Deep Learning Summit Deep Learning Summit schedule

Top 3 Most Complex Projects in the Current Environment

Even if AI provides “state of the art” solutions and research is gaining more ground, in real life Conversational AI doesn’t always have the impact expected by many. Florin and Zsolt Szekely will address the current challenges and will present a live demo of the latest and most complex developments. We at Bosch Service Solutions believe that the time for power point presentations, in which we create unrealistic expectations, need to be put behind us, and the focus should be on projects that can have immediate impact and the capacity to engage users, giving the same importance also to low-resourced languages.

This year, Florin together with his team started the implementation of complex Conversational AI frameworks (available for all languages, channels, and use-cases) designed to improve the AI based solutions within Bosch SO. The target is to continue the deployment of projects in each of the five Conversational AI pillars: Voicebots, Chatbots, Translation AI, Text/Email AI and Agent Assist AI. Prior to this role, Florin was involved in research, start-ups and other summits as keynote speaker - all these experiences giving him a 360 degree perspective on trends and challenges.

Unlock the Business Value of your Multimodal Data via Jina AI

Multimodal AI is a new type of AI that can take in data from multiple modalities (images, text, audio, etc.) and learn from it to make predictions. This makes it very powerful for applications like neural search and creative AI (DALLE and StableDiffusion). Multimodal AI also presents new challenges for engineering. In particular, it can be difficult to deploy and manage multimodal AI models and pipelines and make them scalable in production. This is where MLOps platforms come in. In this talk, we will discuss the benefits of using an MLOps platform for multimodal AI applications in production and help to unlock the business value of your multimodal data.

Dr. Han Xiao is the Founder & CEO of Jina AI, a commercial opensource company based in Berlin. Since its founding in 2020, Jina AI has raised $38M from top investors, including GGV, Cannan, YUNQI, SAP.io. Jina AI is one of the most promising AI startups globally according to CBInsights 2022, 2021 and Forbes DACH 2020. Before Jina AI, Han led a team on neural information retrieval at Tencent AI, laying down the next-gen search infrastructure. Han served as a board member at Linux Foundation AI in 2019, driving the opensource innovation and international collaboration. In 2014-18 Han worked at Zalando Research in Berlin as a Senior Research Scientist. Han received a Ph.D. (2014) and MSc. (2011) in computer science from the Technical University of Munich in Germany.

Dish Catalogue Optimisation: From Unstructured data to knowledge graph

Can you imagine the beauty of the data we work with at Delivery Hero? We hold knowledge of the world’s food culture in 74 countries, as of 2022 including Glovo! To extract the hidden knowledge that our data holds, we need to transform the free and unstructured data into a meaningful catalogue and taxonomy. In this talk we will show you how we use state-of-the-art machine learning models to provide structure, relationship and meaning to our products worldwide. We will also focus on the main challenges, learnings and solutions we applied tackling this complex task, not only from a data science point, but also from a product perspective. Making this data available to our stakeholders at hand we are able to facilitate dominance in innovation, quality and customer satisfaction in our business

Aleksandra did her PhD in the area of complex networks with the goal of knowledge extraction by combining multiple data sources and diverse algorithms. She has passion in bioinformatics and improving health trough food and nutrition data. Currently she works as ML Engineer for the global food delivery service, Delivery Hero.

The legal Concept of Fraud in Machine Learning

In legal terms fraud is not mere lying; it is seeking to obtain advantage, usually monetary, or to put someone else at a disadvantage by lies and deceit. In order to prove a person has acted fraudulently, it is necessary prove that that person was either deliberately or recklessly intended to defraud. There are fine lines to be drawn between 'negligent', 'reckless' and 'deliberate' acts and only the last two suffice to prove fraud. Whilst training machines to detect 'fraud' such fine lines must be observed. Otherwise, framing an action as a 'fraud' whereas there was no intention to 'defraud' might have serious consequences especially for consumers and the relevant actor who is responsible for such misclassification might be found liable to compensate losses suffered by many consumers who have been affected by it. I therefore look forward to discovering more at Berlin Summit how experts define and design 'fraud detection' in their machine training processes.

Özlem Gürses is Professor of Commercial Law at King’s College London. She specialises in insurance and reinsurance law. Özlem is the author of Reinsuring Clauses (Informa), Marine Insurance Law (Routledge), Insurance of Commercial Risks (Sweet and Maxwell), and The Compulsory Motor Vehicle Insurance (Informa) as well as numerous articles published on insurance and reinsurance related topics. Özlem sits in the British Insurance Law Association Committee and the Presidential Council of the International Insurance Law Association (AIDA). She is Vice-Chair of the Reinsurance Working Party of AIDA. Özlem teaches insurance and reinsurance law at King’s College London and abroad, including National University of Singapore, University of Hamburg and World Maritime University, Malmö

We Give it to the Projects, Spread it Across the Countries!" - NLP at Munich Re.

• Since 2018, Munich Re is developing an in-house NLP platform to quickly transfer research insights into production-ready models. While it is quite common that knowledge gets lost from one project to another, we systematically capture reusable components in our own library, standardize and share code and models across projects for reuse and collaboration. This facilitates the creation of new solutions for any of our companies.

• Roshan works as a Lead Architect for Business Technology at Munich Re and consults various projects including the NLP platform. Prior, he worked in the automobile industry handling Big Data, MLOps and data platform topics. He holds a Master’s degree in Software Engineering. He is passionate about serverless technologies and loves helping organizations in their data, AI and cloud journeys .

Applying NLP in the Automotive Domain and Challenges Involved

Virendra is an NLP and Deep Learning engineer at the Artificial Intelligence team of BMW Group. At BMW, he primarily focuses on industrializing AI in the quality domain. He takes particular interest in transforming business requirements into AI problems and using semi-supervised approaches, especially in the absence of labeled data.

He moved to Germany for his Master's in Informatics at TU Munich. Previously, he worked in the semiconductor industry, building Linux drivers and compiler optimizations for HPC processors.

Reinforcement Learning in Transport and Logistics

This presentation is dedicated to the challenges and successes of utilizing reinforcement learning as an optimization tool in transport.

Nima is a Researcher for Machine Learning at Max Planck Institute and Senior RL Research Engineer at InstaDeep Ltd. He is an AI-practitioner and experienced Simulation Scientist with focus on Complex Systems.

Real Time Supervised Anomaly Detection

Anomalies are patterns in data not conforming to expected behavior. Discovery of such patterns leads to actionable insights as anomalies often correspond to undesired states e.g. reduced quality or some failure. While automatic finding of anomalies is a challenge in itself, it is usually followed up by another hard task to understand the causes of their occurrence, in order to prevent them from happening again. The possibility to explain the decision of the anomaly detection system not only helps to establish trust but also to identify the root causes. In this talk I investigate this connection and challenges accompanying Deep Learning approaches.

Dr. Sergei Bobrovskyi is a Data Scientist within the Analytics Accelerator team of the Airbus Digital Transformation Office. His work focuses on applications of AI for anomaly detection in time series, spanning various use-cases across Airbus. Prior to Airbus he worked on automated fraud detection for one of the largest e-commerce companies in Germany. Before that he was engaged in various research related positions in the space industry.

Sergei holds a PhD in theoretical physics as well as a physics Diploma from the University of Hamburg. Besides physics he also studied philosophy with an emphasis on the philosophy of mind.