8 papers
Multi-Way Representation Alignment
Akshit Achara, Tatiana Gaintseva, Mateo Mahaut +5
The Platonic Representation Hypothesis suggests that independently trained neural networks converge to increasingly similar latent spaces. However, current strategies for mapping t…
Adversarial Attacks Leverage Interference Between Features in Superposition
Edward Stevinson, Lucas Prieto, Melih Barsbey +1
Why do adversarial examples exist, and why do they transfer between models? Existing explanations appeal to high-dimensional geometry, non-robust patterns in the input, and decisio…
From Data Statistics to Feature Geometry: How Correlations Shape Superposition
Lucas Prieto, Edward Stevinson, Melih Barsbey +2
A central idea in mechanistic interpretability is that neural networks represent more features than they have dimensions, arranging them in superposition to form an over-complete b…
On the Interaction of Compressibility and Adversarial Robustness
Melih Barsbey, Antônio H. Ribeiro, Umut ÅimÅekli +1
Modern neural networks are expected to simultaneously satisfy a host of desirable properties: accurate fitting to training data, generalization to unseen inputs, parameter and comp…
Large Learning Rates Simultaneously Achieve Robustness to Spurious Correlations and Compressibility
Melih Barsbey, Lucas Prieto, Stefanos Zafeiriou +1
Robustness and resource-efficiency are two highly desirable properties for modern machine learning models. However, achieving them jointly remains a challenge. In this paper, we id…
Grokking at the Edge of Numerical Stability
Lucas Prieto, Melih Barsbey, Pedro A. M. Mediano +1
Grokking, the sudden generalization that occurs after prolonged overfitting, is a surprising phenomenon challenging our understanding of deep learning. Although significant progres…