IITM BS Deep Learning (BSCS3004): Syllabus and Tips

Deep Learning (BSCS3004) is a 4 credit core course at the BSc level of the IITM BS in Data Science. It covers neural networks from the single perceptron up to CNNs, LSTMs, attention and transformers. The DS course page lists no prerequisite, and every DS student entering the BSc level must take it. It is also the gateway to several AI electives and minors.

Deep Learning at a glance

The instructor is Prof. Mitesh M. Khapra of IIT Madras. The course page calls the type "Core Option II", and the DS handbook tags it "Core_BD" with 4 credits and no prerequisite. The table (updated 18 March 2026) marks it as offered in May 2026, September 2026 and January 2027.

One mismatch: the ES handbook lists Deep Learning as an open elective for Electronic Systems students with Machine Learning Techniques (CS2007) as a prerequisite. One ES table also lists Machine Learning Foundations (CS2004) as a corequisite. The DS page lists neither. DS students finish both in the diploma anyway.

What you learn in Deep Learning

The 12 weeks fall into four parts.

• Weeks 1 to 3: the basic neuron to full networks. A short history, the McCulloch Pitts neuron and the perceptron learning algorithm. Then multilayer perceptrons, sigmoid neurons, gradient descent, what feedforward networks can represent, and backpropagation.
• Weeks 4 to 7: training networks well. Gradient descent variants such as momentum, Nesterov, stochastic GD, Adagrad, RMSProp and Adam, plus learning rate schedules. Autoencoders and how they relate to PCA. The bias variance tradeoff and ways to regularise, like L2, early stopping, data augmentation and dropout. Better activation functions, weight initialisation and batch normalisation.
• Weeks 8 and 9: images. Word vectors, then convolutional networks from LeNet and AlexNet to VGGNet, GoogLeNet and ResNet. How to visualise what a CNN has learned, and how CNNs can be fooled.
• Weeks 10 to 12: sequences. Recurrent networks, backpropagation through time and the vanishing gradient problem. GRUs and LSTMs as the fix. Encoder decoder models, attention and transformers.

The suggested books are Deep Learning by Goodfellow, Bengio and Courville, and Neural Networks and Deep Learning by Charu Aggarwal.

How Deep Learning is assessed

The course page lists weekly online assignments, 2 in-person invigilated quizzes and 1 in-person invigilated end term exam. It does not list an OPPE or a project.

Where it counts

Deep Learning is one of the five mandatory courses of the BSc degree level. It also opens up later courses. In the DS handbook table, it is a prerequisite for Large Language Models (BSDA5004), Deep Learning Practice (BSDA5013) and the Data Science and AI Lab (BSDA4001). It is a corequisite for Reinforcement Learning (BSDA5007). Deep Learning Practice in turn sits in the Generative AI and Cloud Computing for AI minors, and is the prerequisite of the Multimodal AI Systems minor. See the Minor in Generative AI for one example.

Who finds it hard and how to prepare

The course is mathematical. Backpropagation is repeated use of the chain rule with vectors and matrices. If you rushed through derivatives or matrix work earlier, the early weeks will feel heavy.

• Revise the chain rule and matrix shapes. Before week 3, practise working out gradients for a two layer network by hand, keeping track of every matrix size.
• Revise PCA from MLT before week 5. Machine Learning Techniques (BSCS2007) starts with PCA, and the autoencoder week builds on it.
• Build a tiny network in NumPy. One hidden layer, trained on a toy dataset. Writing the forward and backward pass yourself makes backpropagation clear.
• Keep an optimiser sheet. Write each gradient descent variant's update rule in one line with a note on what problem it fixes.
• Draw an LSTM cell. Label every gate and say in words what it keeps or throws away.

What to take before and after

Before: the diploma ML courses, mainly MLF and MLT, and Maths 2 for the linear algebra. Next to it: AI: Search Methods for Problem Solving (BSCS3003), the other AI core course. After: Deep Learning Practice, Large Language Models or the Data Science and AI Lab, depending on the path you want.

Common questions

Should I take Deep Learning in my first BSc term?

It is mandatory, and it is the entry point for several electives and three AI minors. Taking it early keeps those options open in later terms.

Does the course page name a framework like PyTorch?

No. The page lists topics, books and the assessment pattern but no software framework. The more hands-on course is Deep Learning Practice, which needs Deep Learning first.

Can Electronic Systems students take Deep Learning?

The ES handbook lists it as an open elective for ES students. There, Machine Learning Techniques is a prerequisite, so plan that course first and confirm with support.

Is this the same as Introduction to Deep Learning and Generative AI?

No. That is BSDA2001, a diploma level option in the Diploma in Data Science. Deep Learning (BSCS3004) is the degree level core course.