IITM BS Deep Learning Practice (BSDA5013): Syllabus and Tips

Deep Learning Practice (DLP, BSDA5013) is a 4 credit elective at the degree level of the IITM BS in Data Science. The handbook places it at level 5. You need Deep Learning (BSCS3004) before you take it. The course is about using deep learning on real tasks: four weeks each on text, speech and images.

The course is taught by Prof. Mitesh M. Khapra (Data Science and AI), Prof. S. Umesh and Dr. Kaushik Mitra (both Electrical Engineering), all at IIT Madras. The course page and the handbook agree on the credits and the prerequisite. The ES handbook also lists DLP as an open elective for ES students, with the same prerequisite.

What you learn in Deep Learning Practice

The page sets goals beyond single models: datasets, frameworks, training hardware, deployment on different devices, interpretability and security. It also stresses training well when data or hardware is limited. The 12 weeks fall into three blocks.

• Weeks 1 to 4: text. Modern NLP with Hugging Face, and how tokenisation works. Then fine-tuning models for specific tasks. Then continuing pre-training on new data and instruction tuning.
• Weeks 5 to 8: speech. Identifying which language is spoken, and working out who spoke when in a conversation. Then speech to text and text to speech. Then wake word detection, like the phrase that wakes a voice assistant, with personalisation.
• Weeks 9 to 12: vision. Image classification and object detection. Then estimating depth from images, and image super-resolution.

How the course is assessed

The course page points to the standard course structure on the programme's Academics page. It does not list a separate assessment plan.

Where it counts

DLP appears in three minor lists:

One course can count for only one minor. The handbook uses DLP as its own example and says it appears in 4 minors, though the lists show it in the three above. Either way, you can claim it in one minor only. The handbook does not say whether using it as the Multimodal prerequisite counts as a claim. Ask support before you plan two of these minors. See the Generative AI minor, the Cloud Computing for AI minor and the Multimodal AI Systems minor.

For the BS level, DLP is tagged BD/BP. The handbook does not say which stream a BD/BP course counts for, so ask support. The DS course table (updated 18 March 2026) marks it for May 2026, September 2026 and January 2027.

Who finds it hard and how to prepare

Each block is a new field. If you have only worked with text and tables, the speech weeks will feel the least familiar, because audio has its own formats and features. The work also needs real compute, since the page talks of training models at real-world scale.

• Get fluent in a framework such as PyTorch before week 1. You should be able to write a training loop without copying one.
• Learn the basics of the Hugging Face libraries: loading a model, a tokenizer and a dataset.
• Before week 5, learn how audio is stored (sample rate, waveform) and what a spectrogram shows.
• Plan your GPU access for the term. Check what the course provides and keep a backup option.
• Keep a short log for each task: data used, model, settings and result. It saves time when later weeks reuse ideas.

What to take before and after

Deep Learning (BSCS3004) is required first. After DLP, pick the path that matches your minor. Large Language Models (BSDA5004) goes deeper into text models. Speech Technology (BSEE4001) is a full course on speech recognition and synthesis, and Prof. Umesh teaches it too.

Common questions

Is DLP the same as the Data Science and AI Lab?

No. The lab (BSDA4001) is a level 4 course that covers many tools, from dashboards to LLM apps and cloud. DLP is a level 5 course focused on training deep models for text, speech and vision tasks.

Can I take DLP and Deep Learning in the same term?

No. Deep Learning is a prerequisite for DLP, not a corequisite. You need to have completed it first.

Which minor should I use DLP for?

Pick the minor whose other two courses you are most likely to finish. Introduction to Big Data, needed for Cloud Computing for AI, is marked only for September 2026 in the March 2026 table. If you also want Multimodal AI Systems, ask support how its DLP prerequisite interacts with claiming DLP elsewhere.