30 Years Max Planck Institute for Informatics

Machine learning has made great advances in all fields of science and technology. Notably, large language models (LLMs) have been in the spotlight of public and scientific attention for some time now, representing an important class of foundation models, that is, models that can be adapted for a wide range of downstream tasks. This is done leveraging word embeddings that are derived from large text corpora. Less known, a specialized class of foundation models have arisen in computational biology: protein language models (PLMs). Similar to LLMs, they produce embeddings for protein sequences, which can be considered biological texts. These embeddings can be used in various protein-related downstream tasks, such as predicting protein functions, effects of mutations, or drug-target interactions.

In my talk, I will present our recent work on development of structure-aware PLMs with a specific application in predicting drug-target interactions. Further, I will introduce natural products as a rich source of novel candidate drugs and talk about applications of PLMs in natural product research.

Advances in networking are driven by evolving needs of society and the development of new applications. At the same time, advances in networking technology have fueled the creation of novel digital connected experiences. This talk will discuss this virtuous cycle of network innovation over the last three decades and the inflection points in the technology landscape that have led us to the current large scale, programmable, virtualized edge-to-cloud networks. The talk will also cover a view into the future network evolution, including the possibilities with generative AI, and the associated new challenges.

Key-value stores and search engines are posing a continuously growing need to efficiently store, retrieve and analyze massive sets of keys under the many and different requirements posed by users, devices, and applications. Such a new level of complexity could not be properly handled by known data structures, so that academic and industrial researchers started recently to devise new approaches that integrate classic data structures with various kinds of advanced techniques drawn from data compression, computational geometry and machine learning, hence originating what are currently called “Learned Data Structures”.

In this talk, I’ll survey the evolution of these novel kinds of data structures, discuss their theoretical and experimental performance, and point out new challenges worth of future research.

The creation of lifelike digital human faces has been pivotal in a range of applications, spanning from healthcare and telepresence to virtual assistants and cinematic visual effects. For decades, the ultimate objective has been to create digital representations so authentic that they are virtually indistinguishable from real faces, while also conveying genuine emotional depth. Overcoming the challenge of the “uncanny valley” has been crucial to this pursuit. In this talk, I will give a 30-year retrospective of pioneering research in digital humans. We will explore the evolution of various elements —including facial capture techniques, geometry, appearance modeling, soft tissue modeling as well as eyes, teeth, and hair.

The talk will also highlight the transformative impact of contemporary machine learning on facial visual effects. As we look toward the future, the focus will shift to real-time facial animation and the symbiotic relationship between digital characters and machine learning algorithms to bring AI avatars to life.

Recent advances in AI have led to a significant improvement in the ability to understand visual content. Given these advances, AI-based solutions have been increasingly adopted in industry. In turn, real-world constraints, such as compute efficiency, data efficiency, model size, and so forth, need to be addressed. This talk will cover how to handle these constraints. In particular, I will introduce techniques for designing and optimizing compute-efficient neural networks. Furthermore, I will discuss data-efficient machine learning methods.

Analyzing large data sets often involves complex query execution plans, where different evaluation strategies can be used to answer a given query. This leads to interesting optimization problems, where we can invest time in optimization to speed up query execution, sometimes by orders of magnitude. On the other hand query optimization itself can become expensive, too, as many optimization problems are NP hard. In this talk we study approaches for practical query optimization that can handle complex, real world data sets and queries and that can be integrated into production query processing engines.

In randomized experiments, we randomly assign the treatment that each experimental subject receives. Randomization can help us accurately estimate the difference in treatment effects with high probability. It also helps ensure that the groups of subjects receiving each treatment are similar. If we have already measured characteristics of our subjects that we think could influence their response to treatment, then we can increase the precision of our estimates of treatment effects by balancing those characteristics between the groups.

We show how to use the recently developed Gram-Schmidt Walk algorithm of Bansal, Dadush, Garg, and Lovett to efficiently assign treatments to subjects in a way that balances known characteristics without sacrificing the benefits of randomization. These allow us to obtain more accurate estimates of treatment effects to the extent that the measured characteristics are predictive of treatment effects, while also bounding the worst-case behavior when they are not.

This is joint work with Chris Harshaw, Fredrik Sävje, and Peng Zhang.

This talk introduces the Emerald system, a new technology that uses wireless sensing and machine learning to implement touchless health monitoring. An Emerald device in the home transmits low-power radio signals and records their reflections. The recorded data is then passed to the cloud for further processing using neural network algorithms. The monitoring system can infer the movements, breathing, heart rates, and sleep stages of people in their homes, without requiring them to wear any sensors. By monitoring physiological signals continuously over a period of months, the system can automatically detect changes in health conditions. The talk will describe the underlying technology and present examples of its application in clinical drug trials.

Everyday human activities are impressive feats of physical intelligence, for instance, the highly coordinated movement of fingers to type this sentence. Robots with even a fraction of this capability would revolutionize our lives but remain elusive.

In this talk, I will discuss my group‘s work on taking a step towards more capable robots by building 3D computer vision and machine learning methods for understanding human manual skills -- the diverse and dexterous ways in which we shape our environment with our hands.I will discuss the techniques we are building to capture, aSnalyze, and generate hand skills by observing adults and children. Our goal is to produce a large repository of high-level manual skills that can lead to a deeper understanding of hands and eventually be transferred to robots.

Machine learning has been making steady progress in database systems for many years, with tasks such as natural language interfaces, data cleaning, integration, imputation, and forecasting. The advent of large language models (LLMs) has accelerated this progress, particularly in the area of natural language interfaces.

In this talk, we will discuss the research challenges that arise in text-to-SQL generation over large, complex schemas. We will present recent research in high-recall schema subsetting for retrieval augmentation, online customization to new schemas, and handling of ambiguities. We will also discuss the potential of LLMs to enrich data analysis with explanations from outside the structured database.

However, custom models for specific data analysis tasks, such as imputation and forecasting, still need to be performed outside the LLM. We will present recent deep learning advances for performing these tasks on multidimensional time series data. The talk will conclude with a discussion of the many challenges that still remain on the path to effectively integrate the orderly self-contained world of database systems and the stochastic open-ended world of AI-driven approaches.

In the last few decades supercomputers were primarily used for applications in scientific computing. However, during the same period, high performance computers were being designed for graphics and imaging. Simulating realistic physical phenomena is also required to create immersive virtual worlds. Similarly, processing images and videos is data intensive, and analyzing them requires fast implementations of compute intensive signal processing, statistics, and machine learning algorithms. There is a large consumer demand and market for these these capabilities. These quests for cost efficient powerful computers led to the modern graphics processing unit (GPU).

In this talk I will give a personal history of my journey developing software and hardware for high performance graphics supercomputers.