Interpreting Convolutional Blocks as Feature Embedding by Template Matching for Image Recognition

Date

2023-7-28

Author

Görgün, Ada

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

161
views

91
downloads

Cite This

Known as the key to the success of many neural networks (CNNs), convolutional blocks serve as local feature extractors. Yet, explicit supervision of intermediate layers becomes a major challenge in image recognition since there are no localized annotations for low-level features in practice. In this thesis, this challenge is addressed by referring back to the template matching paradigm. Firstly, to make local semantic feature embedding rather explicit, convolutional blocks are reformulated as feature selection according to the best-matching kernel. Consequently, typical ResNet blocks are shown to perform local feature embedding via template matching once batch normalization followed by a rectified linear unit is interpreted as an arg-max optimizer. Following this perspective, a residual block is tailored to explicitly force semantically meaningful local feature embedding by using class-label information for shaping the intermediate features of CNNs. This concept is expanded through knowledge distillation (KD), which is regularizing the learning of a limited capacity model (student) by pushing its responses to match a powerful model (teacher). Similarly, to explicitly embed the teacher’s knowledge in feature transform, a learnable KD layer is proposed for the student to gain three distinct abilities: i) learning how to leverage the teacher’s knowledge, ii) enabling to discard nuisance information, and iii) feeding forward the transferred knowledge deeper. Additionally, to facilitate template learning in the intermediate layers, a novel form of supervision based on the teacher’s decisions is proposed. Through rigorous experimentation, the effectiveness of the proposed methods is demonstrated, surpassing several state-of-the-art methods on image recognition.

Subject Keywords

Image Recognition, Template Matching, Feature Embedding, Residual Layer, Knowledge Distillation

URI

https://hdl.handle.net/11511/104866

Collections

Graduate School of Natural and Applied Sciences, Thesis