Awesome Deep Learning Project Ideas

A curated list of practical deep learning and machine learning project ideas

• 30+ ideas
• Relevant to both the academia and industry
• Ranges from beginner friendly to research projects

Contents

• Hackathon Ideas - Project ideas unlocked by use of Large Language Models, specially text to text -- note that a lot of the text to text ideas can also be buit a lot better with LLMs now!

• Text - With some topics about Natural language processing

• Forecasting - Most of the topics in this section is about Time Series and similar forecasting challenges

• Recommendation Systems

• Vision - With topics about image and video processing

• Music and Audio - These topics are about combining ideas from language and audio to understand music

• Conclusion

Hackathon Ideas

• Developer Ideas

  • Text to cmd for terminal: Take user intent in terminal e.g.

    $ask "how to list all files with details"
    > Execute "ls -l"? [y/N] y 
    $ls -l
    

  • Build and edit YAMLs using natural language e.g. Kubernetes and other form of config files

    • Kor for ideas on how this is done for JSON
    • Can be use-case specific. Build pipelines? Kube?

  • Mobile Android/iOS SDK for Stable Diffusion inference

    • Apple has released a CoreML Stable Diffusion Inference

• Voice powered Experiences

  • Audio Conversation with chatGPT, can combine with fast Text-to-Speech e.g. Eleven Labs to have a two-way conversation
  • Telegram/WhatsApp bot to get audio and save as text with metadata into mem.ai or Roam Research or Obsidian

• Edit image by giving instructions of what you want to do: SeeChatGPT and playgroundai.com as examples

  • The underlying mechanism which you can use is called InstructPix2Pix

• Semantic search over any media

  • Can build using CLIP or BLIP-2 embeddings for images and CLAP for all audio including music and speech

• Text to Music Generation

  • See MusicLM for reference

• Knowledge Base QA aka Answer Engines

  • Take any plaintext dataset e.g. State of the Union address and build on top of that
  • Can use this over Video Subtitles to search and QA over videos as well, by mapping back to source

• Guided Summarisation/Rewriting

  • Take specific questions which the user might have about a large text dataset e.g. a novel or book and include that in your summary of the piece
  • Pay attention to specific entities and retell the events which happen in a story with attention to that character

• ControlNet + Stable Diffusion for Aethetic Control

  • Build tooling using diffusers which takes in a set of photos, finetunes a model (LoRA) on a person, detects face and moves it to a new aesthetic e.g. futuristic neon punk, grunge rock, Studio Ghibli. Can also add InstructPix2Pix to give user more control.

• Text to Code/SQL

  • Use code understanding to convert use query to SQL or another executable programming language, including Domain Specific Languages
  • Here is an example of the same: qabot

Text

• Autonomous Tagging of StackOverflow Questions

  • Make a multi-label classification system that automatically assigns tags for questions posted on a forum such as StackOverflow or Quora.
  • Dataset: StackLite or 10% sample

• Keyword/Concept identification

  • Identify keywords from millions of questions
  • Dataset: StackOverflow question samples by Facebook

• Topic identification

  • Multi-label classification of printed media articles to topics
  • Dataset: Greek Media monitoring multi-label classification

Natural Language Understanding

• Sentence to Sentence semantic similarity

  • Can you identify question pairs that have the same intent or meaning?
  • Dataset: Quora question pairs with similar questions marked

• Fight online abuse

  • Can you confidently and accurately tell whether a particular comment is abusive?
  • Dataset: Toxic comments on Kaggle

• Open Domain question answering

  • Can you build a bot which answers questions according to the student's age or her curriculum?
  • Facebook's FAIR is built in a similar way for Wikipedia.
  • Dataset: NCERT books for K-12/school students in India, NarrativeQA by Google DeepMind and SQuAD by Stanford

• Automatic text summarization

  • Can you create a summary with the major points of the original document?
  • Abstractive (write your own summary) and Extractive (select pieces of text from original) are two popular approaches
  • Dataset: CNN and DailyMail News Pieces by Google DeepMind

• Copy-cat Bot

  • Generate plausible new text which looks like some other text
  • Obama Speeches? For instance, you can create a bot which writes some new speeches in Obama's style
  • Trump Bot? Or a Twitter bot which mimics @realDonaldTrump
  • Narendra Modi bot saying "doston"? Start by scrapping off his Hindi speeches from his personal website
  • Example Dataset: English Transcript of Modi speeches

Check mlm/blog for some hints.

• Sentiment Analysis
  • Do Twitter Sentiment Analysis on tweets sorted by geography and timestamp.
  • Dataset: Tweets sentiment tagged by humans

Forecasting

• Univariate Time Series Forecasting

  • How much will it rain this year?
  • Dataset: 45 years of rainfall data

• Multi-variate Time Series Forecasting

  • How polluted will your town's air be? Pollution Level Forecasting
  • Dataset: Air Quality dataset

• Demand/load forecasting

  • Find a short term forecast on electricity consumption of a single home
  • Dataset: Electricity consumption of a household

• Predict Blood Donation

  • We're interested in predicting if a blood donor will donate within a given time window.
  • More on the problem statement at Driven Data.
  • Dataset: UCI ML Datasets Repo

Recommendation systems

• Movie Recommender

  • Can you predict the rating a user will give on a movie?
  • Do this using the movies that user has rated in the past, as well as the ratings similar users have given similar movies.
  • Dataset: Netflix Prize and MovieLens Datasets

• Search + Recommendation System

  • Predict which Xbox game a visitor will be most interested in based on their search query
  • Dataset: BestBuy

• Can you predict Influencers in the Social Network?

  • How can you predict social influencers?
  • Dataset: PeerIndex

Vision

• Image classification

  • Object recognition or image classification task is how Deep Learning shot up to it's present-day resurgence
  • Datasets:
    • CIFAR-10
    • ImageNet
    • MS COCO is the modern replacement to the ImageNet challenge
    • MNIST Handwritten Digit Classification Challenge is the classic entry point
    • Character recognition (digits) is the good old Optical Character Recognition problem
    • Bird Species Identification from an Image using the Caltech-UCSD Birds dataset dataset
  • Diagnosing and Segmenting Brain Tumors and Phenotypes using MRI Scans
    • Dataset: MICCAI Machine Learning Challenge aka MLC 2014
  • Identify endangered right whales in aerial photographs
    • Dataset: MOAA Right Whale
  • Can computer vision spot distracted drivers?
    • Dataset: State Farm Distracted Driver Detection on Kaggle

• Bone X-Ray competition

  • Can you identify if a hand is broken from a X-ray radiographs automatically with better than human performance?
  • Stanford's Bone XRay Deep Learning Competition with MURA Dataset

• Image Captioning

  • Can you caption/explain the photo a way human would?
  • Dataset: MS COCO

• Image Segmentation/Object Detection

  • Can you extract an object of interest from an image?
  • Dataset: MS COCO, Carvana Image Masking Challenge on Kaggle

• Large-Scale Video Understanding

  • Can you produce the best video tag predictions?
  • Dataset: YouTube 8M

• Video Summarization

  • Can you select the semantically relevant/important parts from the video?
  • Example: Fast-Forward Video Based on Semantic Extraction
  • Dataset: Unaware of any standard dataset or agreed upon metrics? I think YouTube 8M might be good starting point.

• Style Transfer

  • Can you recompose images in the style of other images?
  • Dataset: fzliu on GitHub shared target and source images with results

• Chest XRay

  • Can you detect if someone is sick from their chest XRay? Or guess their radiology report?
  • Dataset: MIMIC-CXR at Physionet

• Clinical Diagnostics: Image Identification, classification & segmentation

  • Can you help build an open source software for lung cancer detection to help radiologists?
  • Link: Concept to clinic challenge on DrivenData

• Satellite Imagery Processing for Socioeconomic Analysis

  • Can you estimate the standard of living or energy consumption of a place from night time satellite imagery?
  • Reference for Project details: Stanford Poverty Estimation Project

• Satellite Imagery Processing for Automated Tagging

  • Can you automatically tag satellite images with human features such as buildings, roads, waterways and so on?
  • Help free the manual effort in tagging satellite imagery: Kaggle Dataset by DSTL, UK

Music

• Music/Audio Recommendation Systems

  • Can you tell if two songs are similar using their sound or lyrics?
  • Dataset: Million Songs Dataset and it's 1% sample.
  • Example: Anusha et al

• Music Genre recognition using neural networks

  • Can you identify the musical genre using their spectrograms or other sound information?
  • Datasets: FMA or GTZAN on Keras
  • Get started with Librosa for feature extraction

FAQ

• Can I use the ideas here for my thesis? Yes, totally! I'd love to know how it went.

• Do you have any advice before I start my project? Advice for Short Term Machine Learning Projects by Tim R. is a pretty good starting point!

• How can I add my ideas here? Just send a pull request and we'll discuss?

• Hey, something is wrong here! Yikes, I am sorry. Please tell me by raising a GitHub issue.

  I'll fix it as soon as possible.

Acknowledgements

Problems are motivated by the ones shared at:

• CMU Machine Learning
• Stanford CS229 Machine Learning Projects
• swyx

Credit

Built with lots of keyboard smashing and copy-pasta love by NirantK. Find me on Twitter!

License

This repository is licensed under the MIT License. Please see the LICENSE file for more details.