NLMK, “The prediction of rolls wear in the hot strip mill”

Room 425

Full description on Kaggle

Prerequisites: Python, Anaconda (sklearn), Excel
Abstract. The metal processing in the hot strip mill is divided into the series of so-called rolling batches. After finishing each batch, the mill stops and the worn work rolls are replaced with the new ones. The removed rolls are then polished on the grinding machine in order to remove the surface defects and restore their original geometrical profile. After that, the rolls are installed back into the hot strip mill during its next stop.
The work rolls are utilized until their work surface, which is typically several centimeters thick, fully wears out. The wear is defined as the reduction of the roll diameter during its operation and the subsequent polishing. The wear depends on multiple factors such as the roll’s material, the roll’s position in the mill and the processed steel grade. The task is to create the mathematical model of the work roll wear by using the provided statistical data, as well as some physical reasoning. The model should be then used to predict the roll wear after processing each rolling batch.

CSD Lab, MIPT, “Computer vision and reinforcement learning”

Room 424

Registration, Presentation, GitHub

Abstract. Creation of an improved program for the accident-free movement of a car at an unregulated crossroad using video from an unmanned aerial vehicle. This will require the usage of deep learning approaches to recognize cars on the images and reinforcement training to plan their movement.

Huawei, “Metric learning for facial descriptors”

Registration, Presentation, Kaggle, GitHub, Telegram

The problem: Construct and implement the metric (similarity function) which brings us the better percent of recognition (identification\verification) of human faces on photos in comparison to L2 baseline.

Prerequisites: Python 3.

Abstract: State-of-the-art recognition (identification) methods generally use the following paradigm: firstly vector (descriptor) of some object (e.g. photo) is calculated using deep neural net, and then using some simple metric (e.g. L2 in Euclidian space) we select the closest one(-s) descriptor from the preliminary computed set of descriptors of objects using the same neural net, among which we’ll search the appropriate pair request<->DB. For verification systems the metric is applied between the descriptor of object-request and the preliminary computed for this object descriptor-template, and a result conclusion on whether this is the same object or not is done based on comparison of this metric with some predefined threshold.