Teaching

SVGs are everywhere - logos, icons, UI elements, web graphics. The gap between "AI can generate an SVG" and "AI can generate an SVG a designer can actually use" is massive. This thesis maps that gap argues that the structural quality of generated SVGs is primarily shaped by the interaction between two architectural decisions: how the SVG is represented (Bézier paths, shape primitives, latent codes) and what supervision signal drives the learning (pixel reconstruction, text-guided diffusion, dataset-driven training).

This thesis presents an end-to-end framework for deploying Spiking Neural Networks (SNNs) on the THOR neuromorphic processor. The proposed pipeline transforms quantization-aware trained models into hardware-compatible representations, maps them using five strategies including the novel THOR-Native Adaptive Mapping Algorithm (TAMA 2.0) and evaluates them through cycle-accurate simulation.

This thesis comprehensively explores various loss functions in diffusion models, unifying them under the variational lower bound objective framework. Through theoretical analysis and empirical studies, it examines how different objectives affect performance under various conditions and investigates.

The thesis investigates stress detection in sugarbeet crops using Sentinel-2 image time series and various machine learning and deep learning models, with a focus on autoencoder-based architectures. The primary objective is to build a stress detection pipeline that works with minimal supervision.

This thesis addresses pancreatic tumor treatment challenges by ranking drug combinations and concentrations based on their efficacy. Using SimCLR self-supervised learning, it extracts features from bright-field microscopy images of 3D tumor tissue models to characterize drug-induced alterations over time.

This thesis develops surrogate models to predict Key Performance Indicators for Permanent Magnet Synchronous Machine Electric Motors using geometric and physical parameters. The author creates a tabular representation of motor designs, applies Multi Linear Perceptron models.

This thesis evaluates deep learning models for deforestation detection using weakly annotated Sentinel-2 satellite imagery. U-Net and DeepLabV3+ outperform traditional methods but face precision challenges. Results highlight the complexities of multispectral remote sensing and the need for improved annotation quality.

This paper addresses challenges in electrical machine simulation by enhancing deep learning approaches that bypass resource-intensive finite element analysis. Building on previous work (Parekh et al., 2022; 2023), it refines a variational autoencoder architecture that maps diverse motor topologies into a unified latent space, improving both training efficiency and optimization robustness compared to earlier models.

This thesis explores neural network approaches for step detection using accelerometer and gyroscope data. It compares two methods: CNN models that extract spatial features from time-series data with improved accuracy and speed over traditional LSTM models, and a YOLO-like architecture that predicts step start-end pairs with minimal post-processing. The research evaluates models using both traditional metrics and mean absolute error to assess temporal prediction accuracy.

This research addresses the challenge of estimating global wheat cultivation before harvest seasons, which is difficult due to limited quality data despite wheat's importance for global food security. The study explores deep learning techniques with a focus on transfer learning methods, using satellite imagery combined with US agricultural statistics to predict winter and spring wheat production.