@VicZ
VicZ
Nothing here yet.
Nothing here yet.
Dec 14, 2024 · 6 min read · TL; DR by carefully selecting where to put aliasing mitigation in a CNN it is possible to limit anti-aliasing information loss. Aliasing negatively impact performances in a way that is not evident from validation loss. Another paper on the effect of ...
Join discussionDec 8, 2024 · 4 min read · The last two posts (1, 2) I published were on the topic of spatial robustness, i.e. the stability of a neural network prediction with respect to small translation and rotation. A curious observation from [1] states that first order adversary seem to ...
Join discussionNov 30, 2024 · 4 min read · TL; DR Many operations in a typical convolutional neural network down-sample the signal without accounting for aliasing. Aliasing causes frequency content that makes a classifier vulnerable to rotation and translations. Aliasing in CNNs Aliasing is a...
Join discussionNov 24, 2024 · 4 min read · TL; DR Image classification model (probably vision models in general) are spatially brittle. Small combination of rotation and translation can radically change model output. Introduction The focus in on spatial robustness, i.e. small spatial transfor...
Join discussionNov 12, 2024 · 2 min read · While this is an older paper (2017), it holds significant relevance in the field of robustness. The paper addresses a fundamental and still relevant question: do neural networks and human vision operate similarly? How do they handle distorted image i...
Join discussion