软等变 Transformer 训练后可输出对称性指纹
Symmetry Fingerprinting with Geometric Learnable Invariants Transformers
这篇论文把等变网络的选择难题交给训练过程自己解决:先圈一组对称群,网络自动挑出真正匹配数据的那些,还能读出指纹看结果。做过几何深度学习的工程师会很有共鸣,硬等变选错伤性能,增强又拖慢训练。
arXiv 论文提出 Geometric Learnable Invariants Transformers(GLIT),一种软等变神经网络架构。传统几何深度学习在强结构等变(硬归纳偏置)与低几何偏置(需数据增强、拉长训练)之间二选一,GLIT 选择了中间路线:训练者先选定一组候选对称群,网络在训练过程中自动只对与训练数据匹配的那些对称群变得等变。
训练完成后,可以直接检查网络权重,读出一个 symmetry fingerprint(对称性指纹),精确显示哪些对称性最终被网络采纳、采纳到什么程度。作者在图像分类任务上展示了该架构与指纹提取流程的完整工作方式。
Symmetry Fingerprinting with Geometric Learnable Invariants Transformers
Geometric deep learning has made available a large variety of structurally equivariant neural network architectures, meaning a particular transformation of their input causes a predictable transformation of their output. But structural equivariance is a very strong inductive bias that can limit performance if chosen poorly. In contrast, choosing an architecture with low geometric bias necessitates data augmentation to obtain correct behavior; extending the training process. We propose a novel neural network architecture that is soft equivariant; we select a set of symmetry groups and during training the network will become equivariant only to those symmetry groups that are appropriate given the training data. After training we can inspect the network to obtain a symmetry fingerprint that will tell us exactly which symmetries ended up being relevant and to what degree. We show how our architecture and the fingerprinting process work in the case of an image classification task.
This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include Geometric Deep Learning, Transformer, equivariance, image classification.
Source attribution: curated from arXiv cs.LG (@Bart M. N. Smets, Gijs Bellaard, Pieter Wils).