Olympiad problems
IOAI-style challenges: adversarial attacks, pruning, self-supervised learning, matching, multi-task
In this chapter, you will learn to
- Tackle a new IOAI-style problem under time pressure
- Combine several techniques from the course
- Build your own validation to score yourself honestly
Tools: PyTorch, scikit-learn
9 exercises
Each exercise is graded automatically, in your browser. The full statements are available once you have an account.
What this is for in real life
Real problems never come labelled with the chapter that solves them. Nor do these: each problem starts from a situation, and it is up to you to work out what to do with it. - Model security. In 2018, researchers…
The cheque reader
A bank wants to automate reading the handwritten amounts on cheques. It has trained a small convolutional network that reads each digit from a 16 × 16 pixel grayscale image. On its test digits, it almost never makes a…
The surgeon
Second event. You are given a trained neural network with 45 405 parameters (weights and biases combined). The network has a fixed architecture: input (20) -> hidden layer (200) -> hidden layer (200) -> output (5).…
The self-supervised learner
Fourth event. You have 500 unlabeled samples (dimension 20) and only 10 labeled samples (2 per class, 5 classes). The classes overlap: a classifier trained on the 10 labeled examples alone plateaus around…
The decipherer
Fifth event. You are given 200 encrypted messages and 200 plaintext messages, but in a different order. Each encrypted message corresponds to exactly one plaintext message. The encryption is a…
The final challenge
Last event of the series. You are given images (vectors of dimension 30) and you must predict 3 things simultaneously: 1. Class (5 classes, 0-4): which object is in the image 2. Rotation angle (real, 0 to…
The artificial signal
You join the team of a radio telescope that listens to the sky day and night. Each observation is a spectrogram: a 64 × 64 image where the row is a frequency, the column an instant, and the intensity the received power.…
The forms audit
A start-up sells software that reads handwritten digits on forms. Its model was trained on the scanned forms of 1,000 customers. A consumer association suspects that forms were used without their authors' consent, and…
The right to be forgotten
A year after the audit, the company has learned its lesson: when a customer asks for their forms to be erased, the software must, from the next delivery on, behave as if it had never seen them. Requests arrive in waves…