AUTUMN
September marked the start of my Masters degree at the University of Bristol. A degree entitled Machine Learning, Data Mining, and High Performance Computing. A bit of a mouthful, but reasons exist for this particular choice. My interest in the fields of data science and artificial intelligence built up over time. A kind of interest akin to an ocean wave breaking upon the shore. A wave which was there all along, but only became visible and tall when it reached its destination. A point by which you, the beachgoer, get smacked in the chest and taken by surprise. That’s kind of what happened with me, one day making the decision that this is what I wanted to do. So there I was. Now, let me tell you about what I did.
Performed Image Manipulations
One of my most enjoyable courses would be one taught by Tilo Burghardt. He is the most energetic professor I have ever come across. The annunciation and variation in his voice is phenomenal. Tilo aside, his course, Image Processing and Computer Vision covers a lot in a small amount of time. The course outlines what makes up an image, and how one can encode them. He describes Computer Vision as attempting to bridge the semantic gap between pixels and meaning.
With the details aside, my teammate and I created a program which counts coins in an image. It does this by encoding images within a matrix representation. We then performed successive transformations on the matrix to detect coin locations. Each detected location forms a prediction on where the center of a coin is within the image. Also encoded is the amount of support there is for a given circle radius. Given that we have the center points, and a predicted radius, we then draw the circles overlaid on the original image.
The techniques used during this process are:
- The derivation of the image in the x and y direction.
- Calculations of the gradient direction and magnitude.
- Thresholding operations.
- Encoding the shape of cirlces using the Generalised Hough Transform.
- Detecting the existence of coins using the same methodology.
Solved Problems with Genetic Algorithms
- Learning in Autonomous Systems.
- Taught by Tim Kovacs.
- Simplifies what could’ve been a complex topic.
- Involved Evolutionary Computing and Reinforcement Learning (Unsupervised).
- Explored how EC and RL are different, but synergistic approaches.
- Defined what Evolutionary Computing is.
- Defined what Reinforcement Learning is.
- Explored how they can be used together.
- Solved real-world problems using Genetic Algorithms.
- Solved the OneMax problem as a sample example.
- Solved the Traveling Salesman Problem.
- Solved a specific incarnation of the CVRP problem.
Learned How to Derive Bayes’ Theorem
There is another class which, although unenjoyable, has some interesting content. The module is Uncertainty Modelling for Intelligent Systems, taught by Jonathan Lawry. It covers many modelling techniques ranging from probability and information theory, to Dempster-Shafer theory. It is more mathematical than it needs to be, yet the probability aspects are quite intriguing.
One such interesting probabilistic technique is the infamous Bayes’ theorem. Since becoming a rationalist, I’ve been meaning to learn it in an intuitive manner. If you’re in a similar position you usually get pointed towards Eliezer Yudkowsky’s post. So, I tried reading it. It makes sense as you go through it, but when I tried to think back on it, I became even more confused. His article even spawned an article explaining it! This lead to the discovery of a metaphor based approach. Although this was better, it still failed in the same way. I then came across a Venn Diagram based approach. It was at this point that things finally clicked. It seems we all learn things in different ways.
Now here’s the magical part. I was then able to derive Bayes’ theorem, without any prompts. Not only that, but I attempted it a few days later with the same results. Although, to be honest, I still need to learn how to actually apply it in everyday situations. I guess that leaves something to do in the future.
