Joint embedding (JE) architectures have emerged as a promising avenue for buying transferable knowledge representations. A key impediment to utilizing JE strategies, nonetheless, is the inherent problem of evaluating realized representations with out entry to a downstream activity, and an annotated dataset. With out environment friendly and dependable analysis, it’s troublesome to iterate on architectural and coaching decisions for JE strategies. On this paper, we introduce LiDAR (Linear Discriminant Evaluation Rank), a metric designed to measure the standard of representations inside JE architectures. Our metric addresses a number of shortcomings of latest approaches primarily based on characteristic covariance rank by discriminating between informative and uninformative options. In essence, LiDAR quantifies the rank of the Linear Discriminant Evaluation (LDA) matrix related to the surrogate SSL activity—a measure that intuitively captures the data content material because it pertains to fixing the SSL activity. We empirically exhibit that LiDAR considerably surpasses naive rank primarily based approaches in its predictive energy of optimum hyperparameters. Our proposed criterion presents a extra strong and intuitive technique of assessing the standard of representations inside JE architectures, which we hope facilitates broader adoption of those highly effective methods in numerous domains.