Reference no: EM133064081
6G6Z1105 Artificial Intelligence - Manchester Metropolitan University
Learning outcome 1: Analyse a real-world problem and select an appropriate combination of algorithms, building blocks and techniques from AI to compose a solution.
Learning outcome 2: Demonstrate the capability of capturing knowledge and preparing it in a suitably revised form for creating an AI classifier.
Learning outcome 3: Appraise and evaluate theoretical and practical issues underpinning AI and justify design choices for AI problem solving strategies.
Learning outcome 4: Design, execute and evaluate an experimental plan to create and optimise a small real-world system incorporating AI techniques.
Scenario
You're working for a company who want to include a new automatic image classification feature in their software application. Your line manager has asked you to undertake an initial investigation based on a particular classifier and shortlist of image features you're already familiar with, plus an additional investigation into a new kind of image feature you haven't met before. The relevant part of her initial email is reproduced below:
"[...] I'm particularly interested in what kind of features we should use for classification. We need respectable performance, but it doesn't have to be perfect. I'm wondering what the simplest kind of image feature we could get away with is. I probably need to see some performance metrics, along with some idea of how tricky each feature is to extract. I've drawn up an initial short-list of features I'm interested in, in order of increasing complexity:
1. Filesize-based features
2. Brightness-based features
3. Edge-based features
4. HOG1-based features
5. BoVW2-based features
6. CNN3-based features
Please will you produce a Matlab live script that uses the MerchData dataset to investigate the following two issues for each of the features on my short-list:
A. What kind of classification performance is it possible to get using Matlab's built-in functionality? Sticking with a single, simple classifier (please use k-NN, with k=3) and calling on any bit of built-in Matlab functionality that you need, what kind of performance is it possible to get with the MerchData dataset? Feature 1 is possibly a bit too simple, but I've thrown together a quick .csv file you can use to investigate, so hopefully it'll be quick for you to get some results either way.
B. How difficult each feature extraction process will be for us to implement for ourselves in the main codebase? How much of the built-in Matlab functionality you use in A are you able to understand and re- implement for yourself? You should demonstrate identical results with the built-in functionality wherever possible. I'm happy for you to work in Matlab, producing your re-implementations using standalone functions (.m files), but if you would prefer, I'm also happy for you to work in any other language you like, as long your work is still called upon directly from your Matlab live script.4 Don't worry about coding up extraction of feature 1 - I was able to do that very easily - but if you could code up the k-NN classifier instead, that would be great.
I know that you haven't worked with feature 6 before, but I've done some initial investigations myself and will write
to you with some further guidance5.
I would like to share your investigation with the rest of the team. They are all strong coders, but not necessarily strong on the Machine Learning side, so please ensure that your code is appropriately commented so that they can follow what you are doing, and why, at each stage of the investigation. [...]"
Attachment:- Artificial Intelligence.rar