Talk outline: We first see an overview of recent approaches to extract simpler abstractions of complex neural networks using Angluin’s exact learning framework, from computational learning theory. The aim of constructing such abstractions is to obtain high level information from machine learning models, which can be useful to interpret their behavior, detect harmful biases, among others. We then discuss in more detail algorithms for learning logical theories in Angluin’s framework, in particular, those for learning rules in Horn logic. We highlight the benefits and shortcomings of these approaches. Finally, we present promising possible next steps and applications of these approaches for extracting high level information from complex machine learning models such as large language models
Talk outline:In recent decades, several approaches have drawn analogies between pattern mining tasks and constraint programming. The development of modern constraint solvers (SAT, CP, Linear Programming, Answer Set Programming, etc.) has demonstrated the efficiency of these approaches on real-world datasets. More than the efficiency of declarative programming, we argue that the true benefit of this approach lies in its versatility and its ability to integrate expert knowledge. This is especially evident in the case of Answer Set Programming (ASP), originally designed as a tool for knowledge reasoning. In this presentation, we will focus on the task of sequential pattern mining, which involves discovering interesting patterns in a collection of sequences. Firstly, we will introduce the concept of ASP encoding for sequential pattern mining tasks and showcase the advantages of declarative programming in fast-prototyping complex mining tasks. Next, we will highlight the capability of the Clingo solver to efficiently combine reasoning and procedural programming to address the mining of chronicles. Finally, we will conclude this presentation by exploring the potential development of epistemic measures of interestingness at the crossroad of declarative pattern mining and knowledge reasoning. To illustrate the effectiveness of a knowledge-centric approach, we will provide practical examples of these advanced features, focusing specifically on their application in analyzing care pathways to answer pharmaco-epidemiological questions.
Talk outline: TBD
Talk outline: TBD
Talk outline: Not too long ago most headlines talked about our fear of AI. Today, AI is ubiquitous, and the conversation has moved on from whether we should use AI to how we can make the AI systems that we use in our daily lives trustworthy. In this talk I look at some key technical ingredients that help us build confidence and trust in using intelligent technology. I argue that intuitiveness, adaptability, explainability and inclusion of human domain knowledge are essential in building this trust. I present some of the techniques and methods we are building for making AI systems that think and interact with humans in more insightful and personalised ways, enabling humans to better understand the solutions produced by machines, and enabling machines to incorporate human domain knowledge in their reasoning and learning processes.
Talk outline: TBD
Talk outline: The ‘AI alignment problem’ refers to the possibility that AI systems may develop goals that are in conflict with those of their human creators. Part of the problem, I will argue, stems from the fact that computational theories of cognitive science lack a viable account of goal-based behavior in humans and other organisms. Without understanding what our goals are, where they come from, and how we align them with one another, there is little hope of designing AI systems that align with our goals. I argue that non-representationalist, “4E” theories of cognitive science can shed some light on this issue. These theories emphasize that biological intelligence is, at its core, about the dynamic coordination of brain, body, and environment. I will present some of my own work on trying to model the mechanisms of coordination at multiple scales of analysis, and propose that this line of work may help us design AI systems that spontaneously align with the world around them, with each other, and with their human users.
Talk outline: The ability to learn programs from a few examples allows users to automate repetitive tasks in an intuitive way. Existing frameworks on inductive synthesis only perform syntactic manipulations, where they rely on the syntactic structure of the given examples and not their meaning. Any semantic manipulations, such as transforming dates, have to be manually encoded by the designer of the inductive programming framework. In this talk, I will highlight two methods of integrating semantic information with programming by example. First, we extract properties about the input public from public knowledge graphs and provide these properties as inputs to the model. Second, we use the in-context learning capabilities of large language models to solve small subproblems that the deductive backpropagation algorithm behind FlashFill cannot solve. I will conclude with a comparison of advantages and disadvantages of both approaches.