$BT^2$: Backward-compatible Training with Basis Transformation

TL;DR

This paper introduces $BT^2$, a novel basis transformation method that enhances backward-compatible training by efficiently adding dimensions, enabling models to update representations without costly backfilling while maintaining performance across various model evolutions.

Contribution

The paper proposes $BT^2$, a basis transformation approach that improves backward-compatible training by selectively adding dimensions, addressing limitations of naive dimension increase and supporting diverse model updates.

Findings

$BT^2$ outperforms state-of-the-art methods in various settings.

Adding dimensions via basis transformation preserves information and improves compatibility.

$BT^2$ adapts to model architecture changes, modality shifts, and multiple updates.

Abstract

Modern retrieval system often requires recomputing the representation of every piece of data in the gallery when updating to a better representation model. This process is known as backfilling and can be especially costly in the real world where the gallery often contains billions of samples. Recently, researchers have proposed the idea of Backward Compatible Training (BCT) where the new representation model can be trained with an auxiliary loss to make it backward compatible with the old representation. In this way, the new representation can be directly compared with the old representation, in principle avoiding the need for any backfilling. However, followup work shows that there is an inherent tradeoff where a backward compatible representation model cannot simultaneously maintain the performance of the new model itself. This paper reports our ``not-so-surprising'' finding that adding extra dimensions to the representation can help here. However, we also found that naively increasing the dimension of the representation did not work. To deal with this, we propose Backward-compatible Training with a novel Basis Transformation ($BT^2$). A basis transformation (BT) is basically a learnable set of parameters that applies an orthonormal transformation. Such a transformation possesses an important property whereby the original information contained in its input is retained in its output. We show in this paper how a BT can be utilized to add only the necessary amount of additional dimensions. We empirically verify the advantage of $BT^2$ over other state-of-the-art methods in a wide range of settings. We then further extend $BT^2$ to other challenging yet more practical settings, including significant change in model architecture (CNN to Transformers), modality change, and even a series of updates in the model architecture mimicking the evolution of deep learning models.