2026 - Thesis & Poster - SSL

USZ


References

  • ViT: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, ICLR 2021.
  • Large-scale pancreatic cancer detection via non-contrast CT and deep learning, Nature 2023.
  • BYOL: Bootstrap your own latent: A new approach to self-supervised Learning, GDM, NeurIPS 2020.
  • CLIP: Learning Transferable Visual Models From Natural Language Supervision, ICML 2021.
  • AlexNet: ImageNet Classification with Deep Convolutional Neural Networks, NeurIPS 2012.
  • 📍 ResNet: Deep Residual Learning, CVPR 2015.
  • 2017 - FLAME: Learning a model of facial shape and expression from 4D scans ( + SMPL)
  • Epochs
        Representation Learning
                   ▲
                   │
Small / Limited Supervision ──── Tabular / Medical Data
  • Moderate clustering metrics across PCA, t-SNE, and UMAP indicate non-random latent structure but insufficient outcome separability, highlighting the need for representation learning beyond geometric proximity in raw tabular space.


The ‘Right Abstraction’ for A System

Category                          | Focus Area / Topics
----------------------------------|------------------------------------------------------------
Representation & Inductive Bias   | Network architecture design
                                  | Multimodal alignment (vision–language–action)
                                  | World models
                                  | 
                                  | Impact of good inductive bias:
                                  | - 10× less data required
                                  | - 100× lower training cost

Learning × Control                | Model-based reinforcement learning
                                  | Differentiable MPC
                                  | Latent dynamics models
                                  | Structured sim-to-real methods
                                  |
                                  | Key insight:
                                  | - Control background is a major advantage
                                  | - Not brute-force learning

System-level AI                   | ML compilers
                                  | Scheduling on heterogeneous hardware
                                  | Memory-aware training
                                  | Inference optimization
                                  |
                                  | These areas strongly reward intelligent
                                  | system and structure design

Robotics / Embodied AI            | Not actuators, SEA, or motors
                                  | (These are constrained by physics)
                                  |
                                  | Instead focus on:
                                  | - Contact representation
                                  | - Hybrid system abstraction
                                  | - Task decomposition
                                  | - Perception–control interfaces
                                  | - Failure-aware planning
                                  |
                                  | Goal:
                                  | - Replace large sets of heuristics with
                                  |   a unified abstraction





















References