VariedVision: Employing Adaptive Patching Mechanisms in Vision Transformer Models
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Vision Transformers; Adaptive patching mechanisms; Multi-resolution information;
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Leveraging self-attention mechanisms, Vision Transformers (ViTs) have recently achieved unprecedented success across diverse vision tasks. However, they encounter significant computational overheads, exacerbated by the increasing number of patches, attention heads, and transformer blocks. The predominant uniform patching strategy in ViTs overlooks the disparate detail levels within different image regions, leading to inefficiencies. Addressing these constraints, we introduce VariedVision, a groundbreaking framework that incorporates adaptive patching mechanisms, allowing the model to tap into multi-resolution information by allocating variable patch sizes to different regions based on their intrinsic details. This enables more refined attention to areas requiring finer granularity and more comprehensive coverage for areas with lesser details. VariedVision employs an auxiliary network to dynamically assign 'detail scores' to various image regions, thereby determining the optimal patch size for each, ensuring uniformity in the number of patches while enhancing adaptability to diverse long-range dependencies between them. The integration of dynamic positional embeddings caters to the varying patch sizes and multi-resolution nature of the input, enriching the model's spatial and contextual awareness. This multi-resolution approach enables VariedVision to perceive and process details at different scales effectively, optimizing the representation of both macro and micro-level features within the image. VariedVision is optimized end-to-end, harmonizing the adaptive patching mechanism with the transformer backbone, unveiling an enhancement in both efficiency and accuracy. Rigorous experiments on ImageNet validate VariedVision's efficacy, showcasing enhanced efficiency/accuracy trade-offs and revealing insights into potential redundancies and inefficiencies in conventional uniform patching methodologies.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4669
Loading