Tim Berti

About Me

Hi, I'm Tim. After completing my bachelor's degree, spending a year in the military and co-founding StocksOnView, a start-up that brought statistical analysis to retail investors, I'm now back at university doing my master's in physics at KIT. My master's studies focus on theoretical particle physics, optics and financial mathematics. Currently, I am working as a research assistant at the Forschungszentrum Informatik and am writing my thesis on the detection of color vision deficiencies using time series classification on electroencephalography data.

Besides work and studying, I also enjoy doing projects in the field of data science, some of which can be found below. To learn more about a project, just click on the "Read more" button, which will take you to its Github page. Feel free to contact me if you would like to discuss my current research or potential opportunities.

LinkedIn | GitHub

Projects

Identifying Color Vision Deficiency in EEG Data

Stimulating steady-state visually evoked potentials (SSVEP) in subjects with and without color vision deficiencies to investigate differences in brain activity using signal processing and machine learning techniques like RANSAC detrending, Butterworth filters, Canonical Correlation Analysis, Wavelet Transforms, Principal Component Analysis and Support Vector Machines.

Zeiss Computer Vision Hackathon 2024

We placed 2nd as a team of five people who had just met at the hackathon with the development of a camera, image processing and deep learning setup on a Raspberry Pi to recognize free seats in public places such as libraries.

Simulated Wind Power Generation

Simulating different regimes of wind power generation with Hidden Markov Models (HMM) and Generative Adversarial Networks (GAN).

Diffusion-Limited Aggregation

Using convolutional neural networks (CNN) and graph neural networks (GNN), to predict the growth patterns of diffusion-limited aggregation (DLA) clusters, which are fractal structures that form when particles aggregate in a random walk process.

Donut Augmentation

Introduction of a data augmentation technique for CNNs by wrapping images in torus form, which enables shift transformations without loss of information and thus promotes translation invariance in models trained on images with features near the edge.

Constructing Custom Kernels

A notebook that demonstrates employing domain knowledge to construct custom kernels for Support Vector Machines (SVM) to improve classification performance on a synthetic dataset.

Pharynx CV

Analysis of images acquired with a low-cost endoscope and a 3D-printed holder using transfer learning with a ResNet18 model, image segmentation and LIME for feature importance to investigate muscle activation in the mouth and throat during Frenzel equalization, a pressure equalization technique commonly used by scuba divers.

Magnetic Dipole Lattice

Simulation of magnetic dipoles on a 2D lattice, employing Verlet integration and a Nosé thermostat to explore behaviors like magnetic domain formation, phase transitions and hysteresis effects.

Deep Learning Cheat Sheet

A cheat sheet that provides an overview of deep learning concepts, covering some of the core aspects of modern architectures, optimization techniques and regularization methods.