Hypersphere anchor loss for K-Nearest neighbors

Abstract

Learning effective feature spaces for KNN (K-Nearest Neighbor) classifiers is critical for their performance. Existing KNN loss functions designed to optimize CNNs in $\mathbb {R}^n$ feature spaces for specific KNN classifiers greatly boost the performance. However, these loss functions need to compute the pairwise distances within each batch, which requires large computational resource, GPU and memory. This paper aims to exploit lightweight KNN loss functions in order to reduce the computational cost while achieving comparable to or even better performance than existing KNN loss functions. To this end, an anchor loss function is proposed that assigns each category an anchor vector in KNN feature spaces and introduces the distances between training samples and anchor vectors in the NCA (Neighborhood Component Analysis) function. The proposed anchor loss function largely reduces the required computation by existing KNN loss functions. In addition, instead of optimizing CNNs in $\mathbb {R}^n$ feature spaces, this paper proposed to optimize them in hypersphere feature spaces for faster convergence and better performance. The proposed anchor loss optimized in the hypersphere feature space is called HAL (Hypersphere Anchor Loss). Experiments on various image classification benchmarks show that HAL reduces the computational cost and achieves better performance: on CIFAR-10 and Fashion-MNIST datasets, compared with existing KNN loss functions, HAL improves the accuracy by over $1\%$, and the computational cost decreases to less than $10\%$.