IITM BS Machine Learning Techniques (BSCS2007): Syllabus and Tips

Machine Learning Techniques (BSCS2007), or MLT, is a 4 credit diploma level course in the IITM BS in Data Science. It is a fixed course in the Diploma in Data Science. It has no prerequisite, but Machine Learning Foundations (BSCS2004) is a corequisite, so you take MLF before MLT or in the same term. MLT teaches the main methods for regression, classification and clustering, and when each one suits a problem.

The course page names Arun Rajkumar, Assistant Professor in the Department of Data Science and AI at IIT Madras, as the instructor. He is also one of the MLF instructors.

What you learn

• Weeks 1 to 4: learning without labels. Principal component analysis and its kernel version, K-means and kernel K-means clustering, then a recap of maximum likelihood and Bayesian estimation leading to Gaussian mixture models and the EM algorithm.
• Weeks 5 and 6: regression. Least squares and its Bayesian view, then ridge and LASSO regression, which add a penalty to control overfitting.
• Weeks 7 to 9: classification. K nearest neighbours and decision trees, naive Bayes as a generative model, then the perceptron and logistic regression as discriminative models.
• Weeks 10 to 12: stronger classifiers. Support vector machines, ensembles through bagging and boosting (AdaBoost), and neural networks for multiclass problems.

For every method, the course page expects you to know four things: the model, the loss function, the optimisation method and how it is evaluated. It also wants you to fix underfitting and overfitting and to choose between methods for a task.

The suggested books are Pattern Classification by Duda, Hart and Stork, Pattern Recognition and Machine Learning by Christopher Bishop, and The Elements of Statistical Learning by Hastie, Tibshirani and Friedman.

How it is assessed

The course page lists the standard pattern: 12 weeks of coursework, weekly online assignments, 2 in-person invigilated quizzes and 1 in-person invigilated end term exam. Your term's grading document has the weights.

Where it counts

• Diploma in Data Science. MLT (4 credits) is one of the fixed courses. See the Diploma in Data Science.
• Machine Learning Practice. The MLP course page lists MLT as a prerequisite.
• Re-entry. The handbook lists MLT and MLP as the final exams for a student who exited with the Diploma in Data Science and rejoins at the Diploma in Programming level.
• Electronic Systems students. The ES handbook lists MLT (CS2007, 4 credits) as an open elective and in the ES Minor in Machine Learning.
• Terms. The DS handbook says all diploma courses are offered every term.

Mismatch between sources. The DS course page lists MLF as a corequisite of MLT, with no prerequisite. The ES handbook's table lists MLF (CS2004) as a prerequisite instead. DS students should follow the course page. ES students should confirm before taking both in one term.

Who finds it hard

MLT covers a new method almost every week, so falling behind by two weeks hurts. It uses MLF maths throughout. If you take MLF in the same term, some MLT weeks will need ideas you have not fully settled yet, such as eigenvectors for PCA or Lagrange multipliers for support vector machines.

How to prepare

1. Keep one page per method. Four lines: the model, the loss, how it is trained, and when it overfits or underfits. Before a quiz, cover three lines and recall them from the fourth.
2. Work tiny examples. K-means on the points 1, 2, 10 and 11 with two clusters ends with centres 1.5 and 10.5. Do small cases like this by hand for K nearest neighbours, naive Bayes and the perceptron too.
3. Compare methods side by side. For a given problem, ask which method you would try first and why. The course page names choosing between algorithms as a goal.
4. Code a few methods from scratch. A short K-means or perceptron in Python makes the maths real, and it prepares you for MLP.
5. Revisit MLF each week. Before each MLT week, re-read the MLF topic it builds on.

What to take before or after

Take Machine Learning Foundations (BSCS2004) first or alongside. After MLT, Machine Learning Practice (BSCS2008) applies the same methods with scikit-learn. At the degree level, Deep Learning (BSCS3004) is one of the mandatory BSc courses. For how the three ML courses link, read MLF, MLT and MLP explained.

Common questions

Can I take MLT and MLP in the same term?

No. MLP lists MLT as a prerequisite on its course page, so MLT must be completed first. The ES handbook shows MLT as a corequisite of MLP instead, so ES students should confirm.

Is MLT mostly coding or mostly theory?

The course page focuses on understanding methods: their models, losses, optimisation and evaluation. The practical, library-based side is the job of MLP. Writing small implementations yourself still helps a lot.

Which book should I start with?

The course page lists three books without ranking them. Use the lectures as the main source and dip into a book only when a topic is not clear to you.