Recursive generative query network application to real data - a neural architecture based on a generative query network adjusted to deal with time series briefly described here https://docs.mlinpl.org/virtual-event/2020/posters/38-Application_of_Generative_Query_Networks_for_industrial_time_series.pdf It was verified on synthetic data and some preliminary test were performed on real data but to verify it the approach works it shall be thoroughly tested using real-world data.

Time series invariants - an equivalent of Hu moment invariants but for time series instead of images. Find a way to transform a time series to a vector that will be the same or very similar if the series got shifted, scaled, stretched, or shrunk. It will allow easier time series clustering or using them in classical ML approaches with constant input shape.

Improved ELO score - ELO ranking is used in many online and offline games to compare players. There are a few drawbacks of this approach like a need to manually set parameters, not being stable, or finally not fitting the real distribution perfectly. With modern hardware and algorithms, an improved version can be created.

Aircraft deplane strategy optimization - when boarding a plane people get a boarding group assigned. It used to aim to minimize the time needed to get everyone inside, now it's used to maximize gain by charging more prioritized groups. However, deplaning is mostly not organized right now and often ends in the front rows first strategy which is the worst one in terms of time. Can you find a better strategy taking into account multiple factors like who has a piece of additional baggage, who booked and sat together, a fraction of people not following the rules, some people moving slower, etc?

Technology prediction - based on a 3D CAD model of an item to be produced predict steps and times needed like 5 minutes of welding, 3 minutes of bending, and 7 minutes of painting.

Genetic neural network optimization - using genetic procedure to find the best neural network for a given problem. Both, the architecture and the weights are to be determined during the genetic optimization.

Comparison of RL algorithms on the DSJ game - CNN based vs manual features extraction, different ways of training the model

Automatic feature engineering - most models can't perform some operations e.g. multiplication, thus neither neural network nor xgboost can e.g. predict BMI based on weight and height. The aim of the project would be to automatically create derived features improving the quality of the prediction.

Find a song based on someone humming it. Information retrieval system with sounds instead of words used as tokens to find relevant songs (documents) based on humming (query)