M.Tech Thesis Topics in AI and Machine Learning

Picking an AI thesis that will survive the next model release

If your thesis says only that a new model scored higher on a dataset, a newer model can make that finding irrelevant in months. A more durable thesis studies why methods fail, how to measure them properly, or how to make them cheaper and safer to use.

Also be honest about compute. Training a large model from scratch is rarely practical for a master's student, but evaluation, retrieval and on-device studies can all produce sound research on a small budget.

Large language models and generative AI

1. Detecting unsupported claims in RAG answers — Open-weight LLMs, NLI models, FAISS, RAGAS as baseline
   Retrieval-augmented systems can state things their sources do not support. Build a sentence-level entailment checker and evaluate how accurately it flags such claims compared with existing faithfulness metrics.
2. Chunking strategy effects on retrieval for long technical documents — Python, LlamaIndex, BM25 and dense retrieval
   Fixed-size chunks can break tables and cross-references. Propose structure-aware chunking and measure retrieval quality on standards documents and manuals, with a labelled question set you create.
3. Quantisation versus reliability in small language models — llama.cpp, GGUF formats, benchmark question sets
   Compressed models are smaller, but the effect on answer quality is uneven. Study which task types degrade first under quantisation, and propose a test to predict that degradation cheaply.
4. Prompt injection defence for tool-using agents — Agent frameworks, sandboxed tools, red-team prompt sets
   Agents reading untrusted text can be tricked into wrong actions. Design a layer that separates data from instructions, and report attack success rate against the loss of task completion.
5. Machine unlearning verification for fine-tuned models — PyTorch, small open models, membership inference attacks
   Removing specific training data from a model is hard to confirm. Compare unlearning methods and develop a verification test that goes beyond asking the model direct questions.
6. Uncertainty-aware answering with calibrated abstention — Conformal prediction, open-weight LLMs, domain QA datasets
   Models often answer confidently when they should decline. Calibrate an abstention threshold using conformal prediction, then measure the coverage and error trade-off across domains.

Indian-language and code-mixed NLP

1. Marathi-English code-mixed sentiment and intent analysis — IndicBERT, MuRIL, transliteration tools, self-annotated data
   Romanised, code-mixed social text confuses models trained on standard script. Compare transliteration-first and direct approaches, and measure which errors each one reduces.
2. Low-resource speech recognition for regional dialects — Whisper fine-tuning, data augmentation, WER analysis
   Models trained on standard speech struggle with dialects. Study how much a small amount of dialect data improves performance when fine-tuning, using your own carefully consented recordings.
3. Cross-lingual retrieval for Indian government documents — IndicTrans, multilingual embeddings, FAISS
   Users may ask in one language while documents are in another. Compare translation-based and multilingual embedding retrieval and study failures on named entities and legal terms.

Computer vision

1. Few-shot defect detection for small manufacturers — DINOv2 features, PatchCore, MVTec-style datasets
   Small factories cannot label thousands of defects. Evaluate self-supervised pretraining with few labelled samples and find how few are needed for stable detection on industrial anomaly data.
2. Domain shift in crop disease diagnosis — PyTorch, PlantVillage-style data, domain adaptation
   Models trained on lab images often fail on field photographs. Measure this gap, then test domain adaptation methods using a small field set you collect.
3. Detecting AI-generated images after social media compression — CNN and ViT detectors, diffusion image sets, JPEG simulation
   Detectors can lose accuracy after resizing and recompression. Train with simulated platform degradation and report robustness across generators the detector has not seen.
4. Object detection under Indian traffic density — YOLO family, IDD-style datasets, edge inference testing
   Dense mixed traffic with autorickshaws and heavy occlusion stresses standard detectors. Propose an occlusion-aware training approach and evaluate it on regional driving data with a small-object focus.

Learning methods, trust and efficiency

1. Federated learning with heterogeneous clients — Flower, PyTorch, FedAvg and FedProx baselines
   Non-identical data and uneven hardware slow convergence. Design a client selection and aggregation strategy, then compare accuracy, fairness and communication cost across simulated clients.
2. Stability of explanations under small input changes — SHAP, LIME, gradient boosting, credit or health tabular data
   Explanations such as SHAP values can vary when inputs barely change. Define a stability metric, test it on tabular risk models and propose a regularisation that improves it.
3. Continual learning without storing raw data — Avalanche library, PyTorch, standard incremental benchmarks
   Replay buffers can raise privacy concerns. Compare feature-based or generated replay with raw replay, and measure forgetting against the privacy and storage benefit.
4. Graph neural networks for fraud under label delay — PyTorch Geometric, public transaction datasets
   Fraud labels arrive weeks after the transaction. Build a temporal graph model that handles delayed labels and evaluate it with a strict time-based split.
5. Fairness audit of ML-based resume screening — Fairlearn, sentence embeddings, synthetic resume generation
   Screening models can learn proxies for gender or region. Create a counterfactual testing method using synthetic resumes, then evaluate mitigation methods and their effect on ranking quality.
6. Time-series forecasting with missing sensor data — Transformers for time series, PatchTST, public energy data
   Forecasting models often assume complete input. Compare imputation-first pipelines with models that handle gaps directly, across different missingness patterns.
7. Clinical note summarisation with factuality checks — Open-weight LLMs, MIMIC-style data with approved access
   Summaries of medical notes must not add facts. Combine a summariser with a claim verifier on de-identified public data and measure hallucination rate alongside readability, without making any clinical claims.

What examiners look for in an AI thesis

Weak AI theses often report a single impressive accuracy figure. Stronger ones show how it was obtained.

• Fair baselines: include simple methods as well as recent ones. If logistic regression nearly matches your model, that is worth reporting.
• No data leakage: split by time, user or document where relevant, and fit preprocessing on the training set only.
• Repeated runs: report mean and spread over several random seeds, and use a suitable statistical test before claiming an improvement.
• Suitable metrics: use precision, recall or F1 on imbalanced data, and report inference time and memory for anything meant to be deployed.
• Error analysis: show where your method fails and try to explain why.
• Ethics and licensing: check dataset and model licences, and get any required approval and consent before collecting data from people.

Getting started on limited compute

Get the whole pipeline running end to end on a small data subset and a small model before you scale up. Log every experiment from the first day, using MLflow or a simple spreadsheet. On free cloud notebooks, save checkpoints often since sessions can end without warning.

Road2Tech in Pune mentors M.Tech students on reference papers and algorithm understanding. The thesis and its findings remain your own work.