ANC Workshop - 27/05/2025

Speaker: Antonio Vergari (https://april-tools.github.io/)

Title: Trustworthy AI via Tractable Probabilistic Inference

Abstract: What is tractability and why should you care in ML where everything seems to be geared towards just approximations and performance? Well, first of all we always preform tractable computations even when we are computing approximate inference! Understanding tractability and its limits helps us design models that scale better, are more reliable and can better approximate intractable quantities. In this first talk from the april lab, I will introduce many of the ingredients of trustworthy AI and give a perspective on how to realize them through tractable probabilistic models. I will provide a gentle introduction to circuits, a unifying framework for tractable models and show how this framework enables connecting many different formalisms in ML, from mixture models to logical formulas, thus introducing the basics for the other talks of the day.

Speaker: Lorenzo Loconte (https://loreloc.github.io/)

Title: Uncovering the Building Blocks for Expressive Tensor Factorizations

Abstract: Tensor factorizations represent a "Swiss army knife" framework in machine learning, allowing us to compress data, accelerate deep learning models, simulate physics phenomena, and estimate high-dimensional distributions while enabling efficient inference. Many tensor factorization methods have been proposed, each of them exhibiting a different trade-off in terms of efficiency and compression rate (or expressiveness). Although tensor factorizations and circuits appear very different and have different applications, In this talk I will show they are two representations of the same thing: hierarchical decompositions of functions. Under this interpretation, popular factorization methods for probabilistic modelling can be reduced into just two simple characteristics: their circuit structure and how they are parameterized. Then, by navigating different combinations of structures and parameterizations, I will briefly uncover an expressiveness hierarchy surrounding circuits, allowing us to generalize known expressiveness results while suggesting new tensor factorization methods.

Speaker: Emile Van Krieken (https://www.emilevankrieken.com/)

Title: An Introduction to Reliable Neurosymbolic Methods

Abstract: Neurosymbolic (NeSy) AI promises methods that combine neural networks for neural perception with symbolic reasoning to solve tasks like visual reasoning _reliably_. In this talk, I will introduce probabilistic NeSy predictors, some of which developed in our lab. These methods have a strong theoretical backing to, at least partially, guarantee this reliability, and ensure compliance with safety constraints. Then, I will describe several issues with NeSy predictors, starting with Reasoning Shortcuts. This is the phenomenon that models may completely minimise the training loss without learning the correct underlying concepts. I will also discuss issues with a conditional independence assumption that is frequently taken in practical setups, and end with a brief introduction of state-of-the-art methods we worked on for tackling these issues.

Speaker: Leander Kurscheidt

Title: Probabilistic Predictions under Algebraic Constraints

Abstract: Deep learning owes its success to flexibility and scalability—not to respecting constraints. But in safety-critical applications, constraint compliance is a requirement that cannot be overlooked. E.g. if a robot has a chance of hitting a pedestrian or a wall, this chance is not acceptable even if it is 0.1%. In this talk, I will discuss how to combine an expressive probability distribution modeled by a neural network with a rich, flexible language for specifying algebraic constraints, which we can marginalize over efficiently and exactly. This enables us to produce reliable probabilistic predictions using standard neural networks that one can implement and train in pytorch that only predicts outcomes that are feasible according to some given algebraic constraint. We demonstrate the effectiveness of this approach on both synthetic constraint benchmarks and real-world trajectory prediction tasks, where respecting constraints is important.

Speaker: Adrian Javaloy (https://javaloy.netlify.app/)

Title: COPA: Comparing the incomparable in multi-objective model evaluation

Abstract: We often find ourselves asking one embarrassingly simple question: Which of my trained models should I select, if I have many objectives to account for? Perhaps due to shame, we also often sweep this dilemma under the rug and just aggregate all objectives by adding them together. In this talk, we will overcome our fears and investigate how objectives can be automatically normalized and aggregated to systematically navigate their optimal trade-offs. The resulting method, COPA, makes incomparable objectives comparable using their CDFs, similar to copulas, and aggregate them to match user-specific preferences. As a result, COPA allows practitioners to meaningfully navigate and search for models in the Pareto front, which we empirically demonstrate with use-cases in areas such as fair ML, domain generalization, AutoML, or LLM model selection.