Introduction
This document serves as the editorial preface to the proceedings of two landmark 2022 workshops: the ICML Workshop on Topology, Algebra, and Geometry in Machine Learning (TAG:ML) and the ICLR Workshop on Geometrical and Topological Representation Learning (GTRL). Rather than presenting a single experimental result, this preface acts as an editorial manifesto for an emerging field that seeks to ground the empirical successes of deep learning in the rigorous language of advanced mathematics.
The Theoretical Gap in Modern AI
Despite the resounding success of deep learning in domains such as computer vision and natural language processing, these advancements often lack a fundamental theoretical foundation. The authors identify a "dire need" to move beyond the "black box" nature of current models.
Modern data is characterized by high dimensionality, structural complexity, and strong nonlinearity. This creates a significant gap between two distinct intellectual communities:
- The Machine Learning Community: Primarily data-driven and focused on empirical performance and benchmark optimization.
- The Mathematical Community: Possessors of over a century of "finely-developed machinery"—specifically topology, algebra, and geometry—designed to describe and analyze complex spaces.
The primary objective of these workshops was to bridge this divide, fostering a "cross-pollination" of ideas to understand why specific algorithms succeed and to identify the next generation of mathematical tools for representation learning.
A Taxonomy of Mathematical Approaches
While the preface does not introduce a new algorithm, it categorizes the methodologies championed across the two workshops as the pillars of this new field:
- Topological Data Analysis (TDA): Utilizing the "shape" of data, such as persistent homology, to extract robust features that are invariant to noise.
- Manifold Learning: Uncovering the low-dimensional intrinsic structures that exist within high-dimensional measurement spaces.
- Geometric Deep Learning: Extending neural network architectures to non-Euclidean domains, including graphs and manifolds.
- Equivariant Machine Learning: Constructing models that respect physical or mathematical symmetries through the lens of classical invariant theory.
- Optimal Transport: Leveraging geometric distances between probability distributions to inform and regularize learning processes.
Key Insights: Beyond the Benchmark
A central theme of the TAG:ML workshop was the encouragement of research that addresses critical, domain-specific challenges, even when such approaches do not immediately yield "improved performance on mainstream, data-rich benchmarks." This represents a philosophical shift from "SOTA-chasing" toward theoretical depth and structural understanding.
The authors describe mathematics as a "veritable cornucopia" of methods capable of describing everything from the boundaries of neural network architectures to more expressive models for graph learning. A key conceptual anchor is the "Magnificent Manifold" perspective—the idea that the intrinsic structure of data is often a manifold, and that identifying these structures is the key to the next generation of representation learning.
Workshop Outcomes and Community Building
The scale of these academic gatherings underscores the growing interest in this intersection:
- TAG:ML (ICML 2022): Received 44 submissions (33 accepted) and hosted over 125 in-person participants. Keynote topics included the "Memory of Persistence" and "Equivariant Machine Learning."
- GTRL (ICLR 2022): Received 54 submissions (31 accepted), utilizing a structure of "Foundation Talks" for broad overviews and "Invited Talks" for specific research contributions.
- Curation: From the 64 total accepted papers across both events, 34 were selected for inclusion in this formal volume based on their contribution to the field's theoretical rigor.
Beyond the workshops themselves, the preface highlights the Applied Algebraic Topology Research Network (AATRN) as a vital resource for researchers, providing global access to the tools and community necessary to advance this vibrant field.
Future Directions
The authors acknowledge that efforts to integrate the ML and mathematical communities remain in their infancy. The primary limitation is the historical disjointedness of these computational and theoretical approaches. Future work focuses on breaking down these remaining barriers, developing "mathematically-informed deep learning" that moves beyond empirical observation toward a predictive, structural science of artificial intelligence.
Summary updated