Verification of Size Invariance in DNN Activations Using Concept Embeddings - Artificial Intelligence Applications and Innovations Access content directly
Conference Papers Year : 2021

Verification of Size Invariance in DNN Activations Using Concept Embeddings

Abstract

The benefits of deep neural networks (DNNs) have become of interest for safety critical applications like medical ones or automated driving. Here, however, quantitative insights into the DNN inner representations are mandatory [10]. One approach to this is concept analysis, which aims to establish a mapping between the internal representation of a DNN and intuitive semantic concepts. Such can be sub-objects like human body parts that are valuable for validation of pedestrian detection. To our knowledge, concept analysis has not yet been applied to large object detectors, specifically not for sub-parts. Therefore, this work first suggests a substantially improved version of the Net2Vec approach [5] for post-hoc segmentation of sub-objects. Its practical applicability is then demonstrated on a new concept dataset by two exemplary assessments of three standard networks, including the larger Mask R-CNN model [9]: (1) the consistency of body part similarity, and (2) the invariance of internal representations of body parts with respect to the size in pixels of the depicted person. The findings show that the representation of body parts is mostly size invariant, which may suggest an early intelligent fusion of information in different size categories.
Fichier principal
Vignette du fichier
509922_1_En_30_Chapter.pdf (3.47 Mo) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03287709 , version 1 (15-07-2021)

Licence

Attribution

Identifiers

Cite

Gesina Schwalbe. Verification of Size Invariance in DNN Activations Using Concept Embeddings. 17th IFIP International Conference on Artificial Intelligence Applications and Innovations (AIAI), Jun 2021, Hersonissos, Crete, Greece. pp.374-386, ⟨10.1007/978-3-030-79150-6_30⟩. ⟨hal-03287709⟩
27 View
24 Download

Altmetric

Share

Gmail Facebook X LinkedIn More