Adv_ML — Reproducibility study of "Vision Transformers Need Registers"

Group project for the USI ATML course (2026). We reproduce the central claims of Darcet et al., ICLR 2024.

Structure

src/Kun/        # Section 2 — outlier characterisation (Figs. 2, 3, 5, Tab. 1)
src/Paolo/      # Section 3.2 — linear evaluation + Figure 8 ablation (Tabs. 2a, 2b)
src/Raffaele/   # Section 4 — LOST object discovery (Tab. 3, Fig. 9) + Analysis of LOST performance (Figs. 13, 14, 7) + Behavior of models with registers (Fig. 15, 16, Tabs. 4, 5)
report/ATML_Report_2026/   # LaTeX report
assets/         # shared figures from the paper

Each member's section is self-contained: scripts produce numerical results and figures under src/<member>/results/, which are then referenced from report/.

Contributors

• Kun Zhan — Section 2
• Paolo Deidda — Section 3.2
• Raffaele Perri — Section 4