Conférences  >  Mathématiques  >  Réseaux de neurones et intelligence artificielle, apprentissage automatique  >  États-Unis

California Institute of Technology (Caltech)

Démonstration automatique de théorèmes,    Optimisation, recherche opérationnelle

California Institute of Technology (Caltech)

This annual workshop welcomes academics and industry practitioners at the interface of theoretical physics, mathematics, and computer science to discuss research directions centred on the application of machine learning and data science to open problems in fundamental theory, and equivalently the application of techniques from theory to describe the successes of machine learning.

Institute for Pure and Applied Mathematics (IPAM), Los Angeles, CA

The purpose of this workshop is to bring together neuroscientists who collect and study these data sets with mathematicians and other theorists who develop techniques to model and analyze networks, network dynamics, and dynamical processes on networks. We expect that crossing disciplinary boundaries will greatly facilitate progress, as neuroscientists working in connectomics often lack exposure to recent mathematical developments, while the biological and technical details that underlie connectomic data may not be familiar to mathematicians. This workshop will help define directions of future work in connectomics, with deep links to neuroscience, mathematics, and data science.

Institute for Pure and Applied Mathematics (IPAM), UCLA

In this workshop, we will explore the statistical and computational requirements for solving various learning problems. The statistical limit is the minimum number of samples needed to solve a learning problem. In contrast, the computational limit is the minimum number of samples required for the problem to be solvable by an efficient algorithm. There is much research on the statistical requirements for many important learning problems, but the computational requirements are less well-understood. We often have large gaps between the two for several important problems (e.g., sparse linear regression). In addition, there are also gaps in our understanding of the costs of various constraints on learning, such as privacy, fairness, interpretability, robustness, and parallelization. This workshop will provide a forum to discuss the latest research and develop new ideas on the above questions. It will help build bridges between different disciplines, such as applied mathematics, statistics, optimization, and theoretical computer science, which will lead to more effective solutions to challenges in statistical inference.

Institute for Pure and Applied Mathematics (IPAM), UCLA

The study of biological intelligences has been transformed by breakthroughs in experimental and observational methods. New data have expanded our appreciation of the sophisticated behaviors exhibited by human and non-human animals across a range of taxa. Traces of intelligent behavior are typically high-dimensional, with complex structure in space and time. They often involve multiple data streams: for example, positional information from individual members of a bird flock, as well as “point-of-view” recordings showing what birds are looking at, moment by moment; or electrocorticographic recordings of human brain activity, as well as detailed traces of motor behavior. These traces are often buried beneath complex noise environments: individual animal movements must be pulled out of extensive visual clutter; whale song must be isolated from ocean noise. This complex, high-dimensional data requires entirely new approaches to data representation, integration, and analysis. This workshop will focus on the challenges raised by high-throughput, high-dimensional studies of intelligent behavior. It will bring together experts in animal cognition, computational neuroscience, and cognitive science with mathematicians and computer scientists focused on relevant methods in machine learning, network science, high-dimensional statistics, and information theory.

Part of the Long Program Mathematics of Intelligences

Institute for Pure and Applied Mathematics (IPAM), UCLA

Neural networks have confounded traditional ML beliefs about the dangers of overfitting and the need for regularization. They have also given rise to many new empirical findings around transfer learning, adversarial examples, compressibility, scaling laws (relating the size of datasets, models, and compute), grokking, and so on. What is needed to explain and predict all this is a rich new theory of learning capable of addressing the delicate interplay between model, data, and optimizers at large scale. This workshop will bring together top researchers driving the frontiers of this work with experts in both theory and experiment for natural intelligence. The result will be a scholarly discussion on how to frame questions about learning and how to distill the similarities and differences between learning with biological and artificial systems.

Part of the Long Program Mathematics of Intelligences

Institute for Pure and Applied Mathematics (IPAM), UCLA

This workshop will draw together researchers creating new algorithms and architectures (e.g., active symbol architectures, evolutionary programming approaches, neural program synthesis) with mathematicians and theoretical computer scientists who specialize in non-convex optimization, the theory of programming languages, type theory, proof theory, and category theory. It aims to promote cross-fertilization between these paradigms and more traditional approaches, while stimulating the development of rigorous foundations for evolutionary computing, program synthesis, and other naturalistic approaches to AI.

Part of the Long Program Mathematics of Intelligences

Institute for Pure and Applied Mathematics (IPAM), UCLA

This workshop will bring together leading experts in the mathematical modeling of collective intelligence. While the workshop will explore collective intelligence in many different systems, substrates (both natural and artificial), and scales, it will focus on three broadly applicable modeling frameworks: dynamical systems, which have been applied to systems as various as brains and insect swarms; statistical physics, which has illuminated the behavior of flocking birds and deep learning architectures; and game theory, which has been applied to animal, human, and AI collectives. The collective intelligence domain cultivates the further development of formal tools for treating compositionality; the emergence of new capacities and collective degrees of freedom at multiple scales; and the higher-level institutions that coordinate the intelligent behavior of lower-level parts.

Part of the Long Program Mathematics of Intelligences