Tasks - SprawdzAI

A complete introductory course on the PyTorch library: interactive tasks, intuitions, tips, and tricks. You will learn all the topics needed for olympiads as well as indi...

Difficulty:

1,6

MIKO

MIKO write-up

PyTorch

Casino MIKO

Based on a sequence of cards from a casino game, implement an algorithm that maximizes your winnings.

The lost epic

Repair the tokenizer's damaged mapping using its original BERT model and the text of "Pan Tadeusz".

Interworld olympiad

Help organize an interplanetary olympiad by determining a time and planetary motions that minimize the cost of their relative positions.

Self-reconstruction

Reconstruct every missing pixel in the images without making use of any training dataset.

Adversarial defense

Protect the model from adversarial attacks by modifying only its loss function.

Antique painting authentication

Using the digital features of antique paintings, help an art critic determine whether they are originals or forgeries.

Image reconstruction

A task demonstrating the capabilities of models that produce images as output (e.g. U-Net), involving the denoising of cat photos.

MikoMark

Fix the model so that it works correctly when the text MIKO is added to the images, without training any parameters!

Titanic

One of the best-known examples of classical machine learning - use passenger information to try to predict who managed to survive the Titanic disaster. Deal with missing ...

Word segmentation

Split compound words in German sentences into short words separated by spaces.

Adversarial Attacks

Attack a convolutional network - modify the images (imperceptibly to the human eye) so they completely fool a pre-trained model!

Cream-roll trees

Create a model that processes photographs of physical structures made of tubes with different flavors into their digital representation.

Anomaly Detection

Can a model be taught to recognize problems it has never seen before? Teach a neural network to understand 'normality' and use this knowledge to find anomalies.

Pixel parsimony challenge

Identify the most important areas in wildlife photographs to maintain high CLIP classification quality after removing most of the pixels from the image.

Image autoencoder

Write and train a model to compress image sketches.

Dimensionality reduction

PyTorch

Sentiment analysis

A simple introductory task in the field of natural language processing, using the classification of the sentiment of film reviews as an example.

Classifier in two minutes

Teach a model to recognize activity based on multivariate time series. Use unlabelled data to build a representation that remains effective even with a small set of label...

Data Prototypes

Create a small set of embeddings that achieves the highest possible accuracy for the one-nearest-neighbor algorithm.

Hallucination Detection

Language models can be very convincingly wrong. Create a solution that assesses the reliability of their answers to factual questions using data from the generation proce...

Data splitting

Estimate the "scaling law" of a machine learning model whose training set leaked into the test set.

Photos of cars and cats

Train a model to classify images using a very small labeled dataset, but a huge unlabeled dataset.

Mikomarket

Using a pretrained BERT model, assign politicians' tweets to their authors.

Operation "Retreat"

Use a huge dataset and train language models to solve the fascinating problem of embedding model inversion - you have certain vectors at your disposal and must guess whic...

One

A purely implementation-focused task involving the generation of new samples of the handwritten digit “1” and testing your algorithm yourself with a KNN model trained to ...

Detecting ECG Signal Abnormalities

Design a lightweight solution for an embedded device: four meta-features, a small random forest, and five classes describing normal rhythm and ECG abnormalities.

High-flying designs

Design a model that can quickly evaluate thousands of rocket ideas. Based on twelve design parameters, predict four key flight simulation results.

Clustering

A simple task demonstrating a possible application of clustering techniques.

Hidden Subsequences

Three important subsequences are hidden in binary sequences. Train a model that recognizes their influence and reproduces the secret function as accurately as possible.

Double Agent Dilemma

Make two accurate image classifiers disagree using tiny pixel changes. Fool each model while keeping the other correct, and earn more points for less visible perturbation...

Continuous sentiments

Use pretrained word embeddings to learn from a small number of comparisons and recognize the sentiment of new synonyms. Build a ranking whose sentence order agrees with t...

Training task

Train a CNN to recognize 16 classes of small images. Plan the validation yourself, limit overfitting, and ensure good performance for the rarest classes as well.

Difficulty:

?

Ghost of the Machine

Find the moment a human-written passage becomes machine-generated. Analyze seemingly coherent stories and pinpoint the exact character where the author changes from a per...

Coin Counter

Teach a model to calculate the value of scattered coins by detecting each one, recognizing its denomination, marking a precise rectangle, and determining the prediction c...

The Madarian Cow Mystery

Modify the input to the image-generating diffusion model so that cows always appear with a fire hydrant - without affecting other objects.

Restroom icon matching

Using a pretrained CLIP model, write an algorithm capable of matching restroom icons with other icons of the same type but of the opposite gender.

Machine (un)learning

Unlearn a specified class from the model so that it can no longer classify it correctly, while preserving prediction quality on the remaining classes.

Pruning

Zero out some of the model's weights without reducing the quality of its predictions.

Color Quantization

Create an algorithm that reduces the number of colors in an image while preserving its quality and preferring bright, vivid colors.

Chicken counting

Help farmers and refine a model that can count the number of chickens in a given image using density estimation techniques.

Collages

Train a model to classify collages using a very small amount of training data.

Non-normal Distribution

Create an image-denoising model that also identifies the type of noise and its parameters.

Motion Classification

Reduce the dimensionality of the human body coordinates, and then determine what type of motion the person performed.

Dimensionality reduction

Find the Order

Reconstruct the chronological order of the English dialogues by arranging the shuffled audio clips. Use the provided opening of the conversation, as well as the connectio...

Artificial inception

Some images in the dataset were created by a generative model. Use clever features and data analysis to discover which of them are real and which were generated.

Credit Decisions

Create an algorithm for interpreting the decisions of a classification model.

Search for Dark Matter Halos

Design a machine learning model to predict the positions of dark matter halos based on weak gravitational lensing convergence maps.

Radar (on-site)

Develop a model that performs radar semantic segmentation, that is, predicts where people and other objects are based on actual radar readings.

Painter's Optimizer

Decompose the image into the parameters of all the circles that make it up.

Noise in Data Labels

Training data can be misleading. Design a sample selection strategy for two models so that they can learn despite label noise and class imbalance.

Robot Chasing

Learn how six robots respond to instructions and predict their next actions from room snapshots. Each robot follows its own consistent pattern.

Chemical Reaction Kinetics Simulation

Create a model that learns to perfectly reproduce organometallic reaction curves based on time and concentration measurements.

Potato

Find a hidden word in at most 30 turns. Use semantic comparisons and public word embeddings to choose each guess, earning more points for earlier discoveries.

Inpainting

Complete missing parts of a video using an INR network.

Source Extraction

Develop an embedding algorithm for efficiently searching for source documents matching scientific queries.

IOAI Field

Train a neural network to recreate a stylized city logo from coordinates. Capture sharp boundaries, enormous gradients, spiral patterns, and a letter whose values change ...

Translation Styling

Are you tired of strange translations of AI terms into Polish? Based on a pretrained translation model, build a system that preserves the original English terms and gives...

Decision trees

Select data preprocessing and tree parameters to classify many different point datasets well. Each dataset places slightly different requirements on the model.

Semantic changes in words

A word’s meaning can change over the course of decades. Use historical word representations from two periods to identify those whose meaning has shifted significantly.

Convolutional filters

Design a filter that repairs distorted photos, then combine image enlargement with convolution. Check how both steps affect the reconstruction error.

Pixels

How much can be recognized from just ten visible pixels? Develop a method for classifying digits and explore how to use a model trained on complete images.

Colorization with a GAN

How can you restore color to black-and-white portraits? Use a face generator and propose a colorization method that preserves image details and produces plausible skin, h...

Token predictor

Predict which tokens will appear after a fragment of Shakespeare’s text. Use the available language model and data to distinguish genuine continuations from misleading ca...

The Void

Discover the five pairs of information that the mysterious model responds to. Use the ship reports, recorded activations, and the final layer's weight to determine the ru...

Multilabel classification

Several items of clothing may appear in one image. Train a model that recognizes all the classes present and handles their different combinations.

Hidden categories

Read product categories from the vectors used by the recommendation system. Teach a model to assign consistent sets of labels where the previous employee left no instruct...

Decision localization

Does the classifier look at the object or at random background? Prepare saliency maps and compare them with the actual object areas in the images.

Whisper or scream?

Your voice reveals more than the word itself. Teach a model to identify normal speech, whispers, and screams in short recordings of Polish words with equalized volume.

Multispectral data segmentation

From the twelve bands in the satellite image, choose those that help identify the terrain. Train a model to assign each pixel to one of four classes and check the quality...