mics-lab/deeplearningcourse

Ressources for Deep Learning Course (VSE).

Deep Learning Course

Ressources for Deep Learning Course (VSE).

Grade breakdown:

• Kaggle challenges: 30% of the total grade
  • Deadline on the 15th of March 2023, 5pm
  • Fitting functions (1D-1D & 5D-5D)
  • Binary face classification (with glasses or without glasses)
  • Multi (10) classes classification
  • Transfer Learning
• Group Project: 70% of the grade
  • 2-3 per group
  • Deadline on the 2nd of April 2023, 11:59 pm
  • Must demonstrate understanding of deep learning & ability to lead a project
  • Easiest projects that are thoroughly understood are preferred to complicated projects that are only scratched the surface
  • Slides and written report must be in English. Presentation doesn't need to.

Projects Ideas

"Mathy" projects:

• deep to shallow: deep (>2 hidden layers) to shallow (1 hidden layer) network
• perturb weights of a network to assess which neurons are the most “sensitives” (maybe compare wide and deep?)

CNN projects:

• follow up on digits classification: hand written characters classification
• measure how far two handwritings are from images samples (use siamese networks to create a distance)
• follow up on TD8: from medical decathlon (http://medicaldecathlon.com/), try to segment other organs (we will do prostate and hippocampus in TD8 & TD8bis)
• realistic person not existing: this website generates artificial images of people some are more realistic than others; create a network that discriminate realistic from unrealistic ones

RNN projects:

• predict the language spoken from some audio record
• predict the area of math to which belong the latex formula

GAN projects:

• conditional gans on pokemons: generate intermediate frames for the evolution of pokemons (this could be a beginning: https://becominghuman.ai/generating-new-pokemons-using-gans-ceba1c6dc676)
• mixing faces: take images of faces and create a mixture of them, you can then create the "average face" of your group, your family, etc...
• how will our babies look like: take two face images and generate a younger version of the mixed faces

NLP projects:

• autocomplete based on previous messages of the conversation
• automatic texting in a conversation (similar to the above, but more ambitious... How long can you hold before the other person finds out? this is called the Turing test)
• automatic group conversation image generation based on conversation title/messages

Drive folder with datasets:

https://drive.google.com/drive/folders/1_KqzmBVvZKwvLi0a1W3P6wPa-IgwnmLP?usp=share_link

Colab notebook example that mounts your drive:

https://colab.research.google.com/drive/1Ummw01RU9TeSnmoUIBAPR88wfTpLvZUD?usp=sharing

Kaggle links:

• Fitting function 1D-1D: https://www.kaggle.com/t/81c8fa412fb840c0b34fe86e7dab174f
• Fitting function 5D-5D: https://www.kaggle.com/t/3eb9e2aa517a414aacd0c9468281a7c2
• Binary Classification (Glasses/NoGlasses): https://www.kaggle.com/t/a232e4d3385b46c68921719e547b45e1
• Decimal Classification (MICS MNIST): https://www.kaggle.com/t/073954a6c84a4970903d5a340df28695
• Transfer Learning (Sword video classification): https://www.kaggle.com/t/d6ca6181f7eb4acc93f418f7b43cbece

---

Course ressources from last year: https://m2dsupsdlclass.github.io/lectures-labs/