A Unified Batch Selection Policy for Active Metric Learning

Active metric learning is the problem of incrementally selecting high-utility batches training data (typically, ordered triplets) to annotate, in order progressively improve a learned model over some input domain as rapidly possible. Standard approaches, which independently assess informativeness each triplet batch, are susceptible highly correlated with many redundant triplets and hence low overall utility. While recent work [20] proposes batch-decorrelation strategies for learning, they rely on ad hoc heuristics estimate correlation between two at time. We present novel batch active method that leverages Maximum Entropy Principle learn least biased distribution given set prior constraints. To avoid redundancy triplets, our collectively selects maximum joint entropy, simultaneously captures both diversity. take advantage submodularity entropy function construct tractable solution using an efficient greedy algorithm based Gram-Schmidt orthogonalization provably \(\left( 1 - \frac{1}{e} \right) \)-optimal. Our approach first define unified score balances diversity entire triplets. Experiments several real-world datasets demonstrate robust, generalizes well different applications modalities, consistently outperforms state-of-the-art.