⚡ If you’d like to know more about a certain topic, send me an email.

⚡ If you’d like to discuss your own AI-related idea, send me an email.

⚠️ If I don’t reply within two days, ping me.

⚠️ All topics below involve a non-trivial amount of coding. For this reason, I prefer candidates who have experience with deep learning libraries (e.g. pytorch) and machine learning workloads (training, hyperparameter selection, interpretation of results, debugging). Please, be upfront about your prior experience, so as to avoid surprises down the line. It may be possible to tailor a project to your specific background.

nesy Improving Hierarchical Classification in NLP with Guarantees

Hierarchical classification is a tough machine learning task: given a document we want to predict multiple, related labels. E.g., reviews of Amazon products can refer to “Electronics”, “Toys”, “Plushies”, “Dog Plushies”, “Seal Plushies”: if a model predicts “Plushies”, it also has to predict either “Dog Plushies” or “Seal Plushies” but not both. Violating these so-called hierarchical constraints could make predictions unusable.

We want to smartly integrate hierarchical NLP classifiers with recent Neuro-Symbolic techniques so as to improve their ability of satisfying hierarchical constraints.

ⓘ References: 1.

ⓘ Requirements: Basic knowledge of (propositional) logic. Preliminary code/data are available.

nesy llm Scaling Up Neuro-Symbolic Autoformalization

Neuro-Symbolic predictors like SPL (see above) can ensure their predictions conform to given symbolic constraints. This is useful in situations in which we want the model absolutely not to output invalid configuration, e.g., in autonomous driving we absolutely do not want cars to cross red lights – with guarantees – as this is essential for safety.

The question is: who gives us the symbolic constraints? Typically, these are provided by human experts, yet this step is expensive and prevents scaling NeSY AI to tasks in which symbolic constraints are not readily available (which are most of them).

In recent work, we evaluated whether LLMs can “autoformalize” symbolic constraints from high-level natural descriptions given by non-experts. The results are promising, but many aspects are not yet tested. How does autoformalization fare when the requested constraint is complex, malformed, incomplete, ambiguous, or out-of-distribution? Can we catch such issues before the constraint is used? And how can we fix them? Can we ask the non-expert to improve their natural language description? Can we make use of imperfect constraints in down-stream NeSy processing? This is what this thesis is about.

ⓘ References: 1.

ⓘ Requirements: Some experience with LLM prompting and basic knowledge of (propositional) logic. Preliminary code/data are available.

xai Integrating Causal and Formal Explainability

Explainable AI develops techniques for understanding the inner workings of machine learning models. Out of the (very many) existing techniques, two particular clusters stand out: causal XAI and formal XAI. Causal approaches leverage ideas from the theory of causality to formalize the explanation process, making it possible also to integrate causal models of how the world works to produce more actionable explanations. Formal approaches neglect actionability, but provide guarantees about the properties of the explanations they produce.

The idea is to take the best of the two fields and come up with algorithms that produce actionable causal explanations with guarantees.

ⓘ References: 1, 2.

ⓘ Requirements: Basic knowledge of linear programming.

xai Finding Compact Faithful Explanations of Large Concept Vocabularies

Concept-bottleneck models are “explainable-by-design” neural networks that route predictions through a bottleneck of high-level concepts. The idea is that users can understand what the model thinks by looking at the bottleneck, which can however be very large (e.g., hundreds of concepts), making interpretation basically impossible. Moreover, we know that summarizing the bottleneck can miss out concepts essential for decision making, also making interpretation difficult.

We came up with algorithms that solve this issue, but the explanations they spit out can be very large if the bottleneck is large. This work is about testing ideas for tackling this issue.

ⓘ References: 2.

ⓘ Requirements: Basic knowledge of linear programming.

xai llm Building a Truly Explainable Zero-Shot Concept-based Model

Concept-based models are neural networks that route class predictions through a bottleneck of high-level concepts. These are predicted by a black-box neural network, but then the class prediction is obtained form the concepts in a white-box manner. In principle, this is great as it can help practitioners interpret and steer the model’s doing.

There are a couple of major issues, though. First, it turns out it is difficult to ensure the learned concepts are in fact human aligned: there is no guarantee that what the model calls a “red car” is (always) a “red car” also for us. Second, when there are many concepts, humans find it difficult to interpret them all at once, and existing concept “summarization” strategies are flawed.

The thesis is about combining the latest findings from large world models, disentangled representation learning, and formal explainability to solve all these problems in one big gulp.

ⓘ References: 1, 2.

ⓘ Requirements: Basic knowledge of linear programming.

xai llm Debugging models with LLMs

Neural nets often solve tasks by exploiting quirks of the data itself. Even if the model looks sensible on the surface, there is a chance it has misunderstood some essential aspect of the prediction problem, e.g., what features to use, how to embed the data, how to use the features/embeddings for prediction.

Typical solutions include unsupervised penalties that aim to prevent models from misunderstanding something (but these often have limited success) or asking experts to inspect and fix the models themselves (but, despite being effective, doing so is expensive).

We want to assess whether some fo the heavy lifting can be carried out by LLMs. There are several sub-questions worth answering: 1) Are LLMs able to spot likely bugs? 2) If so, what kinds of corrective information can we glean from LLMs? 3) Is this information always good, or should we be careful with replacing experts with LLMs?

ⓘ References: 1.

ⓘ Requires further thinking and expertise with LLM prompting.

xai Are GNN Explanations Understandable?

Graph Neural Networks take a graph – representing, for instance, a molecular structure or relational data – and produce a prediction for the graph as a whole or for its elements (nodes, edges, etc.). Researchers explain GNN predictions by attempting to identify a specific sub-graph of the input graph that is somehow causally responsible for the prediction itself. A multitude of techniques for doing so exist.

We want to understand whether this sub-graphs can indeed play the role of explanations: do stakeholders understand them at all? Can they make use of them for replicating and anticipating the model’s reasoning?

ⓘ References: 1.

ⓘ Requirements: Basic knowledge of GNNs and some thinking.