AAAI 2025 Tutorial

Inferential Machine Learning: Towards Human-collaborative Vision and Language Models

Presented by: Ghassan AlRegib¹, Mohit Prabhushankar¹, and Xiaoqian Wang²
1. Georgia Institute of Technology, 2. Purdue University

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

Duration: Half Day (3 hours, 30 mins)

Neural network driven applications like ChatGPT suffer from hallucinations where they confidently provide inaccurate information. A fundamental reason for this inaccuracy is the feed-forward nature of inductive decisions taken by neural networks. Such decisions are a result of training schemes that do not allow networks to deviate from and creatively abduce reasons at inference. With the advent of foundation models that are adapted across applications and data, humans can directly intervene and prompt vision-language foundation models. However, without understanding the operational limits of the underlying networks, human interventions often lead to unfair, inaccurate, hallucinated and unintelligible outputs. These outputs undermine the trust in foundation models, thereby causing roadblocks to their adoption in everyday lives. In this tutorial, we provide a systematic way to analyze and understand human interventions in neural network functionality at inference. Specifically, our insights are the following: 1) decision theory must be abductive rather than deductive or inductive, 2) interventions must be analyzed as a function of the ‘not-taken’ residual interventions, 3) interventions are not always positive and networks must be equipped to detect unfair and adversarial decisions and interactions. The end goal is to promote a human-AI collaborative environment via inferential machine learning techniques.

The Association of Advancement of Artificial Intelligence has historically been at the forefront of advancing research in AI adoption. As neural network-based AI systems transition from academia to everyday life, prompt-based architectures that allow limited inputs from users during inference stage have gained prominence. However, users employ a ‘trial and error’ policy that may exceed the operational limits as designed by researchers during training. Hence, a larger emphasis must be placed regarding trustworthy measures at inference. This is a timely tutorial that emphasizes inferential machine learning measures which aid AI’s applicability. Since AAAI is attended by stakeholders at all levels, including students, researchers, entrepreneurs, policy makers, media outlets, and other interested parties in academia and industry, it provides an ideal opportunity to emphasize inferential human-collaborative concepts at inference.

Tutorial Location and Time

Venue: AAAI 2025 held at Philadelphia convention center

Room: Room 118C

Date and Time: Feb. 26, 2025 at 2-6 PM

Presenters

Tutorial Content

Neural networks provide generalizable and task-independent representation spaces that have garnered widespread applicability in vision, language, and multimodal understanding applications. The complicated semantics of feature interactions within data has been broken down into a set of non-linear functions, convolution parameters, attention, as well as multi-modal inputs among others. The complexity of these operations has introduced multiple vulnerabilities within neural network architectures. These vulnerabilities include adversarial samples, confidence calibration issues, and catastrophic forgetting among others. Given that AI promises to herald the fourth industrial revolution, it is critical to understand and overcome these vulnerabilities. Doing so requires creating trustworthy neural networks that drive AI systems. Defining trustworthiness, however, is not trivial. Traditionally, trust is a function of trustworthiness attributes including simple invariance to noise and perturbations. However, these measures are neither applicable nor communicable to casual users at inference. In this tutorial, we provide a human-interventionist approach to understanding neural network interactions that allow AI to function in society. Doing so allows us to state the following: 1) neural networks must know when and what they don’t know, 2) neural networks must be amenable to being intervened upon by humans at any stage of their decision-making process. These statements call for trustworthy neural networks that are explainable, intervenable, and fair.

The tutorial is comprised of three major parts. Part 1 discusses some recent results regarding training neural networks with out-of-distribution data, the conclusions of which are that it is not always clear when and how to utilize new and uncertain data in a neural network training setup. We use this as a motivation for trustworthiness measures at inference. A key goal at inference, is for users to garner trust in the neural network predictions. Part 1 of the tutorial tackles this via post-hoc contextual and relevant explanations as shown in Fig.~\ref{fig:Intro}. Context, in Explainability research, arises due to the stakeholders asking the questions. Hence, the goal of Explainability at inference must be to satisfy multiple stakeholders at various levels of expertise, including researchers, engineers, policymakers, and everyday users among others. Technically, this calls for a larger emphasis on contrastive and counterfactual explanations. Part 2 introduces the basic mathematical framework for Intervenability in neural networks. We specifically discuss interventions for the applications of causality, privacy, prompting, interpretability, and robustness. In all cases, humans interact and intervene within the neural network functionality. Some of these interventions maybe positive, for instance, to promote privacy in machine unlearning. On the other hand, without understanding how models function, users may overprompt in a ‘trial-and-error’ fashion leading to worse results (Fig. 2). Part 3 discusses Fairness in trained neural networks, specifically, the dichotomy between demographic privacy and fairness. Several image understanding and robustness applications including anomaly, novelty, adversarial, and out-of-distribution image detection, among others will be discussed. In this tutorial, we examine the types, visual meanings, and interpretations of trustworthy attributes as a human-centric measure of the utility of large-scale neural networks.

History

The organizers have organized tutorials and short courses at recent venues including AAAI’24. Some selected references:

• Tutorial G. AlRegib, M. Prabhushankar, “Robustness at Inference: Towards Explainability, Uncertainty, and Intervenability,” at IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 17, 2024. [Website], [Youtube], [PDF]
  
• Tutorial G. AlRegib, M. Prabhushankar, “Formalizing Robustness in Neural Networks: Explainabilty, Uncertainty, and Intervenability,” at 38th Annual AAAI Conference on Artificial Intelligence (AAAI), Feb. 20, 2024. [Website]
  
• Tutorial G. AlRegib, M. Prabhushankar, “Robust Neural Networks: Towards Explainabilty, Uncertainty, and Intervenability” at IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan. 08, 2024. [Website]
  
• Short Course G. AlRegib and M. Prabhushankar, “Visual Explainability in Machine Learning” at IEEE Signal Processing Society Educational Short Course, Virtual, Dec. 5-7, 2023. [Website]
  
• Tutorial G. AlRegib and M. Prabhushankar, “Robust Neural Networks: Explainability, Uncertainty, and Intervenability,” at IEEE Big Data 2023, Sorrento, Italy, Dec. 15, 2023. [PDF]
  

Audience are expected to have a basic understanding of neural networks and robustness applications including image recognition and detection.

Recent Relevant Publications

•Explainability [1, 2]
•Out-of-distribution Detection [3, 6]
•Adversarial Detection [4]
•Anomaly Detection [5]
•Corruption Detection [3]
•Misprediction Detection [6]
•Causal Analysis [7]
•Open-set Recognition [8]
•Noise Robustness [9]
•Uncertainty Visualization [10]
•Image Quality Assessment [11, 12]
•Saliency Detection [13]
•Novelty Detection [14]
•Disease Severity Detection [15]

[1] AlRegib, G., & Prabhushankar, M. (2022). Explanatory Paradigms in Neural Networks: Towards relevant and contextual explanations. IEEE Signal Processing Magazine, 39(4), 59-72.
[2] Prabhushankar, M., Kwon, G., Temel, D., & AlRegib, G. (2020, October). Contrastive explanations in neural networks. In 2020 IEEE International Conference on Image Processing (ICIP) (pp. 3289-3293). IEEE.
[3] J. Lee, C. Lehman, M. Prabhushankar, and G. AlRegib, “Probing the Purview of Neural Networks via Gradient Analysis,” in IEEE Access, Mar. 21 2023.
[4] 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.
[5] Kwon, G., Prabhushankar, M., Temel, D., & AlRegib, G. (2020, August). Backpropagated gradient representations for anomaly detection. In European Conference on Computer Vision (pp. 206-226). Springer, Cham.
[6] Prabhushankar, M., & AlRegib, G. (2024, August). Counterfactual Gradients-based Quantification of Prediction Trust in Neural Networks. In 2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR) (pp. 529-535). IEEE.
[7] M. Prabhushankar, and G. AlRegib, ”Extracting Causal Visual Features for Limited Label Classification,” in IEEE International Conference on Image Processing (ICIP), Sept. 2021.
[8] Lee, Jinsol, and Ghassan AlRegib. “Open-Set Recognition With Gradient-Based Representations.” 2021 IEEE International Conference on Image Processing (ICIP). IEEE, 2021.
[9] 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
[10] Prabhushankar, M., & AlRegib, G. (2024). Voice: Variance of induced contrastive explanations to quantify uncertainty in neural network interpretability. IEEE Journal of Selected Topics in Signal Processing.
[11] M. Prabhushankar and G. AlRegib, “Stochastic Surprisal: An Inferential Measurement of Free Energy in Neural Networks,” in Frontiers in Neuroscience, Perception Science, Volume 17, Feb. 09 2023.
[12] G. Kwon*, M. Prabhushankar*, D. Temel, and G. AlRegib, “Distorted Representation Space Characterization Through Backpropagated Gradients,” in IEEE International Conference on Image Processing (ICIP), Taipei, Taiwan, Sep. 2019.
[13] 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.
[14] Kwon, G., Prabhushankar, M., Temel, D., & AlRegib, G. (2020, October). Novelty detection through model-based characterization of neural networks. In 2020 IEEE International Conference on Image Processing (ICIP) (pp. 3179-3183). IEEE.
[15] K. Kokilepersaud, M. Prabhushankar, G. AlRegib, S. Trejo Corona, C. Wykoff, “Gradient Based Labeling for Biomarker Classification in OCT,” in IEEE International Conference on Image Processing (ICIP), Bordeaux, France, Oct. 16-19 2022