An Idea Board for ML Students Tired of MNIST

What it does

This repository is a well-organized idea board for machine learning and deep learning practitioners who have compute cycles but no target. It catalogs more than thirty project concepts across NLP, computer vision, time-series forecasting, recommender systems, and audio, pairing each suggestion with publicly available datasets and relevant references. The scope runs from beginner-friendly classifiers to research-grade medical imaging tasks.

The interesting bit

Instead of dumping links, the author frames every idea as an answerable question—“Can you predict if a blood donor will donate?” or “Can computer vision spot distracted drivers?"—which forces the scope to fit in a sentence. A newer “Hackathon Ideas” section folds in LLM-era tooling like InstructPix2Pix, BLIP-2 embeddings, and ControlNet LoRA fine-tuning, so the list is aging into the generative-AI wave rather than fossilizing in 2017.

Key highlights

• Six domains: Text, Vision, Forecasting, Recommendation Systems, Music/Audio, and LLM-powered Hackathon Ideas
• Each entry typically includes a specific dataset (e.g., Netflix Prize, MURA, YouTube 8M, Quora question pairs)
• Explicitly pitched at both academia and industry
• Difficulty ranges from entry-level (rainfall forecasting, MNIST) to research-grade (clinical lung-cancer detection, video summarization)
• Includes contemporary generative-AI angles like text-to-music, semantic media search, and natural-language YAML editing

Caveats

• This is a reading list, not a framework; there is no code, boilerplate, or training scripts provided
• Some dataset links point to external competition pages that may have shifted or closed since curation
• The README admits that video-summarization lacks “standard dataset or agreed upon metrics,” making that section more speculative than prescriptive

Verdict

Grab this if you are a student, bootcamp grad, or hackathon team starving for a viable project brief with data already attached. Skip it if you are looking for reference implementations, libraries, or copy-paste training loops.