OpenTalks.AI /
6-7 March 2023
Yerevan, Armenia

OpenTalks.AI 2023

version from 02.03.2023
Yerevan time, GMT+4
Welcome drinks and networking
The evening before the conference is a great time to drink a glass of wine and meet familiar faces in an informal setting!) And of course, to meet new people!)

Welcome drinks will start at 19:00 at IBIS Yerevan Central Hotel, 2nd floor. You will need your badge and need to be 18+ to enter if you want to join, so, please, do not forget your id document)

Specially for this evening, we invited winemaker Samvel Machanyan from Alluria winery to introduce his wines to our participants. A wine of this winery was selected by Igor Pivovarov and Elena Chinarina to be the speakers gift at OpenTalks.AI-2023! We will have several bottles of this wine to taste on this event)

Computer Vision & Reinforcement Learning Day

Monday, March 6
09:00 – 10:00
Registration & coffee
10:00 – 10:05
Opening remarks
Igor Pivovarov, OpenTalks.AI
Conference spotlights, main ideas, figures
10:05 – 10:10
Opening remarks
Habet Madoyan, AUA
Welcome speech from the DS AUA Program Chair
10:10 – 11:25
Plenary 1 - reviews
Big Conference Hall
10.10 – 10.40
AI Technologies for Digital Characters and Avatars
Dmitry Korobchenko, NVIDIA
Modern advanced digital characters and avatars are powered by AI technologies in multiple different ways and on all stages of their creation and usage. From avatar appearance synthesis and reconstruction using generative AI and computer vision, to their rendering, emotional facial animation and body animation, speech synthesis and speech understanding by conversational interactive avatars, and complex environment-aware character animation with interaction with objects. In this talk I will provide a comprehensive overview of the field and related task, core AI methods to solve them, and corresponding target applications, including NVIDIA products.
10.40 – 11.25
Computer vision - main in 2022
Alexey Dosovitskiy, Google Brain
The talk will provide a quick overview of some trends and results in computer vision in 2022. Covered topics will include: multimodal learning (images and text, video and text/audio, etc), self-supervised learning (masked modeling etc), approaches to fine-grained tasks (detection, segmentation) - in particular open-vocabulary, universal vision models, scaling of vision models, 3D modeling. Advances in generative modeling will only be mentioned briefly, since they will be discussed in more depth in other talks.
11:25 – 11:45
Coffee break
11:45 – 12:45
Computer Vision in healthcare
Small hall
Manoogian hall
Big Conference Hall
Boris Zingerman,
AI in legal practice
Datasets, markup and testing
Daria Suslova,
Exploring for new AI-application in healthcare
The ML model through the eyes of a lawyer: legal nature, protection of rights, responsibility
Elena Melnikova,
Autonomous director and autonomous BoD systems based on AI and ML

Anna Romanova,
AI-based claims analysis system
Ignat Postny,
LLC "TAG Consulting"
Session partner
Victoria Dochkina, Gazprombank
Testing machine learning systems
Lavrenty Grigoryan, Gazprombank
It is planned to discuss computer vision methods for working with medical research on images of different projections of the same area of interest. The talk will touch upon the following modalities: mammography, chest X-ray, etc.
There will be discussion of geometric methods of matching finds on different projections and neural network architectures that allow consideration of information from multiple projections at the same time.

Roman Kucev, TrainingData.Pro
How to get high quality labeled data
The issue of core legal function automation though AI is a daunting, but, nevertheless, a feasible task. In his presentation Ignat Postny will talk about practical experience of developing a system for automatic analysis of legal documents (court documents), which is capable of:

- analyzing a set of incoming documents (court documents);

- drafting a report on all the shortcomings of the received set of documents;

- drafting, at the user's request, a set of necessary response documents.

Sometimes creating an ML-model requires significant resources, such a model can turn out to be unique. The ML developer's rights must be protected somehow, and there are legal remedies for this. In addition, during the working application of the ML-model, harm may be caused, therefore it is necessary to create a method for finding the tortfeasor when such harm is caused by artificial intelligence applications.

My talk will focus on quality control methods and how to build a data labelling pipeline within the company. We will discuss the main mistakes in the organization of the labeling process and will find out how to avoid them.
• Differences between Data-Centric and Model-Centric approaches
• Iterative approach to data labeling: pros and cons
• Building an effective learning process for data assessors
• Quality control methods
• Basic errors in data labelling management

We present AI service for synthetic data generation - SyntData. The service provides relevant, valid synthetic data, generated with deep insights from real data. Synthetic data guarantees safety of clients data, and providing DEVs, QA, DS opportunity to work with data similar to real data.
Polina Postnikova, Research Institute of Rheumatology. V.A. Nasonova
Development of a No-code platform for creating search microservices
Maxim Puchin,
Доклад содержит в себе обзор результатов разработки и внедрения no-code платформы для создания поисковых микросервисов в компании ГАРАНТ.
Learning annotator's style in medical imaging
Evgeny Nikitin,
The World Economic Forum (WEF) 2015 report "Technology Tipping Points and Societal Impact" predicts that by 2026 the first artificial intelligence system will take a seat on the corporate board of directors. The first official announcement about the artificial intelligence system in the board of directors was published in 2014 Position of a corporate director is one of the few that are required for execution by a "natural" person only. The main prerequisites for full automation of management decisions made at the level of the board of directors are formed in the field of corporate law, machine learning, and rules of non-discrimination, transparency, and accountability of decisions made and algorithms applied.
Radiologists disagree when they annotate medical images. There are many reasons why this could happen - human error, different skill level, bad instructions. Some of these factors can be mitigated and accounted for, but sometimes doctors just have different opinions. In my talk, I want to tell you about different ways of working with intra-observer variability, and to propose a novel method of learning annotator style.

Armen Manasyan,
Generating synthetic data and training models on them.
Fattakhova Yulduz,
Evgeny Sidorov,
Third Opinion AI
Multi-view pathology detection
Fedor Zhdanov,
"Test test test" - comprehensive testing of ML models is one of the mandatory steps according to the "Responsible AI" concept. We will discuss different machine learning tests that go far beyond evaluating the models' performance on the data subset, such as: pre-train tests, post-train tests (incl. behavioral testing, metamorphic testing, etc.) and data drift detection tests.
We will also introduce the NLP models' testing pipeline that is used in Gazprombank and talk about the ML testing tools (including a list of libraries).
Only 21 AI-based Medical devices are registered in Russia, and more than half of them are CV in radiology. Nevertheless there is a vast variety of other types of studies and clinical tasks in healthcare, so what are the main drivers and limitations of AI automation in healthcare? Discussion is based upon 13 research projects in various fields of healthcare (cardiology, colonoscopy, obstetrics, endocrinology) and types of studies (ultrasound, MRI, slide microscopy, ECG, endoscopy) for the leading Medical Research Centers.
Elizaveta Dakhova,
Automated evaluation of hand radiographs in patients with rheumatoid arthritis
Radiographic progression in rheumatoid arthritis gives an objective measure of anatomical damage that defines the course of the disease and the effects of treatment. In studies and clinical trials to assess radiological progression used a very time consuming method. We are developing a deep convolutional neural network (CNN) model to automatically evaluate hand radiographs in RA patients. Moreover, we present a prototype of a web application that can be used by radiologists to accelerate the formation of a study protocol.
Sergey Morozov,
Going global: how to deploy AI at EU27 hospitals.
How we deployed 22 algorithms from 12 vendors into EU27 hospitals
12:45 – 13:00
Coffee break
13:00 – 14:00
Generative models in business
Small hall
Manoogian hall
Big Conference Hall
Sergey Lukashkin,
Natural Language Processing - research & development
Robots and drones - research & development
Andrey Kuznetsov,
Creative AI models design. New trends and applications.
Topology meets BERTology: Topological Data Analysis for the understanding of Transformers

Irina Piontkovskaya, Huawei Noah's Ark Lab
The report presents our experience gained during the development of the apple picking robot. Particular attention is paid to the computer vision system for detecting apples. We will also talk about the positioning system relative to the camera and the robotic arm. This compares several stereo cameras, such as the Intel Real Sense Depth Camera D415/D455 and ZED2. What is the error in estimating the coordinates, why is the Internet of Things here and how did you manage to achieve recall at the level of 95%. It will be about problems, and about difficulties, as well as about the joy of the first picked apple.
Alexey Postnikov,
Sber Robotics Laboratory
What can large sequential models bring to robotics?
This talk will explore the ways in which generative artificial intelligence (AI) is being used to augment and enhance the creative process in a variety of industries. The talk will cover the basics of generative AI, including some history, key concepts, and current state of the art. We will discuss specific applications of generative AI in fields such as music, film, and video games. I'll share some nuances of adapting conventional ML lifecycle to fit the requirements of creative industries, and how we overcame them at Deepcake. Overall, I'll try to provide a comprehensive understanding of the role of generative AI in the creative industries and its potential to shape the future of creativity and innovation from perspective of AI startup in the field.
The development of self-driving cars has been a major focus in the field of artificial intelligence. To achieve this goal, large amounts of data are required for training machine learning algorithms. However, collecting and labeling real-world data can be time-consuming and expensive. To overcome these challenges, this paper proposes using synthetic data for learning self-driving cars, including the ability to generate unlimited amounts of diverse and controllable data. We developed a solution for efficient and stable integration of RLLib with Carla simulator. We present end2end solution for learning self-driving cars in Carla simulation environment with GYM-interface. The results demonstrate the effectiveness of using synthetic data in training RL-agents for autonomous vehicles. The findings suggest that synthetic data has the potential to significantly accelerate the deployment of self-driving cars by providing a cost-effective and scalable solution for training machine learning models.
This presentation provides an overview of the current state of robotics and the latest developments in the application of large sequential models (such as GPT-3) to the field. The focus is on how these models can enhance the capabilities of robots and enable them to perform a wider range of tasks and interact with humans in new ways. The talk covers the latest trends in the field, including new models, such as SayCan, that are designed to enable more natural human-robot interaction, as well as the potential benefits and challenges of using large language models in robotics. The presentation concludes by exploring some of the future directions and opportunities in this rapidly evolving field.
Расскажу про реализацию системы восприятия на основе лидаров и камер в нашем беспилотном грузовике. Расскажу, как мы преодолели ограничения промышленного вычислителя для эксплуатации на объектах заказчиков.
Практическое применение генеративных нейронных сетей в практике работы компаний должно получать конкретные прикладные реализации. В своем докладе мы показываем на примере работы крупного Digital агентства, каким образом современные генеративные нейронные сети, будучи дообученными на исторических, маркетинговых, аналитических и финансовых данных компании, могут стать нативным инструментарием для самых различных ролей внутри компании, будучи интегрированным во внутреннюю ERP систему. Покажем реальный опыт внедрения и постараемся оценить результат и оказанный эффект на бизнес компании, порассуждаем о развитии инструментария.
Svetlana Korobkova, Docet TI
Image generation for social media content
"Capture and share the world's wonderful moments" is the slogan of Instagram, this states that images are the dominant point of communication in contemporary social media.
We present a technology for image generation for social media, which can help bloggers who have to produce huge amount of visual content daily to maintain high level of engagement rate for a blog.
Modern out of the box image generation technologies are mostly based on simple textual (and/or visual) "prompt", that is not able to take in consideration a lot of details, which determine blog style.
Our approach allows performing automatic detailed analysis of blog content and use all extracted details as a complex prompt to produce new content which is semantically close to the original and vary the proximity of the original and generated visual blog content style.
Anastasia Semenova, CleverData
Disassembly and Modification of TiSASRec
The process of creating a script for a voice robot operator involves a number of routine operations performed by trained specialists. Our experience in creating such scripts allow to confirm that almost the entire path of creating a robot script can be automated to the magic button "Create script", which will allow programming the robot without special knowledge to solve communication problems over the phone. Let's talk about experiments with AI generator to automate the creation of a script based on real dialogues of live operators with subscribers.
Maria Tikhonova, SberDevices, HSE
Overview of Controllable Text Style Transfer
Text Style Transfer is an important task in NLP, which aims to control certain attributes in the generated text, and to generate or paraphrase text in a specific style. This talk concentrates on a specific style transfer approach known as controllable text style transfer, where one aims to generate a text in a specific style by controlling the generation of a language model so that the generated text is written in a desired style. The presentation gives the broad overview of the controllable text style transfer methods, covering such approaches as CTLR, GeDi, ParaGeDI, FUDGE, DExperts, and CIAF, highlighting possible ways of the developing of this area of research.
Autonomous Truck Perception System for Closed Areas
Alexey Voropaev,
Andrey Kuzminykh,
Synthetic data: Learning self-driving cars in simulation

Roman Doronin, EORA
Computer vision for an agrobot-manipulator for picking apples
Nikita Andriyanov,
Fin. University
Alexander Notchenko, Deepcake
Generative AI for Creative Industries

Alexander Platonov,
Vladimir Novoselov, Realweb
Development and practice of using tools based on generative neural networks in the work of a Digital agency
Anastasia Myshkina, Realweb
We apply topological data analysis (TDA) to speech classification problems and to the introspection of a pretrained transformer models, namely, BERT and RoBERTa in NLP area, and HuBERT for Speech data. Our results demonstrate that TDA is a promising new approach for speech and language analysis, especially for tasks that require structural prediction. We also show that topological features are able to reveal functional roles of Transformer heads; e.g., we find the heads capable to distinguish between pairs of sample sources (natural/synthetic) or voices without any downstream fine-tuning.
The talk will cover one of the main topics in the international AI community - Creative Artificial Intelligence. First, I will speak about the task itself and its history, how we started with classic CV tasks and proceeded to text2image models. Further I will describe the main trends in multimedia data synthesis in 2022-2023 and observe current SoTA architectures, giving a brief description of our diffusion-based text2image model Kandinsky 2.0. After that we will speak about different applications of Creative AI today and in the nearest future in terms of my vision. And finally I will show how we proceed in Creative AI for high fidelity face swap on images and video, describe our current SoTA solution - the GHOST model, and show our marketing applications in movie production, advertising, etc.
14:00 – 15:00
15:00 – 16:30
Plenary 2 - reviews
Big Conference Hall
Pitch session of startups
Manoogian hall
Big Conference Hall
Reinforcement Learning - main in 2022
Alexander Panov, MIPT, AIRI, FRC CSC RAS
Andrey Voynov, Google
In my talk I'm planning to discuss the recent advances in text-to-image generative models development. We will be primarily focused on the diffusion models, and dive deep into how they work, what controls do they have, and how they can be applied to a variety of tasks.
The talk will go through the most exciting results in the field of reinforcement learning (RL) obtained in 2022. We will see how this area of research has changed with the use of autonomous learning, transformers, and an environment model. We will also touch on using RL as an auxiliary tool for other tasks in the field of ML, for example, for additional training of large language models.
Generative and diffusion models — main in 2022
15:00 – 15:45
15:45 – 16:30
16:30 – 16:45
Coffee break
16:45 – 17:45
Reinforcement Learning in business
Small hall
Manoogian hall
Big Conference Hall
Fedor Tsarev, WorldQuant
Machine learning - academic talks
Computer Vision - research & development
Oleg Svidchenko, "AI in Industry"
Reinforcement Learning in Real Life: applications, cases and challenges
Neural networks for the problem of finding anomalies in time series in industry
Iurii Katser,, Skoltech
Fast Simulation of a Data Storage System based on Generative Models

Mikhail Hushchyn,
Neural networks penetrate deeper into business and hold more of the information space. But despite the fact that about ten years have passed since neural networks boom in computer vision, there are still not so many products that are really able to make money with such technologies. Much more often the topic of neural networks sounds for PR or attracting investments. The main problem is the high cost of development and the cost of hardware. I will tell you about our product, which, despite the large and complex CV engine, broke in the long-established advertising market, managed to compete with classic solutions in price and be cost-effective.
Exploring principles of multi-agent simulation, collaborative decision making and experiential learning to active portfolio management relying on crypto-finance as a reference.
FusionBrain — это исследовательский проект, основными задачами которого являются разработка эффективных мультизадачных и мультимодальных моделей и применение их для решения широкого круга практических задач. Общая цель и идея проекта — научиться создавать модели, которые смогут как можно более эффективно извлекать дополнительные важные знания из большого количества модальностей и задач при обучении и за счет этого лучше решать другие задачи. Исследования проводятся во многих модальностях: тексты, изображения, аудио, видео, языки программирования, графы (например, молекулярные структуры), временные ряды. Список решаемых задач очень большой: от классических задач CV и NLP до задач, вовлекающих разные модальности: VideoQA, Visual Commonsense Reasoning, IQ tests (эти задачи сложны даже для человека). Изучается также способность моделей решать задачи, сформулированные на естественном языке (в частности, в формате инструктивной генерации с применением методов RLHF), и даже справляться со скрытыми задачами (для которых в обучающей выборке отсутствовали примеры). Исследования сосредоточены в том числе на сокращении данных, человеческих и вычислительных ресурсов, необходимых для обучения и инференса различных моделей. В рамках доклада будут рассказаны результаты исследований: в частности, речь пойдет о некоторых разработанных архитектурах, таких как ruCLIP, ruDALL-E (Kandinsky), Kandinsky 2.0, RUDOLPH, а также о проведенных соревнованиях, таких как FusionBrain Challenge и FusionBrain Challenge 2.0, и о разработке мультимодального бенчмарка
Interpretable Anomaly Detection Models in Cyber-Physical Systems
Yuri Chernyshov,
Описан метод интерпретируемого обнаружения аномалий с использованием сетей глубокого обучения автокодировщиков, RBM. Рассматривается вопрос обеспечения интерпретируемости показаний модели с использованием анализа значений нейронов скрытого слоя автокодировщика. Приводятся результаты применения модели на синтетических данных и на открытых датасетах.
Case-driven CV in satellite image processing
Alexey Trekin,
Geoalert LLC
Dmitry Anzhiganov,
MSU, Research Institute of Nuclear Physics
Hunting for ultraviolet transients with a neural network

High-precision modeling of systems is one of the main areas of industrial data analysis today. Models of the systems, their digital twins, are used to predict their behavior under various conditions. We have developed a model of a data storage system using generative models of machine learning. The system consists of several types of components: HDD and SSD disks, disk pools with different RAID schemas, and cache. We represent each component by a probabilistic model that describes the probability distribution of component performance values depending on their configuration and external data load parameters. Machine learning helps to get a highly accurate digital twin of a particular system, spending less time and resources than other analogues. It quickly predicts the performance of the system, which significantly speeds up the development of new data storage systems. Also, comparing the forecasts of the model with the real performance helps to diagnose failures and anomalies in the system, increasing its reliability.
Edward Pogossian, Institute for Informatics and Automation Problems of NAS RA
Computer Vision and Artificial Intelligence In Advertising
Maksim Kuprashevich, SberDevices
Anton Kolonin,
Adaptive Multi-Agent Active Portfolio Management
Anton Ganichev,
Nowadays, the relevance of information security of industrial automation systems is no longer in doubt. In our report, we will describe the developed method of detecting anomalies in such systems based on the analysis of a copy of network traffic. This approach is compatible with any industrial automation systems and doesn`t require information about the topology, network protocols and algorithms. An APRE algorithm is proposed, which allows to extract packet headers of unknown network protocols and determine their semantics without a priori knowledge of the protocol structure, based on changes in entropy and mutual information of packet bytes. A multi-agent modeling approach is used to detect anomalies in the automation system operation. For each component of the system, the creation of an agent capable to predict response on input signals extracted from a copy of the network traffic in the previous step is performed. Several ways of representing and training agents in the form of different types of automata are proposed.
Autonomous multi-agent system for detecting attacks on industrial networks: analysis of unknown protocols and search for anomalies
Denis Komarov, CyberLympha
Alexey Sinadsky, CyberLympha
Data-driven methods showed significant results in solving different tasks in many industrial applications. There are recent works that show NNs achieving state-of-the-art results in anomaly detection problems overperforming traditional algorithms and methods. In my speech, I will review some works and NN architectures related to the anomaly detection problem in industrial time-series data.
Despite being proven in the image processing field, neural networks still are a tricky tool for cartography. Can we trust the results? What should we do with the errors? Should we rely on selling ready models as a service, or stick to on-demand development?
In this talk I will share some practical cases: how do we derive the model for the particular task and area from the general off-the-shelf model, how do we collaborate with human cartographers and how to handle user's feedback.
Reinforcement Learning is a field of Machine Learning that solves interactive tasks via continuously learning an agent to pick actions that should be optimal in the long-term interaction horizon. This is a broad problem setup and many different tasks (both theoretical and practical) can be formulated in terms of RL. However, Reinforcement Learning algorithms (especially out-of-box solutions) often lose in terms of efficiency to the specialized ML or optimization methods. Nevertheless, there are also plenty of successful cases of applying RL to a variety of tasks. This talk is devoted to Reinforcement Learning and its applications in real-life tasks. We will briefly talk about common reinforcement learning approaches. Then, we will discuss some successful cases of RL applications both for real-life and digital twins tasks and also the challenges of developing robust solutions with RL.
Начиная с 2019 г. на Международной космической станции работает российско-итальянский эксперимент "УФ атмосфера" (Mini-EUSO), основным инструментом которого является широкоугольный телескоп, направленный в надир. Главной целью эксперимента является получение карты излучения ночной атмосферы Земли в ультрафиолетовом (УФ) диапазоне, что является необходимым элементом подготовки крупномасштабного эксперимента по изучению космических лучей предельно высоких энергий с помощью орбитального телескопа. Как и более ранний эксперимент ТУС, прибор "УФ атмосфера" регистрирует сигналы разнообразных процессов, происходящих в атмосфере в УФ диапазоне, и среди них - свечение метеоров. Мы описываем две простые нейронные сети, которые позволяют эффективно выделять сигналы метеоров в общем потоке данных. Реализованный подход может быть применён для поиска трекоподобных сигналов различной природы в данных флуоресцентных и черенковских телескопов.
FusionBrain: research project on multimodal and multitasking learning
Denis Dimitrov,
17:45 – 18:00
Coffee break
18:00 – 19:00
Computer Vision in business
Small hall
Manoogian hall
Big Conference Hall
Alexey Sidoryuk,
ANO "Digital Economy"
Machine learning - academic talks
Dubai, Almaty, Yerevan, Tbilisi, London, Singapore - Russian experience.
Hebbian learning for Convolutional Neural Networks: Overview
Alexander Demidovsky, HSE
Continuous Deep Q-Learning in Optimal Control Problems: Normalized Advantage Functions Analysis
Anton Plaksin,
Yandex, IMM UB RAS
Kartashev Oleg, Severstal Digital
Metric Learning, Anomaly Detection and Synthetic Data for preventing chain conveyor outages
Dmitry Pshichenko,
Mining industry cases on the application of machine learning and computer vision: a business perspective
The implementation of artificial intelligence (AI) and machine learning in the mining industry provide many economic benefits for the mining industry through cost reduction, efficiency, and improving productivity, reducing exposure of workers to hazardous conditions, continuous production, and improved safety. However, the implementation of these technologies has faced economic, financial, technological, workforce, and social challenges. This report discusses the current status of AI, machine learning implementation in the mining industry and highlights potential areas of future application. The report also presents some cases of implementation these technologies and what are some of the steps needed for successful implementation of these technologies in this sector.
Chain coneveyor monitoring is one of technically complex CV tasks solved at Severstal. We will describe what challenges did we have, how we dealt with lack of data, what ML pipeline did we create and how it is deployed and works on 39 cameras throughout 3 factory shops.

At the session, we will share our experience of moving to different countries - the cost of living in the country, working conditions, visas, the market, what kind of support you can get, etc. Real case studies first hand
Roman Doronin, Dubai
Victor Lempitsky, Yerevan
Arkady Sandler, Spain, Israel
Alexey Dral, Kazakhstan
One of the most effective continuous deep reinforcement learning algorithms is normalized advantage functions (NAF). The main idea of NAF consists in the approximation of the Q-function by functions quadratic with respect to the action variable. This idea allows to apply the algorithm to continuous reinforcement learning problems, but on the other hand, it brings up the question of classes of problems in which this approximation is acceptable. The presented paper describes one such class. We consider reinforcement learning problems obtained by the time-discretization of certain optimal control problems. Based on the idea of NAF, we present a new family of quadratic functions and prove its suitable approximation properties. Taking these properties into account, we provide several ways to improve NAF. The experimental results confirm the efficiency of our improvements.

Training acceleration is one of the prominent research directions in the field of deep learning. Among other directions in this field, Hebbian learning is considered to be a highly prospective approach. Although Hebbian learning does not produce models of accuracy comparable to training with a traditional backpropagation approach, there is an emerging trend of applying Hebbian learning as a part of mixed training strategies that might include various backpropagation methods. Also, Hebbian learning is plausible for neuromorphic hardware due to its locality and highly parallel nature. In this paper, we overview existing approaches of applying Hebbian learning to training one of the largest and most demanded classes of deep neural networks - Convolutional Neural Networks. We analyze the availability of existing software solutions for Hebbian learning. More importantly, we investigate various approaches to the implementation of Hebbian learning to convolutional and linear layers as they are foundational for modern deep neural networks. This paper will be interesting both for researchers who want to accelerate training and for engineering practitioners who might be interested in exploring new ways of training Convolutional Neural Networks on new types of hardware.
Marina Kazyulina,
Continual Learning or overcoming catastrophic forgetting in neural networks

Dmitry Ivanov,
Tsifrum, MSU
Alexey Trutnev,
Artyom Tugarev,
Sergey Kuznetsov,
Разработан программный продукт для прогнозирования потребления электроэнергии на каждый час следующих суток. На основе метода машинного обучения Huber regressor разработана новая, полезная и качественная математическая модель, связывающая потребление электроэнергии с выявленными факторами. Регрессионная модель позволяет получать прогнозные оценки на каждый час следующих суток с ошибкой 3,03% на тестовой выборке данных и прогнозировать на каждый до трех суток c относительной ошибкой в 4,82%.
Software product for predicting electricity consumption for every hour of the day.
Alan Dzgoev,
Stanislav Karatsev, SKGMI (STU)
Konstantin Panfilov,
CG Samolet
Review of cases of application of machine learning models in development problems
Несмотря на высокий объем работ и высокую долю ВВП, производительность в строительной отрасли росла медленнее, чем в других сферах (в среднем, 1% ежегодно за последние 20 лет). За счет цифровизации, Самолету уже удалось увеличить производительность на 60% и впереди еще много работы в этом направлении
Neural networks trained using the backprop are prone to catastrophic forgetting. If we first teach the network to recognize cats and then start teaching it to recognize dogs, then it will forget some amount of information about cats. This problem is especially evident when new data, that needs to be learned, appears continuously during the work of the neural network. This sub-area of machine learning is called Continual Learning. There is a wide variety of approaches to this problem, ranging from the simplest ones, such as remembering all previous data, to sophisticated weight updates that reduce the forgetting of learned knowledge. We will talk about these and other methods in detail in this report.
Armen Manasyan, Armenia

Natural Language Processing & Hardware

Tuesday, March 7
09:00 – 10:00
Registration & coffee
10:00 – 10:10
Opening remarks
Igor Pivovarov, OpenTalks.AI
10:10 – 11:25
Plenary 3 - reviews
Big Conference Hall
10:10 – 10:50
Natural Language Processing - main in 2022
Mikhail Burtsev, DeepPavlov
Review of the main results in Natural Language Processing in 2022 - achievements and trends. Large Language models, etc.
10:50 – 11:25
Investments in AI - a crisis or a time of new opportunities?
Arkady Sandler
The situation in AI with investments and business in general. What markets are promising, what about revenues, rounds, where to go for startups, etc.
11:25 – 11:45
Coffee break
11:45 – 12:45
Computing resources for AI
Small hall
Manoogian hall
Big Conference Hall
Anton Mosharov, SberDevices
Diffusion models - tutorial. Part 1
Natural Language Processing in business
El-Hajj Khalil,
JSC STC "Module"
Hardware for AI: domestic hardware platform NeuroMatrix.
Introduction to diffusion models. From stochastic differential equations to star-shaped models

Dmitry Vetrov,
Session partner
Traditionally, the variety of lexical analysis and thematization tools for speech analytics cases has been limited to searching manually created lists of key words or phrases. Meanwhile, the growing volume and pace of data creation, as well as the increasing complexity of the cases being resolved by contact center, is placing greater demands on both the performance of algorithms and the sophistication of their decisions. When simple full-text search is no longer sufficient for most tasks, advanced ML and DL thematization algorithms come to the analysts' aid.

The report describes which new algorithms have appeared in modern speech analytics systems and for which tasks they can be used.

Alexander Borzunov,
Yandex, HSE
Petals: Collaborative Inference and Fine-tuning of Large Models
Many NLP tasks benefit from using large language models (LLMs) that often have more than 100 billion parameters. With the release of BLOOM-176B and OPT-175B, everyone can download pretrained models of this scale. Still, using these models requires high-end hardware unavailable to many researchers. In some cases, LLMs can be used more affordably via RAM offloading or hosted APIs. However, these techniques have innate limitations: offloading is too slow for interactive inference, while APIs are not flexible enough for research. In this work, we propose Petals − a system for inference and fine-tuning of large models collaboratively by joining the resources of multiple parties trusted to process client's data. We demonstrate that this strategy is more than 10x faster than offloading for 100B+ models, running inference of BLOOM-176B on consumer GPUs with ≈ 1 step per second. Unlike most inference APIs, Petals also natively exposes the hidden states of served models, allowing its users to train and share custom model extensions based on efficient fine-tuning methods.
Greg Tkachenko,
How AI Brings 375M Users Together Every Day
Dmitry Matov,
Barista in a coffee machine: NLP technologies in vending

Roman Doronin,
Daniel Korneev,
Re-designing DeepPavlov Dream around Large Language Models

We will provide derivation and description of diffusion models which are now one of the most promising techniques for generative modeling. Diffusion model will be considered from different angles that highlight its advantages over analogues. We will discuss several facts from the theory of stochastic differential equations that allow better understanding the logic of diffusion models and its attractive properties. In the last we will present a generalisation of diffusion model that may deal with non-gaussian noise and can be especially useful when there are additional manifold constraints on data.
Evolution of approaches to thematization in cases of speech analytics
Inna Lizunova,
Speech Technology Center
Ksenia Melnikova,
Make your AI compute 10x-1000x faster
Igor Pivovarov, OpenTalks.AI
Taxonomy of Federated Learning methods, overview of existing platforms and major players, existing challenges and industry development trends

Denis Afanasiev,
CleverData, SberDevices
Таксономия методов Federation Learning, обзор существующих платформ и основных игроков, существующих вызовов и трендов развития индустрии
A key enabling factor in the innovative AI work you see from organizations such as DeepMind, FAIR, OpenAI is powerful computing infrastructure available to their DL researchers to train large scale neural networks. We believe that this kind of computing infrastructure should be not restricted only to a few privileged companies. Rather, such infrastructure should be available to startups, researchers, universities and non-profits at low costs and without the system engineering chops required. Scaletorch speeds up your deep learning training between 10x-1000x by leveraging GPU capacity across multiple clouds in a fault-tolerant manner. With Scaletorch accelerate your AI training by 100x and at a 98% lower cost with ZERO CODE CHANGES.

Доклад посвящен актуальным и перспективным разработкам в области реализации алгоритмов машинного зрения и нейронных сетей в медицине, промышленности и других сферах человеческой деятельности, построенных на базе отечественных процессоров НТЦ Модуль.
Доклад о кейсе разработки голосового ассистента для кофейных автоматов компании Unicum. Ассистент отвечает на вопросы пользователей, позволяет принимать заказ и оплачивать напитки с использованием голоса.
DeepPavlov Dream is an open-source multiskill AI assistant platform emerged after the DeepPavlov's DREAM team participation in Amazon Alexa Prize Socialbot Grand Challenges 3 & 4 in 2019-2021. While we used various large language models at DeepPavlov, like DialoGPT, GPT-2, BlenderBot etc., recent developments like ChatGPT and GPT3.5 (davinci-003) made us re-think our design approach to the development of the AI assistants. In this talk you will learn how you can tame the wild power of the generative AI and build your own generative assistants with large language models and DeepPavlov Dream.
12:45 – 13:00
Coffee break
13:00 – 14:00