基于统计的机器学习推理优化方法

A Statistics-based Optimization Method for Machine Learning Inference

近年来,机器学习推理技术已经广泛应用在人们的日常生产和生活中,但现有技术通常使用面向传统数据服务的优化方法,而未考虑机器学习推理的统计学特征。本文设计了一种面向机器学习推理的优化方法,通过引入基于统计的优化策略以解决特征计算的性能瓶颈问题。首先,使用自动构造的近似模型丢弃得分较低的输入以进行准确的Top-K查询。然后,使用增强模型对剩余部分进行排序,以最小精度损失为代价提高查询性能。同时,自动调整优化参数,以最大化提高查询性能的同时满足准确性目标。此外,使用编译器优化来优化以上技术,从而为机器学习应用自动快速生成推理代码。

In recent years,the inference technology of machine learning has been widely used in daily production and life,but the existing technologies usually use the optimization method of traditional data services,regardless of the statistical characteristics of inference in machine learning.In this paper,a statistics-based optimization method for machine learning inference is designed to solve the performance bottleneck of feature calculation by introducing a statistics-based optimization strategy.First,the constructed approximation model is used to discard low-scoring inputs for an accurate Top-K query.Then,the enhancement model is used to sort the rest data to improve query performance at the cost of minimum precision loss.At the same time,the optimization parameters are automatically adjusted to maximize the query performance and meet the accuracy.In addition,compiler optimizations are used to complement the above techniques to automatically and rapidly generate inference codes for machine learning applications.