اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدImproving Model Training by Periodic Sampling over Weight Distributions
195
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
In this paper, we explore techniques centered around periodic sampling of model weights that provide convergence improvements on gradient update methods (vanilla acs{SGD}, Momentum, Adam) for a variety of vision problems (classification, detection, segmentation). Importantly, our algorithms provide better, faster and more robust convergence and training performance with only a slight increase in computation time. Our techniques are independent of the neural network model, gradient optimization methods or existing optimal training policies and converge in a less volatile fashion with performance improvements that are approximately monotonic. We conduct a variety of experiments to quantify these improvements and identify scenarios where these techniques could be more useful.
قيم البحث
اقرأ أيضاً
Generative adversarial networks (GAN) have shown remarkable results in image generation tasks. High fidelity class-conditional GAN methods often rely on stabilization techniques by constraining the global Lipschitz continuity. Such regularization lea
ds to less expressive models and slower convergence speed; other techniques, such as the large batch training, require unconventional computing power and are not widely accessible. In this paper, we develop an efficient algorithm, namely FastGAN (Free AdverSarial Training), to improve the speed and quality of GAN training based on the adversarial training technique. We benchmark our method on CIFAR10, a subset of ImageNet, and the full ImageNet datasets. We choose strong baselines such as SNGAN and SAGAN; the results demonstrate that our training algorithm can achieve better generation quality (in terms of the Inception score and Frechet Inception distance) with less overall training time. Most notably, our training algorithm brings ImageNet training to the broader public by requiring 2-4 GPUs.
The inductive bias of a neural network is largely determined by the architecture and the training algorithm. To achieve good generalization, how to effectively train a neural network is of great importance. We propose a novel orthogonal over-paramete
rized training (OPT) framework that can provably minimize the hyperspherical energy which characterizes the diversity of neurons on a hypersphere. By maintaining the minimum hyperspherical energy during training, OPT can greatly improve the empirical generalization. Specifically, OPT fixes the randomly initialized weights of the neurons and learns an orthogonal transformation that applies to these neurons. We consider multiple ways to learn such an orthogonal transformation, including unrolling orthogonalization algorithms, applying orthogonal parameterization, and designing orthogonality-preserving gradient descent. For better scalability, we propose the stochastic OPT which performs orthogonal transformation stochastically for partial dimensions of neurons. Interestingly, OPT reveals that learning a proper coordinate system for neurons is crucial to generalization. We provide some insights on why OPT yields better generalization. Extensive experiments validate the superiority of OPT over the standard training.
Neural networks are commonly used as models for classification for a wide variety of tasks. Typically, a learned affine transformation is placed at the end of such models, yielding a per-class value used for classification. This classifier can have a
vast number of parameters, which grows linearly with the number of possible classes, thus requiring increasingly more resources. In this work we argue that this classifier can be fixed, up to a global scale constant, with little or no loss of accuracy for most tasks, allowing memory and computational benefits. Moreover, we show that by initializing the classifier with a Hadamard matrix we can speed up inference as well. We discuss the implications for current understanding of neural network models.
Sparse deep neural networks have shown their advantages over dense models with fewer parameters and higher computational efficiency. Here we demonstrate constraining the synaptic weights on unit Lp-sphere enables the flexibly control of the sparsity
with p and improves the generalization ability of neural networks. Firstly, to optimize the synaptic weights constrained on unit Lp-sphere, the parameter optimization algorithm, Lp-spherical gradient descent (LpSGD) is derived from the augmented Empirical Risk Minimization condition, which is theoretically proved to be convergent. To understand the mechanism of how p affects Hoyers sparsity, the expectation of Hoyers sparsity under the hypothesis of gamma distribution is given and the predictions are verified at various p under different conditions. In addition, the semi-pruning and threshold adaptation are designed for topology evolution to effectively screen out important connections and lead the neural networks converge from the initial sparsity to the expected sparsity. Our approach is validated by experiments on benchmark datasets covering a wide range of domains. And the theoretical analysis pave the way to future works on training sparse neural networks with constrained optimization.
Sparsification is an efficient approach to accelerate CNN inference, but it is challenging to take advantage of sparsity in training procedure because the involved gradients are dynamically changed. Actually, an important observation shows that most
of the activation gradients in back-propagation are very close to zero and only have a tiny impact on weight-updating. Hence, we consider pruning these very small gradients randomly to accelerate CNN training according to the statistical distribution of activation gradients. Meanwhile, we theoretically analyze the impact of pruning algorithm on the convergence. The proposed approach is evaluated on AlexNet and ResNet-{18, 34, 50, 101, 152} with CIFAR-{10, 100} and ImageNet datasets. Experimental results show that our training approach could substantially achieve up to $5.92 times$ speedups at back-propagation stage with negligible accuracy loss.
الأسئلة المقترحة
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
