Random Decision Stumps for Kernel Learning and Efficient SVM Gemma Roig * Xavier Boix * Luc Van Gool Computer Vision Lab, ETH Zurich, Switzerland {boxavier,gemmar,vangool}@vision.ee.ethz.ch * Both first authors contributed equally. Abstract. We propose to learn the kernel of an SVM as the weighted sum of a large number of simple, randomized binary stumps. Each stump takes one of the extracted features as input. This leads to an efficient and very fast SVM, while also alleviating the task of kernel selection. We demonstrate the capabilities of our kernel on 6 standard vision benchmarks, in which we combine several com- mon image descriptors, namely histograms (Flowers17 and Daimler), attribute- like descriptors (UCI, OSR, and a-VOC08), and Sparse Quantization (ImageNet). Results show that our kernel learning adapts well to these different feature types, achieving the performance of kernels specifically tuned for each, and with an evaluation cost similar to that of efficient SVM methods. 1 Introduction The success of Support Vector Machines (SVMs), e.g. in object recognition, stems from their well-studied optimization and their use of kernels to solve non-linear classification problems.Designing the right kernel in combination with appropriate image descriptors is crucial. Their joint design leads to a chicken-and-egg problem in that the right kernel depends on the image descriptors, while the image descriptors are designed for familiar kernels. Multiple Kernel Learning (MKL) [1] eases kernel selection by automatically learn- ing it as a combination of given base kernels. Although MKL has been successful in var- ious vision tasks (e.g. [2,3]), it might lead to complex and inefficient kernels. Recently, Bazavan et al. [4] introduced an approach to MKL that avoids the explicit computation of the kernel. It efficiently approximates the non-linear mapping of the hand-selected kernels [5,6,7], thus delivering impressive speed-ups. We propose another way around kernel learning that also allows for efficient SVMs. Instead of combining fixed base kernels, we investigate the use of random binary map- pings (BMs). We coin our approach Multiple Binary Kernel Learning (MBKL). Given that other methods based on binary decisions such as Random Forests [8] and Boosting decision stumps [9] have not performed equally well on image classification bench- marks as kernel SVMs, it is all the more important that we will show MBKL does. Not only does MBKL alleviate the task of selecting the right kernel, but the resulting kernel is very efficient to compute and can scale to large datasets. At the end of the paper, we report on MBKL results for 6 computer vision bench- marks, in which we combine several common image descriptors. These descriptors arXiv:1307.5161v2 [cs.CV] 28 Mar 2014