Batch Decorrelation for Active Metric Learning

Abstract

<jats:p>We present an active learning strategy for training parametric models of distance metrics, given triplet-based similarity assessments: object $x_i$ is more similar to object $x_j$ than to $x_k$. In contrast to prior work on class-based learning, where the fundamental goal is classification and any implicit or explicit metric is binary, we focus on perceptual metrics that express the degree of (dis)similarity between objects. We find that standard active learning approaches degrade when annotations are requested for batches of triplets at a time: our studies suggest that correlation among triplets is responsible. In this work, we propose a novel method to decorrelate batches of triplets, that jointly balances informativeness and diversity while decoupling the choice of heuristic for each criterion. Experiments indicate our method is general, adaptable, and outperforms the state-of-the-art.</jats:p>

Active metric learning and classification using similarity queries

“…Furthermore, there seems to be minimal performance difference in selecting random NN queries vs. random triplets on a per-triplet basis, but using NN queries requires far fewer interactions with the oracle. From these experiments, it appears that the methods in [3] require more of a "warm up" to catch up to random query performance, whereas Info-NN can consistently outperform random. Inspecting the visualization in Fig.…”

“…Although class labels may not always be available, very few works consider the case of DML with perceptual similarity queries, especially in an active manner. Recently, active similarity query selection methods for DML that focus on finding batches of non-redundant triplets have been proposed [3] by encouraging both informativeness (measured by entropy) and diversity (through a variety of heuristic approaches) within the selected batch. Our method adopts a similar framework as [3], but we utilize mutual information to find informative NN queries.…”

“…In scenarios where re-training the network many times is computationally expensive, batch methods that select multiple queries at once are preferable. We compare the performance of Info-NN to recently developed triplet batch methods [3]. While Info-NN can identify informative queries, batches of the most informative and TopFraction@21 (center).…”

“…In our experiments, Info-NN-C means the batch variant of Info-NN described above was used to select NN queries of length C, while Batch-Euclidean/Centroid indicate methods proposed in [3]. Finally, Random means the query type (NN or triplet) was constructed by selecting queries uniformly at random from the training set.…”

“…In the case of metric learning, the response is utilized to place the reference closer to the similar item while for classification, the response is equivalent to a label corresponding to the reference (e.g., cake in this case). similarity information, as is done in [3], where triplets of the form "Is item B or item C more similar to item A?" are actively selected for learning an embedding of items.…”