intelligent to avoid the wrath of the thought leaders on the internet as they fight each other over what intelligence is and what is not.

The more I talked to people the more I was realizing the possibilities and opportunities the edge computing opens up for the future: “injecting” intelligence into dumb physical objects seems akin to injecting “life” to dead objects, at least for us, scifi fans, which makes up for interesting vision of the future (for brevity let’s ignore the IoT security issues in this post). The possibilities are even more exciting in light of machine learning and neural networks models getting lighter and more easily deployable in the physical space. And let’s not forget about all the efforts to bring intelligence into some variation of IoT blockchain.

Tensorflow conditionals and while loops

Recntly I found myself needing to implement more advanced control flow in some models I have been hacking on in my free time. In past I never really needed any graph conditionals or loops or any combinations thereof, so I had to dive into documentation and read up on them.

This blog post covers tf.cond and tf.while_loop control flow operations and was written to document and share my experience learning about them. Both of the operations seem intuitive on the first look, but I got bitten by them so I wanted to document their usage on practical examples so I have something as a reference I can return to in the future should I need to.

As I continue to explore the realm of Artificial Neural Networks (ANN) I keep learning about some really cool types of neural networks which don’t get much attention these days. Many of these networks have not found much practical use in the “real world” for various reasons. Nevertheless, I find learning about them in my free time somewhat stimulating as they often make me think of some problems from a different perspective. One such ANN is called Hopfield Network (a.k.a. Hopfield net), named after its inventor John Hopfield.

Couple of months ago I came across a type of Artificial Neural Network I knew very little about: Self-organizing map (SOM). I vaguely remembered the term from my university studies. We scratched upon it when we were learning about data clustering algorithms. So when I re-discovered it again, my knowledge of it was very basic, almost non-existent. It felt like a great opportunity to learn something new and interesting, so I rolled up my sleeves, dived into reading and hacking.

My rekindled interest in Machine Learning turned my attention to Neural Networks or more precisely Artificial Neural Networks (ANN). I started tinkering with ANN by building simple prototypes in R. However, my basic knowledge of the topic only got me so far. I struggled to understand why certain parameters work better than others. I wanted to understand the inner workings of ANN learning better. So I built a long list of questions and started looking for answers.

This is the second post of the series about Principal Component Analysis (PCA). Whilst the first post provided a theoretical background, this post will discuss the actual implementation of the PCA algorithm and its results when applied to some example data.

Theory Recap

In the first post we learnt that PCA looks for a vector basis that can express the analysed data in a better (less redundant) way, whilst retaining as much information from the original data as possible. The vectors that form this vector basis are called principal components.

This is the first of the two post series about Principal Component Analysis (PCA). This post lays down important knowledge bricks that are needed to understand the core principles of the PCA algorithm. The second post will discuss the actual implementation and its results by applying it to various data sets.

Motivation

The world is becoming more data driven than ever before. We collect large amounts of data from arbitrary sources. This is often because we don’t know which data best describes the systems we are trying to understand. The fear of missing out on capturing important features forces us to collect a lot of redundant data.