Tutorials Day
(separate registration, online format)

Replika is the world's largest English speaking chatbot. Let's discuss how one of our key components works, which allows us to choose the most relevant answer from different dialogue models, including GPT-3, significantly improving the quality of the dialogue. Let's go through the entire development path of this component, talk about successful and unsuccessful experiments.

A tutorial on GANs. Detection of changes in the operation mode of the observed object is a well-known task in various contexts: quality control, detection of failures in complex systems, predictive maintenance, patient monitoring in healthcare, voice assistants and video analysis. Various methods of jitter points detection are used in these contexts. In this tutorial, we present two neural network-based approaches for jitter points online detection. These algorithms demonstrate the linear computational complexity of and are suitable for detecting jitter points in long multidimensional time series. We will provide a visual description of these algorithms and their comparison with state-of-the-art methods on various sets of synthetic and real data. In particular, we will consider the task of detecting changes in human physical activity using data from smartwatch sensors and test the algorithms on data from astronomical telescopes and particle detectors in high energy physics.

Four variables make a machine learning model selected: the model parameters and their distribution, the model structure and its distribution. Changes in data implies changes in the model structure. For a complex data set one has to deal with an ensemble of selected models. An ensemble is supposed to collect different models, which fit the data holistically. This talk will discuss principles of model selection for a single model and for an ensemble.

Planning actions according to a known model is an important classical approach in artificial intelligence. By the integration with reinforcement learning methods, it is possible to use an updatable model and achieve truly meaningful results. Starting with simple algorithms, in this tutorial we will understand how AlphaZero works, and then take a look at modern SOTA algorithms from the past two years.

Building good machine learning models often requires large amounts of data. There is a huge amount of data in the world. However, this data is usually owned by someone and may be subject to personal data laws. For example, medical records, financial transactions, photos of people are all examples of data that can be subject to various restrictions. Also, sometimes data takes up a huge amount of space, so it is impossible to collect it in a centralized storage for further training. To develop the AI industry, we want to use maximum of the available data. An approach called Federated Learning helps to solve the lited problems. This approach allows to train models on distributed data without sending the data itself or violating its privacy. Instead of sending data to models, in FL we send models to data. This allows organizations to collaborate on machine learning projects without sharing sensitive data as well as solving the problem of moving across a big data network. As part of the tutorial, we will get acquainted with the Python open-source library for federated learning - OpenFL (https://github.com/intel/openfl). Let's build our federated network for training on distributed data and look at a few examples.

This workshop will provide an overview of Microsoft's cloud machine learning tool, Azure Machine Learning, from running Jupyter Notebooks to planning distributed learning experiments and hyperparameter optimization on a cluster. As an example, we will consider the generation of pictures by generative adversarial networks and the construction of a model for answering questions on a dataset of scientific papers about COVID-19.
For this tutorial you will need a Microsoft Azure subscription, you can get a trial subscription at http://aka.ms/azfree, for students - http://aka.ms/az4stud.

Deploy and launch a docker container on the edge Atlas500 device using the Huawei FusionDirector system. Configuring and running a container with facerecognition demo.