Domino: Discovering Systematic Errors with Cross-Modal Embeddings

Abstract

Machine learning models that achieve high overall accuracy often make systematic errors on important subsets (or slices) of data. Identifying underperforming slices is particularly challenging when working with high-dimensional inputs (e.g. images, audio), where important slices are often unlabeled. In order to address this issue, recent studies have proposed automated slice discovery methods (SDMs), which leverage learned model representations to mine input data for slices on which a model performs poorly. To be useful to a practitioner, these methods must identify slices that are both underperforming and coherent (i.e. united by a human-understandable concept). However, no quantitative evaluation framework currently exists for rigorously assessing SDMs with respect to these criteria. Additionally, prior qualitative evaluations have shown that SDMs often identify slices that are incoherent. In this work, we address these challenges by first designing a principled evaluation framework that enables a quantitative comparison of SDMs across 1,235 slice discovery settings in three input domains (natural images, medical images, and time-series data). Then, motivated by the recent development of powerful cross-modal representation learning approaches, we present Domino, an SDM that leverages cross-modal embeddings and a novel error-aware mixture model to discover and describe coherent slices. We find that Domino accurately identifies 36% of the 1,235 slices in our framework -a 12 percentage point improvement over prior methods. Further, Domino is the first SDM that can provide natural language descriptions of identified slices, correctly generating the exact name of the slice in 35% of settings.

Distilling Model Failures as Directions in Latent Space

“…against the slice detection method Domino [EVS+22], and find that our framework more consistently selects this planted challenging subpopulation.…”

“…It is worth noting that, in contrast to previous works that choose the captions closest to the mean of a selected group of hard examples (c.f [EVS+22]), our method avoids this proxy and directly assigns a caption to the captured failure mode itself.…”

“…Moreover, the SVM selects the AP examples more consistently than ordering the radiographs by the base model's confidence (Figure 14). [EVS+22], which presents an automatic method for identifying challenging subsets ("slices") by fitting a Gaussian mixture model in CLIP embedding space. However, our use of an SVM provides a stronger inductive bias towards simple subclasses, and furthermore imposes a useful geometry on the latent space for use more easily scoring data points, intervening, and selecting captions in a more principled manner.…”

“…Domino [EVS+22] is an automatic method for identifying challenging "slices" by fitting a Gaussian mixture model in CLIP embedding space. They then choose captions generated from a large language model that are closest to the mean of each cluster.…”

Demystify Self-Attention in Vision Transformers from a Semantic Perspective: Analysis and Application

“…Some previous studies [9,23] have shown that traditional image classification tasks suffer from model performance degradation due to the spurious features hidden in images. For example, blue sky background is likely to co-occur in many bird images.…”