I shared in my previous post about how I would like to simply explain my PhD research such that my mum can grasp it. Recently, I worked on a project that involved designing my research into a classroom resource for A-level students. AccessEd is a charity whose vision is to support student attainment and progression to university as a way of challenging the inequality in access to university. They actualise this by recruiting and training researchers to design and/or deliver university-style classroom resources about cutting-edge research to local schools. Last year (2018), I was among the 12 PhD researchers recruited to design teaching resources for local schools around the UK based on our academic research.
The next step after training was to “pin down your pack” which meant to select the resource subject, stage (Key-stage-4/O’level or Key-stage-5/A-level) and topic of interest. I felt computer science was the most related and I chose Key stage 5 as I didn’t want to over-labour myself simplifying my research another level down. However, the response from the organisers caught me off guard. They asked if I could do Key Stage 5 Mathematics as there were already three (3) researchers interested in computer science and none for Maths.
At first I thought it was their way of politely and tactfully laying me off, so I said to myself it was time to bow out and continue with my busy research life, after all who wants to crack their head with Maths. On a second thought, I decided to glance through the Key Stage 5 Mathematics curriculum to see if I could find any links to my research. Interestingly, I found several linked topics like vectors, graphs, numerical methods (approximation), algebra and functions (polynomials), differentiation (gradient), trigonometry, probability, Data presentation and interpretation and statistical distribution.
Here’s a summary of each of the 6-pack module I designed.
Pack 1 - A picture is just numbers!
An image is a 2-dimensional matrix (it has a row and a column). Popularly, an image can be represented in one of these 3 categories: binary, grayscale or RGB. We can convert from RGB to Grayscale image using some well-known techniques:
1. Average method which gives equal weighting to all the colors
Grayscale = (R + G + B) / 3
2. Weighted method: uses a weighted average accounting for human perception. Typically the green color has higher weight as it gives the most soothing effect to the eyes.
New grayscale image = 0.3 R + 0.59 G + 0.11 B
Pack 2: Image processing
As an image is just numbers to a computer, there are a lot of mathematical operations which can be done to transform it from one form to another. An image can be resized by increasing or reducing the total number of pixels using the basic concept of linear interpolation. By interpolation, we are trying to guess the color and intensity based on the known values of the neighbouring pixels. Histogram equalisation is another image processing technique that helps to improve the contrast of the image, i.e. Finding a balance in a picture where some parts are lighter and some darker.
Pack 3: Image characterisation (feature extraction)
Features of an image are one of the building blocks for image recognition operations. Features may be specific structures in the image such as points, edges or objects. Gradient vectors primarily lay the foundation for feature extraction methods. Uniform thresholding is a very basic feature extraction technique. It takes an image as input and segments the image into a binary image using a threshold e.g it’s either black (0) or white (1).
Pack 4: Object recognition (classification)
The simplest technique I found to explain object classification is the K-nearest neighbours (KNN). It ranks the likelihood of being in a group based on its closeness to other points. KNN kind of says “show me your friends and I will tell you who you are”. In the example below where k=5, the polo size of 4 out of 5 of the closest neighbours are Medium (M) hence we guess that the new point falls in M category.
Pack 5: What is anomaly detection
We defined anomaly as a data point that significantly deviates from the rest of the datapoints. If using a Gaussian model or normal distribution like in the figure below, most of the time the value of flour will be very close to 1 kg, and are seldom far from that value. Anomaly is when the value deviates too much from the mean, say if it weighs less than 920g or more than 1080g.
Pack 6: Anomaly detection in images/videos
Given a set of images, we can treat it as a supervised learning problem if they have associated labels for each object (say normal/abnormal, spam/not spam). Example in the MNIST handwritten dataset, each of the images are associated to one of the 10 labels/classes (from 0 to 9). We can apply classification technique as illustrated in Pack 5. It is called unsupervised learning if there are no labels. One of the most common techniques for solving this is the clustering methods like kmeans which uses Euclidean distance as a similarity measure.
