Accuracy at the top is a special class of binary classification problems where the performance is evaluated only on a small number of relevant (top) samples. Applications include information retrieval systems or processes with manual (expensive) postprocessing. This leads to minimizing the number of irrelevant samples above a threshold. We consider classifiers in the form of an arbitrary (deep) network and propose a new method DeepTopPush for minimizing the loss function at the top. Since the threshold depends on all samples, the problem is non-decomposable. We modify the stochastic gradient descent to handle the non-decomposability in an end-to-end training manner and propose a way to estimate the threshold only from values on the current minibatch and one delayed value. We demonstrate the excellent performance of DeepTopPush on visual recognition datasets and two real-world applications. The first one selects a small number of molecules for further drug testing. The second one uses real malware data, where we detected 46\% malware at an extremely low false alarm rate of $10^{-5}$.
翻译:高精度问题是一类特殊的二元分类问题,其中性能仅在少数相关(前)样本上进行评估。应用包括信息检索系统或手动(昂贵)后处理过程。这导致最小化阈值上方的所有无关样本。我们考虑采用任意(深度)网络的分类器,并提出一种新的方法DeepTopPush来最小化顶部的损失函数。由于阈值依赖于所有样本,因此问题是不可分解的。我们修改了随机梯度下降法以处理非可分解性,并提出了一种仅从当前小批量值和一个延迟值估计阈值的方法。我们在视觉识别数据集和两个实际应用程序上展示了DeepTopPush的出色性能。第一个应用程序选择少量分子进行进一步的药物测试。第二个应用程序使用真实恶意软件数据,我们在极低的误警率($10^{-5}$)下检测到46%的恶意软件。