IEEE ICIP 2023 Tutorial

Title: A Multi-Faceted View of Gradients in Neural Networks: Extraction, Interpretation and Applications in Image Understanding

The slides for each section can be found here: [Part1][Part2][Part3][Part4][Part5]

Type / Duration: Half-Day Tutorial (3h)

Presenters: Ghassan AlRegib, Mohit Prabhushankar

alregib@gatech.edu, mohit.p@gatech.edu 

Georgia Institute of Technology 

www.ghassanalregib.info

A description of the tutorial topic, providing a sense of both the scope and depth of the tutorial, along with a tutorial outline.

In this tutorial, we motivate, analyze and apply gradients of neural networks as features to understand image data. Traditionally, gradients are utilized as a computationally effective methodology to learn billions of parameters in large scale neural networks. Recently, gradients in neural networks have shown applicability in understanding and evaluating trained networks. For example, while gradients with respect to network parameters are used for learning image semantics, gradients with respect to input images are used to break the network parameters by creating adversarial data. Similarly, gradients with respect to logits provide predictive explanations while gradients with respect to loss function provide contrastive explanations. We hypothesize that once a neural network is trained, it acts as a knowledge base through which different types of gradients can be used to traverse adversarial, contrastive, explanatory, counterfactual representation spaces. Several image understanding and robustness applications including anomaly, novelty, adversarial, and out-of-distribution image detection, and noise recognition experiments among others use multiple types of gradients as features. In this tutorial, we examine the types, visual meanings, and interpretations of gradients along with their applicability in multiple applications.

The tutorial is composed of four major parts. Part 1 discusses the different interpretations of gradients extracted from trained neural networks with respect to input data, loss, and logits. Part 2 covers in detail a theoretical analysis of gradients. Part 3 describes the utility of gradient types in robustness applications of detection, recognition and explanations. Newer and emerging fields like machine teaching and active learning will be discussed with methodologies that use gradients. Part 4 connects the human visual perception with machine perception. Specifically, we discuss the expectancy-mismatch principle in neuroscience and empirically discuss this principle with respect to gradients. Results from Image Quality Assessment and Human Visual Saliency will be discussed to demonstrate the value of gradient-based methods. The outline as well as the expected time for each part is presented below.

Tutorial Outline

---

Relevance: Address the importance and timeliness of the proposal, and its relevance to researchers and engineers.

Neural networks provide generalizable and task independent representation spaces that have garnered widespread applicability in image understanding applications. The complicated semantics of feature interactions within image data has been broken down into a set of non-linear functions and convolution parameters. Gradients are the computationally effective means of learning these functions and parameters. However, gradients’ effectiveness in network learning has obscured their utility as analysis and inferential tools. Recently, gradients as features has gained relevance with the robustness issues faced by neural networks. Gradients are theoretically interpreted as proxies for information, post-hoc explanations, and epistemic uncertainty. This has increased its applicability, particularly in the last three years. Specific to neural networks, changing the output classes to induce confounding features is a new and exciting topic that can have wide applicability in the immediate future.

Recent Relevant Publications

1. AlRegib, Ghassan, and Mohit Prabhushankar. “Explanatory Paradigms in Neural Networks: Towards relevant and contextual explanations.” IEEE Signal Processing Magazine 39.4 (2022): 59-72.
2. M. Prabhushankar, and G. AlRegib, “Introspective Learning : A Two-Stage Approach for Inference in Neural Networks,” in Advances in Neural Information Processing Systems (NeurIPS), New Orleans, LA,, Nov. 29 – Dec. 1 2022.
3. J. Lee, M. Prabhushankar, and G. AlRegib, “Gradient-Based Adversarial and Out-of-Distribution Detection,” in International Conference on Machine Learning (ICML) Workshop on New Frontiers in Adversarial Machine Learning, Baltimore, MD, Jul. 2022.
4. M. Prabhushankar and G. AlRegib, “Extracting Causal Visual Features for Limited Label Classification,” IEEE International Conference on Image Processing (ICIP), Anchorage, AK, Sept 2021.
5. J. Lee and G. AlRegib, “Open-Set Recognition with Gradient-Based Representations,” IEEE International Conference on Image Processing (ICIP), Anchorage, AK, submitted on Jan. 18 2021.
6. G. Kwon, M. Prabhushankar, D. Temel, and G. AlRegib, “Backpropagated Gradient Representations for Anomaly Detection,” in Proceedings of the European Conference on Computer Vision (ECCV), SEC, Glasgow, Aug. 23-28 2020.
7. M. Prabhushankar, G. Kwon, D. Temel, and G. AlRegib, “Contrastive Explanations in Neural Networks,” in IEEE International Conference on Image Processing (ICIP), Abu Dhabi, United Arab Emirates, Oct. 2020. (Top Viewed Special Session Paper Award)
8. Y. Sun, M. Prabhushankar, and G. AlRegib, “Implicit Saliency in Deep Neural Networks,” in IEEE International Conference on Image Processing (ICIP), Abu Dhabi, United Arab Emirates, Oct. 2020.
9. G. Kwon, M. Prabhushankar, D. Temel, and G. AlRegib, “Novelty Detection Through Model-Based Characterization of Neural Networks,” in IEEE International Conference on Image Processing (ICIP), Abu Dhabi, United Arab Emirates, Oct. 2020.
10. J. Lee and G. AlRegib, “Gradients as a Measure of Uncertainty in Neural Networks,” in IEEE International Conference on Image Processing (ICIP), Abu Dhabi, United Arab Emirates, Oct. 2020.
11. M. Prabhushankar*, G. Kwon*, D. Temel and G. AIRegib, “Distorted Representation Space Characterization Through Backpropagated Gradients,” 2019 IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, 2019, pp. 2651-2655. (* : equal contribution, Best Paper Award (top 0.1%))