Fenil Doshi

Thomas Fel, Matthew Kowal, Mozes Jacobs et al.

What is the geometry of a visual percept? The most widely used protocols for decomposing neural network representations into interpretable parts treat concepts as isolated directions, yet recent work shows that concepts are often realized as geometric structures in low dimensional regions of activation space. We turn to the literature of Structured sparsity to close this gap, and show that block sparsity, which groups directions into blocks, is the prior matched to a generative model in which a representation is a sparse sum of low-dimensional manifolds: the modern, learned form of a classical idea in visual neuroscience, where a visual feature is carried by a coordinated group of neurons rather than a single tuned one. We implement three variants of block-sparse featurizers (BSFs) and, through a minimum-description-length analysis, show that all three describe activations more compactly than direction-based featurizers, with the recovered concepts typically two- to four-dimensional. We then use BSFs to (i) recontextualize prior work, showing that curve detectors in InceptionV1 actually read from a single continuous curve manifold, (ii) discover novel manifolds including shadows and lighting in DINOv3, and (iii) support interpretable control of image generation in diffusion models (SDXL) via manifold steering.

Fenil Doshi, Jimit Gandhi, Deep Gosalia et al.

Social media networks and chatting platforms often use an informal version of natural text. Adversarial spelling attacks also tend to alter the input text by modifying the characters in the text. Normalizing these texts is an essential step for various applications like language translation and text to speech synthesis where the models are trained over clean regular English language. We propose a new robust model to perform text normalization. Our system uses the BERT language model to predict the masked words that correspond to the unnormalized words. We propose two unique masking strategies that try to replace the unnormalized words in the text with their root form using a unique score based on phonetic and string similarity metrics.We use human-centric evaluations where volunteers were asked to rank the normalized text. Our strategies yield an accuracy of 86.7% and 83.2% which indicates the effectiveness of our system in dealing with text normalization.