Skip to yearly menu bar Skip to main content

Workshop: Workshop on Distribution Shifts: New Frontiers with Foundation Models

TiC-CLIP: Continual Training of CLIP Models

Saurabh Garg · Mehrdad Farajtabar · Hadi Pouransari · Raviteja Vemulapalli · Sachin Mehta · Oncel Tuzel · Vaishaal Shankar · Fartash Faghri

Keywords: [ continual learning ] [ benchmark ] [ Distribution Shift ] [ CLIP ]

[ ] [ Project Page ]
Fri 15 Dec 11:15 a.m. PST — 11:25 a.m. PST

Abstract: Keeping large foundation models up to date on latest data is inherently expensive. To avoid the prohibitive costs of constantly retraining, it is imperative to continually train these models. This problem is exacerbated by the lack of any large scale continual learning benchmarks or baselines. We introduce the first set of web-scale Time-Continual (TiC) benchmarks for training vision-language models: TIC-DataComp, TIC-YFCC, and TIC-RedCaps with over 12.7B timestamped image-text pairs spanning 9 years (2014--2022). We first use our benchmarks to curate various dynamic evaluations to measure temporal robustness of existing models. We show OpenAI’s CLIP (trained on data up to 2020) loses $\approx 8%$ zero-shot accuracy on our curated retrieval task from 2021--2022 compared with more recently trained models in OpenCLIP repository. We then study how to efficiently train models on time-continuous data. We demonstrate that a simple rehearsal-based approach that continues training from the last checkpoint and replays old data reduces compute by $2.5\times$ when compared to the standard practice of retraining from scratch.

Chat is not available.