Feature Accentuation: Revealing 'What' Features Respond to in Natural Images

arXiv:2402.10039v1 Announce Type: new
Abstract: Efforts to decode neural network vision models necessitate a comprehensive grasp of both the spatial and semantic facets governing feature responses within images. Most research has primarily centered around attribution methods, which provide explanations in the form of heatmaps, showing where the model directs its attention for a given feature. However, grasping 'where' alone falls short, as numerous studies have highlighted the limitations of those methods and the necessity to understand 'what' the model has …

abstract arxiv attribution cs.cv decode feature features form images natural network neural network research responses semantic spatial type vision vision models