Archives: Speakers

Post Type Description

Dimitri Nowicki

Topic: State representation learning and associative memory for intelligent agents

Language: Russian

Abstract: Most real-world world tasks are hopelessly complex from the point of view of reinforcement learning mechanisms. In this talk we will give an overview of methods for finding efficient state representations form agent’s observations. We will Bayesian inference with infinite capacity models, and explore links representation learning and the machine learning problem of transfer learning. Then we will present models of neural associative memory and their interplay with machine learning.

About: PhD. in Computer Science and Applied Mathematics. Expert in Artificial Intelligence, Neural Networks, Computational Neuroscience, Machine Learning and optimization. Experience as a Senior Research Fellow in US (University of Massachusetts)

Oleh Yudin

Topic: Predicting winner of a horse race

Language: Russian

We’ll discuss the next points:
1. Basic information about horse racing;
2. Betting strategy. Kelly criterion;
3. ‘Raw features’. Feature engineering from raw features;
4. Features and target variable. Problem statement;
5. Mathematical predictive model. Error function;
6. Methods of solving. Linear log regression, XGBoost, CatBoost;
7. Results (accuracy);
8. Ways to improve (Boosting, FM).

About: Data Scientist & Machine Learning engineer, PhD applied mathematics MIPT. Data scientist in EnkeSystem (video classification), Former Data scientist in ThoroughtBet (Horse race prediction) worked with supervise & unsupervise models, feature engineering, classification , regression and ranking tasks.

 

Aleksandr Dolgaryev

Topic: Multiple Object Tracking using Person Re-identification for Video Analytics Platform

Language: Russian

Abstract: There is going to be Video Analytics Platform based on person tracking by re-identification presented. The first part of the speech will be devoted to the overview of Multiple Objects Tracking (MOT) problem and typical challenges. Next, we will describe the framework: building blocks of the system and the main steps of the system development and problem resolving. Also, will demonstrate the business application and real-world examples: person detection, football analysis (time analysis, location analysis, direction analysis).

About:  Aleksandr is a CTO in Quantum, has 17 years of experience, held the next appointments: developer, Tech Lead, Team Lead, Architect, PM, CTO. He has a strong general computer science background, as well as plenty of experience with many technologies. Aleksandr can make architecture and propose the optimal technology for distributed and complicated solutions.

 

Kostiantyn Isaienkov

Workshop topic: Dimension reduction for image clustering using convolutional autoencoder

Language: Russian

Abstract: On the workshop we will describe the solution for the task of the dimension reduction of the image dataset for further clustering using convolutional autoencoder. The program includes the following steps:
1. Building and training autoencoder model (python3, keras).
2. Dimension reduction of real image dataset via autoencoder model.
3. Clustering of image dataset.
4. Comparing precision of convolutional neural network trained on original dataset with another feed-forward neural network trained on compressed dataset.
5. Image dataset clustering (python3, kmeans).
Most clustering algorithms require the linear dimension of the dataset. In the case of images, we are dealing with a 3-dimension matrix (height, width, 3 rgb values). The convolutional autoconcoder allows to reduce the image dataset to 1-ddimensional form. Correct approach for selection of the size of the bottleneck (the dimension of the compressed data) allows to decrease loss of precision.

About: Kostiantyn Isaienkov is a Data Science & Machine Learning Engineer, who works in Akvelon and has 3 years of commercial experience. Kostiantyn is interested in NLP, computer vision and time series forecasting. He has several publications in the field of applied problems of artificial intelligence. Kostiantyn is a kernels expert on Kaggle platform.

Michael Yushchuk

Workshop topic: Simple object detection with Tensorflow

Language: Russian

Abstract: On the workshop we will solve the task of detecting the gesture “heart” on the photo using Tensorflow object detection API. We’ll recover methods that were applied to a similar task in the past. Also, we’ll get a basic understanding of neural networks and how they are applied to the object detection, using one of our projects as the example. Finally, we’ll go through all the steps, starting with the search and markup of the dataset, finishing with the preparation of the model for the application.

About: Michael Yushchuk is a Machine Learning Engineer, who works in Quantum and has 2 years of commercial experience. He has a strong general computer science background, as well as plenty of experience with Python, SQL storages and other server side-related technologies; has worked on Machine Learning, Computer Vision tasks, using scikit-learn, matplotlib, scipy, pandas and other.

Tolga Birdal

Topic:  Sparse and Dense Reconstruction via Detection: CAD Priors for Accurate and Flexible Instance Reconstruction for Industry 4.0

Language: English

Abstract: The first part of this talk will address a coordiante measurement application where manufactured parts are 3D-quality checked in a realtime and online manner. Next, I will propose reconstruction-by-detection framework that is a different and accurate approach to 3D reconstruction. In contrast to conventional scanning, our pipeline uses the nominal CAD model prior as a proxy. It begins with the rough matching of the model to the acquired unordered and unstructured scans, obtaining good initial registration to the common coordinate system. Unlike other conventional models that concentrate on multi-view registration of object scans and require either pre- or post-segmentation of the reconstructed object, our reconstruction by detection allows automatic segmentation of the object of interest from the scans containing clutter and occlusions, thus avoiding any manual interventions. Moreover we introduce automatic computation and update of a global pose graph with live user feedback about the object coverage leading to potential seamless interaction during the data acquisition. Finally we perform global alignment, merging and meshing of the segmented object scans. Our method depends neither on the scan resolution nor the reconstruction volume, allowing reconstruction of objects from small (several cm3) to extremely large sizes (2 to 125m^3) retaining the sub-millimeter scan resolutions.

About: Tolga Birdal is a PhD candidate at the Computer Vision Group at the Chair for Computer Aided Medical Procedures, TUM and a Doktorand at Siemens AG. He has achieved to make many of those memorable with several awards in competitions or conferences.
As being an enthusiast in his specialties in computer vision, pattern recognition and optimization, he has a deep and compelling background, knowledge and creativity as a product of years of effort.
While often developing his own tools, he has a very wide knowledge-spectrum of open source and commercial libraries that he made himself familiar with.

Welf Wustlich

Topic: Cognitive Supercomputing for images, text and speech

Language: English

Abstract: Planet AI is going to present its latest state of the art technology – PlanetBrain.
Following to a general technological overview we will demonstrate how PlanetBrain works in powerful real world applications in various domains as for example: traffic surveillance, automatic address recognition and parcel sortation, document analysis, speech analysis and visual object detection.
Our presentation will end up with a short visionary outlook about what is coming up soon.
We hope you will enjoy and get inspired for your own AI projects!

About: Since years Mr. Wustlich is managing the department of R&D of Planet GMBH (www.planet-ai.de). Besides organizing many international research projects he designs Planets R&D strategy and is responsible for the development of new technologies guiding internal projects and teams. So he was coordinating European research projects with leading European research partners (e.g. FP7-Organic) until 2012 and also national co-operations for example with CITlab from university Rostock (citlab.uni-rostock.de). Besides participating on international scientific research (e.g. Neural Computation Dec 2012), Mr Wustlichs main focus is, integrating newest state-of-the-art technologies into Planets recent product development. Since 2015 Mr. Wustlich is leading Planet AI GmbH in Rostock a R&D dedicated spin-off running a team of selected experts to develop latest state of the art technologies in the area of AI.