This paper presents incremental network quantization (INQ), a novel method, targeting to efficiently convert any pre-trained full-precision convolutional neural network (CNN) model into a low-precision version whose weights are constrained to be either powers of two or zero. Unlike existing methods which are struggled in noticeable accuracy loss, our INQ has the potential to resolve this issue, as benefiting from two innovations. On one hand, we introduce three interdependent operations, namely weight partition, group-wise quantization and re-training. A well-proven measure is employed to divide the weights in each layer of a pre-trained CNN model into two disjoint groups. The weights in the first group are responsible to form a low-precision base, thus they are quantized by a variable-length encoding method. The weights in the other group are responsible to compensate for the accuracy loss from the quantization, thus they are the ones to be re-trained. On the other hand, these three operations are repeated on the latest re-trained group in an iterative manner until all the weights are converted into low-precision ones, acting as an incremental network quantization and accuracy enhancement procedure. Extensive experiments on the ImageNet classification task using almost all known deep CNN architectures including AlexNet, VGG-16, GoogleNet and ResNets well testify the efficacy of the proposed method. Specifically, at 5-bit quantization, our models have improved accuracy than the 32-bit floating-point references. Taking ResNet-18 as an example, we further show that our quantized models with 4-bit, 3-bit and 2-bit ternary weights have improved or very similar accuracy against its 32-bit floating-point baseline. Besides, impressive results with the combination of network pruning and INQ are also reported. The code is available at https://github.com/Zhouaojun/Incremental-Network-Quantization.

Boosting is a learning scheme that combines weak prediction rules to produce a strong composite estimator, with the underlying intuition that one can obtain accurate prediction rules by combining "rough" ones. Although boosting is proved to be consistent and overfitting-resistant, its numerical convergence rate is relatively slow. The aim of this paper is to develop a new boosting strategy, called the re-scale boosting (RBoosting), to accelerate the numerical convergence rate and, consequently, improve the learning performance of boosting. Our studies show that RBoosting possesses the almost optimal numerical convergence rate in the sense that, up to a logarithmic factor, it can reach the minimax nonlinear approximation rate. We then use RBoosting to tackle both the classification and regression problems, and deduce a tight generalization error estimate. The theoretical and experimental results show that RBoosting outperforms boosting in terms of generalization.

Perhaps surprisingly, it is possible to predict how long an algorithm will take to run on a previously unseen input, using machine learning techniques to build a model of the algorithm's runtime as a function of problem-specific instance features. Such models have important applications to algorithm analysis, portfolio-based algorithm selection, and the automatic configuration of parameterized algorithms. Over the past decade, a wide variety of techniques have been studied for building such models. Here, we describe extensions and improvements of existing models, new families of models, and -- perhaps most importantly -- a much more thorough treatment of algorithm parameters as model inputs. We also comprehensively describe new and existing features for predicting algorithm runtime for propositional satisfiability (SAT), travelling salesperson (TSP) and mixed integer programming (MIP) problems. We evaluate these innovations through the largest empirical analysis of its kind, comparing to a wide range of runtime modelling techniques from the literature. Our experiments consider 11 algorithms and 35 instance distributions; they also span a very wide range of SAT, MIP, and TSP instances, with the least structured having been generated uniformly at random and the most structured having emerged from real industrial applications. Overall, we demonstrate that our new models yield substantially better runtime predictions than previous approaches in terms of their generalization to new problem instances, to new algorithms from a parameterized space, and to both simultaneously.

Uniform random 3-SAT at the solubility phase transition is one of the most widely studied and empirically hardest distributions of SAT instances. For 20 years, this distribution has been used extensively for evaluating and comparing algorithms. In this work, we demonstrate that simple rules can predict the solubility of these instances with surprisingly high accuracy. Specifically, we show how classification accuracies of about 70% can be obtained based on cheaply (polynomial-time) computable features on a wide range of instance sizes. We argue in two ways that classification accuracy does not decrease with instance size: first, we show that our models' predictive accuracy remains roughly constant across a wide range of problem sizes; second, we show that a classifier trained on small instances is sufficient to achieve very accurate predictions across the entire range of instance sizes currently solvable by complete methods. Finally, we demonstrate that a simple decision tree based on only two features, and again trained only on the smallest instances, achieves predictive accuracies close to those of our most complex model. We conjecture that this two-feature model outperforms random guessing asymptotically; due to the model's extreme simplicity, we believe that this conjecture is a worthwhile direction for future theoretical work.

The AI community has achieved great success in designing high-performance algorithms for hard combinatorial problems, given both considerable domain knowledge and considerable effort by human experts. Two influential methods aim to automate this process: automated algorithm configuration and portfolio-based algorithm selection. The former has the advantage of requiring virtually no domain knowledge, but produces only a single solver; the latter exploits per-instance variation, but requires a set of relatively uncorrelated candidate solvers. Here, we introduce Hydra, a novel technique for combining these two methods, thereby realizing the benefits of both. Hydra automatically builds a set of solvers with complementary strengths by iteratively configuring new algorithms. It is primarily intended for use in problem domains for which an adequate set of candidate solvers does not already exist. Nevertheless, we tested Hydra on a widely studied domain, stochastic local search algorithms for SAT, in order to characterize its performance against a well-established and highly competitive baseline. We found that Hydra consistently achieved major improvements over the best existing individual algorithms, and always at least roughly matched — and indeed often exceeded — the performance of the best portfolios of these algorithms.