I gave a talk, entitled "Explainability like a support", at the above party that talked over anticipations with regards to explainable AI And the way might be enabled in applications.
Weighted product counting generally assumes that weights are only specified on literals, generally necessitating the necessity to introduce auxillary variables. We look at a whole new technique determined by psuedo-Boolean functions, bringing about a far more standard definition. Empirically, we also get SOTA final results.
Will be Talking on the AIUK occasion on concepts and follow of interpretability in machine Discovering.
I attended the SML workshop inside the Black Forest, and mentioned the connections among explainable AI and statistical relational learning.
Gave a talk this Monday in Edinburgh around the concepts & exercise of device learning, masking motivations & insights from our survey paper. Key thoughts elevated integrated, ways to: extract intelligible explanations + modify the design to fit transforming wants.
A consortia project on dependable systems and goverance was acknowledged late very last year. News url here.
Considering coaching neural networks with logical constraints? We have a new paper that aims to whole pleasure of Boolean and linear arithmetic constraints on instruction at AAAI-2022. Congrats to Nick and Rafael!
The report introduces a common sensible framework for reasoning about discrete and constant probabilistic models in dynamical domains.
A the latest collaboration Together with the NatWest Team on explainable machine Mastering is mentioned from the Scotsman. Hyperlink to short article here. A preprint on the final results will be manufactured available shortly.
Jonathan’s paper considers a lifted approached to weighted product integration, such as circuit building. Paulius’ paper develops a measure-theoretic viewpoint on weighted design counting and proposes a method to encode conditional weights on literals analogously to conditional probabilities, which results in sizeable performance improvements.
With the University of Edinburgh, he directs a investigation lab on artificial intelligence, specialising while in the unification of logic and machine Finding out, with a modern emphasis on explainability and ethics.
The paper discusses how to handle nested features and quantification in relational probabilistic graphical models.
The first introduces https://vaishakbelle.com/ a first-order language for reasoning about probabilities in dynamical domains, and the second considers the automated fixing of likelihood troubles laid out in purely natural language.
Convention link Our Focus on symbolically interpreting variational autoencoders, as well as a new learnability for SMT (satisfiability modulo idea) formulas got recognized at ECAI.