Pre-university student (Zentralschweiz, Switzerland), targeting ETH Informatik.
Interested in NeuroAI and the mathematics of learning — how neural networks represent information, and how that relates to biological systems.
RSA comparing Predictive Coding (Rao & Ballard 1999) and Spiking Neural Networks (snnTorch) against THINGS-fMRI (N=3 subjects, V1–IT). Key finding: PC develops a cortical hierarchy gradient (Δr₀−Δr₃ = +0.266, p=0.007, replicated in all 3 subjects); SNN reaches 92% of PC-IT performance but lacks the cross-over. PC r₃ outperforms ResNet-50 at LOC and IT.
PyTorch snnTorch fMRI RSA Predictive Coding NeuroAI
Hierarchical PC network trained on ResNet-50 features, compared to 7T fMRI (THINGS-fMRI, N=3 subjects) via RSA across six cortical ROIs (V1–IT). PC layers show a crossing gradient: r₀ correlates most strongly with V1 (ρ=0.30), r₃ with IT (ρ=0.16). PC r₀ outperforms ViT-B/16 and CLIP at V1 despite those models being trained on orders of magnitude more data.
PyTorch fMRI RSA Predictive Coding NeuroAI
LIF spiking neural network for gesture classification on the IBM DVS128 dataset (11 classes). DVS cameras fire asynchronous events like retinal ganglion cells — pairing them with an SNN creates a doubly bio-inspired pipeline. Achieves 67% validation accuracy with a fully-connected architecture (chance: 9.1%). Documents known limitations: FC-only ignores spatial structure, rate coding discards spike timing precision.
PyTorch snnTorch Event Camera Neuromorphic NeuroAI
Replication and extension of Nanda et al. (2023) on modular arithmetic (a+b) mod p. Part 1: four lines of evidence that the trained model implements a Fourier multiplication algorithm — sparse W_E spectrum (Gini=0.605), W_L effective rank ~10 (95.9% energy in top-10 SVs), 2D-FFT grid structure, and ablation (keys removed → 1.7% accuracy). Part 2: multi-modulus analysis across p ∈ {71, 83, 97, 113} to test whether key frequencies are structurally bound to p. Finding: frequencies are within-p stable across seeds but do not cluster at universal harmonic ratios — frequency selection is p-specific.
PyTorch Mechanistic Interpretability Transformers Fourier Analysis
Systematic ablation on transformer scaling using nanoGPT on FineWeb-Edu (~104M tokens, 5000 iterations per run). Key findings: depth has a strong, monotonically decreasing effect on val loss (L2→L4: −0.160, L6→L8: −0.033); head count has almost no effect — 1 head vs 6 heads differs by only 0.029 val loss at fixed depth. Depth dominates at small scale.
PyTorch nanoGPT Transformers Ablation Study
Representational Similarity Analysis on THINGS-fMRI. Compared ResNet-50, ViT-B/16, and CLIP across six visual areas (V1–IT). Key finding: visual hierarchy explains IT cortex representations better than language-grounded semantics — a negative result for CLIP.
PyTorch fMRI RSA NeuroAI
Reproduction of Power et al. (2022). 1-layer Transformer trained on (a+b) mod 97. Includes phase diagram mapping the transition between memorisation, grokking, and failure to generalise across dataset size and weight decay.
PyTorch Transformers Generalization
Paper-faithful implementation of Mnih et al. (2015). Replay buffer, target network, frame stacking, reward clipping. Trained for 5M frames, achieving ~17 mean reward.
PyTorch Reinforcement Learning Atari
Fully-connected neural network implemented in NumPy without frameworks. Backpropagation derived and coded manually. Applied to Swiss referendum turnout prediction. Grade: 6/6.
NumPy Backpropagation From Scratch
- Swiss Physics Olympiad, Round 2
- Swiss Informatics Olympiad, Round 2
- Read: Goodfellow et al. Deep Learning; Prince Understanding Deep Learning; Gerstner et al. Neuronal Dynamics; Dayan & Abbott Theoretical Neuroscience
Open to research internship opportunities in NeuroAI or ML theory starting autumn 2026.