Index

xLSTM-HTR-CTC Extended LSTM for Scalable and Efficient Handwritten Text Recognition with CTC

Automatic Data Augmentation for Multi organ Segmentation

Various medical image segmentation models have been created in the past few years for both supervised and semi-supervised tasks. Since variability and diversity of data play a huge role in the accuracy of such models, choosing the most appropriate data augmentation strategy for each organ or tumor is very important in improving the model outcomes. However, most of these data augmentation strategies used in UNet and its variants are simple, custom-made, and not fully optimal. This manual selection of data augmentation strategies limits the possibility for improvement of the accuracy of the medical image segmentation.
As a further improvement on this, Yang et al [1] proposed a technique for the automation data augmentation strategy selection using reinforcement learning. However, this technique requires a high computational time and only tells about the probability of each augmentation strategy. To overcome this issue, Xu et al. [2] proposed an automatic data augmentation strategy framework called ASNG, which searches for the most optimal augmentation strategy by formulating a bi-level optimization algorithm. This framework also designed a search space that includes the fixed magnitude of the operation and the interval of magnitude. Moreover, this work showcases the dynamic change of strategies during training as per requirement, achieving a state-of-the-art performance.
The thesis aims to research the following:
1. How different data augmentation strategies influence the segmentation performance in different data conditions. The data conditions are based on object shape/size, number of objects, mean and standard deviation of the intensity, and so on.
2. Perform analysis of the correlation of the data conditions and the augmentation methods. For this, first, a search space is to be defined which includes typical augmentation strategies with varying hyperparameters, similar to the work done by He et al [3] and Cubuk et al [4]. The performance of different augmentation methods is tested on a segmentation model built using the MONAI framework.
3. Implement an algorithm for searching strategy using machine learning regressor models like support vector regressor that will lead to automatic planning of the augmentation pipeline based on the target dataset. The dataset used for the algorithm development is AMOS [5]. AMOS is a large-scale, diverse, clinical dataset comprising CT and MRI scans for 15 abdominal organ segmentation.

Image2Tikz: Enhancing Image-to-TikZ Code Generation via Large Language Models Knowledge Distillation

Work description
This thesis aims to develop Image2Tikz, a tool that uses neural networks to transform images into TikZ code by segmenting objects, estimating distances between them, and then translating this information into TikZ. Starting with simple TikZ images and progressing to more complex, hand-drawn versions, the tool will be refined to handle increasing noise levels and complexities. The performance will be evaluated by comparing the generated TikZ code against the original images.

The following questions should be considered:

How can neural networks be effectively trained for tikz object segmentation and distance estimation in images?
What techniques can be used to translate segmented objects and their relative distances into TikZ code?
How does the introduction of noise and complexity in images affect the tool’s accuracy and reliability?
What strategies can improve the tool’s performance on more complex, hand-drawn images?

Prerequisites
Candidates should have a strong foundation in machine learning, particularly in neural networks, with practical experience in Python and familiarity with PyTorch. Skills in image processing and an understanding of LaTeX, especially TikZ, are desirable. The ability to work independently and creatively solve problems is essential.

Please include your transcript of record with your application.

TSI Challenge Summer 2024: Heat & Water Demand Forecasting

The Time Series Intelligence group from the Pattern Recognition Lab offers a 5/10 ECTs project in a challenge format. This is a “contest” where the students are expected to use different machine learning and deep learning methods for time series forecasting. The course is limited to 20 students per semester and they can decide whether to work alone or form a group with another student.

Predictive Modeling for Pre-Conditioning in Vehicles

Machine Learning approach for hiring demand forecasting in Large Scale Organizations

In the field of human resources management, the ability to forecast hiring demand with precision is critical for optimizing workforce planning and talent acquisition strategies. As organizations become increasingly complex, traditional forecasting methods, such as simple time series models or heuristic approaches, often fall short of capturing the multifaceted nature of hiring dynamics. In large multinational corporations, forecasting hiring demand requires the consideration of various factors, including macroeconomic indicators, organizational structure, and workforce fluctuations. This thesis proposes the development of a sophisticated machine learning workflow to enhance the accuracy and reliability of hiring demand predictions.

EcoScapes: LLM-powered advice for crafting sustainable cities

Climate adaptation is vital for the sustainability and sometimes the mere survival of our urban
areas [1, chapters TS.C.8 and TS.D.1]. However, small cities often struggle with limited personnel
resources and integrating vast amounts of data from multiple sources for a comprehensive analysis
[1, chapter TS.D.1.3]. Moreover, the complexity of the topic can overwhelm administrative staff and
local politicians alike. To overcome these challenges, this thesis proposes a multi-layered system
combining specialized Large Language Models (LLMs), satellite imagery and a knowledge base to aid
in developing effective climate adaptation strategies.
Initially, the system uses provided location information to request relevant satellite imagery, which can
be used by all subsequent components.
The architecture’s modular core encompasses several LLMs and expert systems that examine the
satellite data to offer insights on different climate adaptation aspects. Examples of potential functions
might include, but are not limited to, the identification of heat islands, areas threatened by flooding, or
the assessment of vegetation cover.
In the last step, the system consolidates the findings from the preceding modules to generate a
comprehensive report on the existing situation and recommend potential adaptation strategies.
In order to assess the system’s performance, we will compare the generated outputs with those of
unaltered LLMs and a model inspired by ChatClimate [2].

References

[1] P¨ortner, H.-O., D.C. Roberts, H. Adams. et al. 2022: Technical Summary. [H.-O. P¨ortner, D.C. Roberts,
E.S. Poloczanska, K. Mintenbeck, M. Tignor, A. Alegr´ıa, M. Craig, S. Langsdorf, S. L¨oschke, V. M¨oller,
A. Okem (eds.)]. In: Climate Change 2022: Impacts, Adaptation, and Vulnerability. Contribution of
Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change
[H.-O. P¨ortner, D.C. Roberts, M. Tignor, E.S. Poloczanska, K. Mintenbeck, A. Alegr´ıa, M. Craig, S.
Langsdorf, S. L¨oschke, V. M¨oller, A. Okem, B. Rama (eds.)]. Cambridge University Press, Cambridge,
UK and New York, NY, USA, pp. 37-118, doi:10.1017/9781009325844.002.

[2] Vaghefi, S.A., Stammbach, D., Muccione, V. et al. ChatClimate: Grounding conversational AI in climate
science. Commun Earth Environ 4, 480 (2023). https://doi.org/10.1038/s43247-023-01084-x