Ahmed Elhag, a PhD student at the University of Oxford, presented a new training procedure that approximates equivariance in unconstrained machine learning models via a multitask objective. The approach adds an equivariance loss to unconstrained models, allowing them to learn approximate symmetries without the computational cost of fully equivariant methods. Formulating equivariance as a flexible learning objective allows control over the extent of symmetry enforced, matching the performance of strictly equivariant baselines at a lower cost. Why it matters: This research from a speaker at MBZUAI balances rigorous theory and practical scalability in geometric deep learning, potentially accelerating drug discovery and design.
This article discusses domain shift in machine learning, where testing data differs from training data, and methods to mitigate it via domain adaptation and generalization. Domain adaptation uses labeled source data and unlabeled target data. Domain generalization uses labeled data from single or multiple source domains to generalize to unseen target domains. Why it matters: Research in mitigating domain shift enhances the robustness and applicability of AI models in diverse real-world scenarios.
A DeepMind researcher presented work on incorporating symmetries into machine learning models, with applications to lattice-QCD and molecular dynamics. The work includes permutation and translation-invariant normalizing flows for free-energy estimation in molecular dynamics. They also presented U(N) and SU(N) Gauge-equivariant normalizing flows for pure Gauge simulations and its extensions to incorporate fermions in lattice-QCD. Why it matters: Applying symmetry principles to generative models could improve AI's ability to model complex physical systems relevant to materials science and other fields in the region.
Bruno Ribeiro from Purdue University presented a talk on Asymmetry Learning and Out-of-Distribution (OOD) Robustness. The talk introduced Asymmetry Learning, a new paradigm that focuses on finding evidence of asymmetries in data to improve classifier performance in both in-distribution and out-of-distribution scenarios. Asymmetry Learning performs a causal structure search to find classifiers that perform well across different environments. Why it matters: This research addresses a key challenge in AI by proposing a novel approach to improve the reliability and generalization of classifiers in unseen environments, potentially leading to more robust AI systems.
The paper introduces the Prism Hypothesis, which posits a correspondence between an encoder's feature spectrum and its functional role, with semantic encoders capturing low-frequency components and pixel encoders retaining high-frequency information. Based on this, the authors propose Unified Autoencoding (UAE), a model that harmonizes semantic structure and pixel details using a frequency-band modulator. Experiments on ImageNet and MS-COCO demonstrate that UAE effectively unifies semantic abstraction and pixel-level fidelity, achieving state-of-the-art performance.
Gregory Chirikjian presented an overview talk on applying probability, harmonic analysis, and geometry to robotics, emphasizing the need for robots to function beyond traditional industrial programming. He discussed a new approach where robots define affordances of objects, using simulation to 'imagine' object use and enabling reasoning about novel objects. Probabilistic methods on Lie-groups, initially developed for mobile robot state estimation, are now adapted for one-shot learning of affordances, with plans to integrate large language models. Why it matters: This research direction aims to enhance robot intelligence and adaptability, crucial for service robots in dynamic environments and aligning with broader goals of advanced AI integration in robotics.
This paper introduces a self-supervised learning method for point cloud analysis using an upsampling autoencoder (UAE). The model uses subsampling and an encoder-decoder architecture to reconstruct the original point cloud, learning both semantic and geometric information. Experiments show the UAE outperforms existing methods in shape classification, part segmentation, and point cloud upsampling tasks.