Michael Neuhold

Compact prediction learning is a training principle that encourages neural networks to reuse a small number of latent representations while maintaining task performance. This bachelor thesis investigates whether such compactness constraints lead to structured and reusable internal representations in neural networks.

The approach is evaluated on a set of synthetic binary prediction tasks and synthetic linear prediction tasks as well as on a convolutional neural network trained on a pairwise MNIST comparison task. Compactness is enforced by an additional loss term applied to the latent representations. The compactness objective is defined as:

E = \sum_{k \neq l} \Bigl( 1 + \log \sum_i (\lvert g_i(x^{(k)}) - g_i(x^{(l)}) \rvert) \Bigr).

The results show that for simple synthetic tasks, compact prediction learning leads to highly structured and low-dimensional latent spaces that reflect task-relevant structure. In contrast, for the convolutional setting, the compact objective causes a strong collapse of the latent space, removing most digit-specific information and reducing classification accuracy.

These findings indicate that compact prediction learning acts as a strong inductive bias that enforces minimal, task-aligned representations, but may limit information retention in more complex perceptual tasks.

Noch nicht ganz so weit ;) (Sommersemester 2027)

This diploma thesis addresses the limited evaluation of battery charging processes in AVL Drive 5, where previously only a few parameters were considered. To over- come this limitation, the existing Python-based analysis was extended to support both AC and DC charging sessions and to automatically calculate a broad range of key performance indicators. The implemented solution standardizes input data, detects charging events, and generates comprehensive results that are directly integrated into AVL Drive 5 and its database. As a result, charging analyses are now more complete, accurate, and immediately available to engineers without additional manual effort. The outcome is an efficient and scalable framework that significantly enhances benchmarking capabilities and provides a solid foundation for future extensions in charging analysis.