IfI Research Talk Series

The IfI Research Talks will be held at the date indicated below, normally in room BIN 1.D.29. They usually start at 13:15 (check exact time in the table below) and last for about 30 minutes.

* please contact the speaker for online access

Speaker: Prof. Dr. Jürgen Bernard

In his Talk, Prof. Jürgen Bernard motivates the benefits of human-machine collaboration to solve data science and analysis challenges. Along these lines, he outlies typical bottlenecks in the human-machine collaboration, and gives indications on how his research helps solving these problems.

In particular, Jürgen Bernard emhpasizes the potentials of leveraging preferences of humans when it comes to data analysis, and discusses how interactive machine learning systems can replay the information that was learned from humans expressing feedback.

In two real-world examples, Jürgen Bernard shows how bi-directional human-machine feedback can be applied in practice, to leverage data preferences: for the preference-based creation of item rankings and human-centered similarity search.

Links: IVDA Group  |  Digital Society Initiative

Speaker: Prof. Dr. Dan Olteanu

Abstract:

In this talk I will overview recent efforts of the DaST (Data Systems and Theory) group on the problem of maintaining real-time analytics over large and continuously evolving relational databases.

We looked at the maintenance problem from different angles. We addressed several fundamental questions, including:

• Which database queries can be maintained in worst-case optimal time?
• Can we trade the query answer time for fast update time?
• How can we maintain models trained over changing relational data?

Our answers include:

• An algorithm with the best possible (worst-case) computational complexity for maintaining the number of triangles in a graph under insertions and deletions of edges.
• The computational complexity trade-off between update time and the answer time for a large class of queries.
• A new algebraic structure called the covariance ring that defines the shared maintenance of the covariance matrix for features in the database.

Equipped with a theoretical understanding of the maintenance problem, we focused on the design and implementation of maintenance systems:

• F-IVM (https://github.com/fdbresearch/FIVM) is our open-source state-of-the-art system. Using one thread of a commodity machine, F-IVM can maintain queries and linear regression models over database joins with a throughput of ten million updates per second.
• The RelationalAI commercial engine (https://relational.ai) incorporated some of our theoretical and system design ideas.

This is joint work with DaST members Ahmet Kara and Haozhe Zhang, Milos Nikolic of U. Edinburgh, and Henrik Barthels, Mohamed ElSeidy and Niko Göbels from the RelationalAI IVM team.

Speaker: Prof. Dr. Alberto Bacchelli

Abstract:

In the majority of software companies and projects, code review is a crucial step in the development process. This involves having developers, other than the original author, manually evaluate any changes to the code before it is integrated into the production software. Despite the widespread adoption of code review in software engineering, there has been a noticeable lack of innovation in the way it is conducted over the past decade or so. This stagnation persists even though there are ample opportunities for refining and rethinking conventional practices.

In this presentation, I will share a series of valuable insights that my team and I have gained through extensive research on code review over the past ten years. Our findings have the potential to significantly enhance current methodologies and tools, leading to more efficient and effective code review processes.

Speaker: Prof. Dr. Manuel Günther

Abstract:

When image classification methods leave academic lab settings, they will be confronted with items that they have never seen before.

Open-set recognition tries to balance between correctly classifying samples from known categories, while not getting confused by unknown items.
In this presentation, I will give a brief introduction to current methods applied to improve open-set recognition.

Furthermore, I will present our new large-scale evaluation framework for open-set image recognition tasks, and the latest findings from these evaluations.

Particularly, when unknown items are very dissimilar to the known classes, they can be rejected easily.
On the other hand, when known and unknown classes come from the same domains, none of our research efforts in the last 10 years provides any real advantage.

Speaker: Prof. Dr. Ingo Scholtes

Abstract:

Graph Neural Networks (GNNs) have become a cornerstone for the application of deep learning to data on complex networks. However, we increasingly have access to time-resolved data that not only capture which nodes are connected to each other, but also when and in which temporal order those connections occur. A number of works have shown how the timing and ordering of links shapes the causal topology of networked systems, i.e., which nodes can influence each other over time. Moreover, higher-order models have been developed that allow us to model patterns in the resulting causal topology. While those works have shed light on the question how the time dimension of temporal graphs influences node centralities, community structures, or diffusion processes, we lack methods to incorporate those insights into state-of-the-art graph learning techniques.
 
Addressing this gap, we introduce De Bruijn Graph Neural Networks (DBGNNs), a time-aware graph neural network architecture for temporal network data. Our approach accounts for temporal-topological patterns that unfold via causal walks, i.e., temporally ordered sequences of links by which nodes can influence each other over time. We develop a graph neural network architecture that utilizes De Bruijn graphs of multiple orders to implement a message passing scheme that follows a non-Markovian dynamics, which enables us to learn patterns in the causal topology of dynamic graphs.

Speaker: Prof. Dr. Martin Volk

Abstract:

In the first part of my talk I will look back on my group’s activities of building and exploiting large multilingual corpora with texts that span more than 100 years, and on our research in machine translation.

In the second part, I will present our recent work on a large collection of 16^th century letters from and to the Zurich reformer Heinrich Bullinger. We investigated handwriting recognition, automatic translation, name recognition and linking, and topic classification for these 12,000 letters. I will show that ChatGPT performs well on most these tasks even though we are dealing with Latin and Early New High German. I will also speculate on how ChatGPT can reconstruct lost letters in this collection.