A course on biological and medical applications of deep learning.
WIP
- Block 1. Classic NLP and Transformers
- Word2Vec: Continuous Bag-of-Words and Skip-Gram. TF-IDF. ELMo
- Revisiting transformers: Encoders vs Decoders vs Full Transformers. BERT family, GPTs, T5.
- Vector Search. Algorithms for fast nearest neighbours search.
- Block 2. Images
- Visual transformers. Self-supervision on images. SimCLR, DINO, V-JEPA
- Segment Anything models.
- Block 3. Graphs
- Classical graph methods. Centrality, clustering, DeepWalk, Node2Vec and Matrix decomposition.
- Graph Neural Networks. GCN, GAT. Heterogeneous Graphs, a bit on Knowledge Graphs, RGCN.
- Proteins. GNN in 3D, equivariance in deep learning. AlphaFold 2, ProteinMPNN.
- Block 4. Reinforcement Learning
- A concise and shallow introduction to RL. Markov Decision Processes. Policy and Value. Q-learning. DQN. ()
- Policy Gradient Methods. REINFORCE. PPO
- Block 5. Generative AI
- LLM Applications. RAG.
- Diffusion. Generating images. Schedules.
- Diffusion for chemical compounds, 3D structures (proteins and peptides). RFDiffusion. Latent Diffusion Models
As homeworks are on the Kaggle platform, here is instruction on how to register for a competition (from Russia):
- Log in to your account on Kaggle (or create one)
- Open the link to a competition
- Click on the "Join competition" button and accept the rules
- If a phone number is required, do the following (Russian phone numbers are not accepted):
- Rent a phone number on any suitable service, for example:
- Register for onlinesim.io
- Top up the balance in the left corner (15.60 RUB is sufficient - Yes, you have to invest your money to the course... but Kaggle and onlinesim accounts are yours and you can use them in the future)
- Open Receive SMS
- Find country: Sweden +46, find a site: Kaggle, rent a phone number by clicking on the price button
- You have a number on the right side that you need to enter into kaggle to verify your account
- Now wait for the SMS with the code in the box on the right, update if necessary
- Rent a phone number on any suitable service, for example:
- Now you have a Kaggle account! Let's see
- There are different tabs in the competition, the most important: Overview, Data, Leaderboard:
- Overview: competition description
- Data: here you can download train data for training and test data for competition, as well as sample_submission.csv (see below)
- Leaderboard: list of competitors and their scores
- To send a solution click on the Submit predictions button on the right and upload a .csv file the same as sample_submission.csv