Vision Transformers convert images to sequences by slicing them into patches. The size of these patches controls a speed/accuracy tradeoff, with smaller patches leading to higher accuracy at greater computational cost, but changing the patch size typically requires retraining the model. In this paper, we demonstrate that simply randomizing the patch size at training time leads to a single set of weights that performs well across a wide range of patch sizes, making it possible to tailor the model to different compute budgets at deployment time. We extensively evaluate the resulting model, which we call FlexiViT, on a wide range of tasks, including classification, image-text retrieval, open-world detection, panoptic segmentation, and semantic segmentation, concluding that it usually matches, and sometimes outperforms, standard ViT models trained at a single patch size in an otherwise identical setup. Hence, FlexiViT training is a simple drop-in improvement for ViT that makes it easy to add compute-adaptive capabilities to most models relying on a ViT backbone architecture. Code and pre-trained models are available at https://github.com/google-research/big_vision
translated by 谷歌翻译
对于神经网络的近似贝叶斯推断被认为是标准培训的强大替代品，通常在分发数据上提供良好的性能。然而，贝叶斯神经网络（BNNS）具有高保真近似推断的全批汉密尔顿蒙特卡罗在协变速下实现了较差的普遍，甚至表现不佳的经典估算。我们解释了这种令人惊讶的结果，展示了贝叶斯模型平均值实际上如何存在于协变量的情况下，特别是在输入特征中的线性依赖性导致缺乏后退的情况下。我们还展示了为什么相同的问题不会影响许多近似推理程序，或古典最大A-Bouthiori（地图）培训。最后，我们提出了改善BNN的鲁棒性的新型前锋，对许多协变量转变来源。
translated by 谷歌翻译
知识蒸馏是一种培训小型学生网络的流行技术，以模仿更大的教师模型，例如网络的集合。我们表明，虽然知识蒸馏可以改善学生泛化，但它通常不得如此普遍地工作：虽然在教师和学生的预测分布之间，甚至在学生容量的情况下，通常仍然存在令人惊讶的差异完美地匹配老师。我们认为优化的困难是为什么学生无法与老师匹配的关键原因。我们还展示了用于蒸馏的数据集的细节如何在学生与老师匹配的紧密关系中发挥作用 - 以及教师矛盾的教师并不总是导致更好的学生泛化。
translated by 谷歌翻译
We propose SWA-Gaussian (SWAG), a simple, scalable, and general purpose approach for uncertainty representation and calibration in deep learning. Stochastic Weight Averaging (SWA), which computes the first moment of stochastic gradient descent (SGD) iterates with a modified learning rate schedule, has recently been shown to improve generalization in deep learning. With SWAG, we fit a Gaussian using the SWA solution as the first moment and a low rank plus diagonal covariance also derived from the SGD iterates, forming an approximate posterior distribution over neural network weights; we then sample from this Gaussian distribution to perform Bayesian model averaging. We empirically find that SWAG approximates the shape of the true posterior, in accordance with results describing the stationary distribution of SGD iterates. Moreover, we demonstrate that SWAG performs well on a wide variety of tasks, including out of sample detection, calibration, and transfer learning, in comparison to many popular alternatives including MC dropout, KFAC Laplace, SGLD, and temperature scaling.
translated by 谷歌翻译
Deep neural networks are typically trained by optimizing a loss function with an SGD variant, in conjunction with a decaying learning rate, until convergence. We show that simple averaging of multiple points along the trajectory of SGD, with a cyclical or constant learning rate, leads to better generalization than conventional training. We also show that this Stochastic Weight Averaging (SWA) procedure finds much flatter solutions than SGD, and approximates the recent Fast Geometric Ensembling (FGE) approach with a single model. Using SWA we achieve notable improvement in test accuracy over conventional SGD training on a range of state-of-the-art residual networks, PyramidNets, DenseNets, and Shake-Shake networks on CIFAR-10, CIFAR-100, and ImageNet. In short, SWA is extremely easy to implement, improves generalization, and has almost no computational overhead.
translated by 谷歌翻译
The loss functions of deep neural networks are complex and their geometric properties are not well understood. We show that the optima of these complex loss functions are in fact connected by simple curves over which training and test accuracy are nearly constant. We introduce a training procedure to discover these high-accuracy pathways between modes. Inspired by this new geometric insight, we also propose a new ensembling method entitled Fast Geometric Ensembling (FGE). Using FGE we can train high-performing ensembles in the time required to train a single model. We achieve improved performance compared to the recent state-of-the-art Snapshot Ensembles, on CIFAR-10, CIFAR-100, and ImageNet. * Equal contribution. 1 Suppose we have three weight vectors w1, w2, w3. We set u = (w2 − w1), v = (w3 − w1) − w3 − w1, w2 − w1 / w2 − w1 2 • (w2 − w1). Then the normalized vectors û = u/ u , v = v/ v form an orthonormal basis in the plane containing w1, w2, w3. To visualize the loss in this plane, we define a Cartesian grid in the basis û, v and evaluate the networks corresponding to each of the points in the grid. A point P with coordinates (x, y) in the plane would then be given by P = w1 + x • û + y • v.
translated by 谷歌翻译
In this paper, we propose a new neural network architecture based on the H2 matrix. Even though networks with H2-inspired architecture already exist, and our approach is designed to reduce memory costs and improve performance by taking into account the sparsity template of the H2 matrix. In numerical comparison with alternative neural networks, including the known H2-based ones, our architecture showed itself as beneficial in terms of performance, memory, and scalability.
translated by 谷歌翻译
t-SNE remains one of the most popular embedding techniques for visualizing high-dimensional data. Most standard packages of t-SNE, such as scikit-learn, use the Barnes-Hut t-SNE (BH t-SNE) algorithm for large datasets. However, existing CPU implementations of this algorithm are inefficient. In this work, we accelerate the BH t-SNE on CPUs via cache optimizations, SIMD, parallelizing sequential steps, and improving parallelization of multithreaded steps. Our implementation (Acc-t-SNE) is up to 261x and 4x faster than scikit-learn and the state-of-the-art BH t-SNE implementation from daal4py, respectively, on a 32-core Intel(R) Icelake cloud instance.
translated by 谷歌翻译
We investigate a model for image/video quality assessment based on building a set of codevectors representing in a sense some basic properties of images, similar to well-known CORNIA model. We analyze the codebook building method and propose some modifications for it. Also the algorithm is investigated from the point of inference time reduction. Both natural and synthetic images are used for building codebooks and some analysis of synthetic images used for codebooks is provided. It is demonstrated the results on quality assessment may be improves with the use if synthetic images for codebook construction. We also demonstrate regimes of the algorithm in which real time execution on CPU is possible for sufficiently high correlations with mean opinion score (MOS). Various pooling strategies are considered as well as the problem of metric sensitivity to bitrate.
translated by 谷歌翻译
Online controlled experiments (A/B tests) have become the gold standard for learning the impact of new product features in technology companies. Randomization enables the inference of causality from an A/B test. The randomized assignment maps end users to experiment buckets and balances user characteristics between the groups. Therefore, experiments can attribute any outcome differences between the experiment groups to the product feature under experiment. Technology companies run A/B tests at scale -- hundreds if not thousands of A/B tests concurrently, each with millions of users. The large scale poses unique challenges to randomization. First, the randomized assignment must be fast since the experiment service receives hundreds of thousands of queries per second. Second, the variant assignments must be independent between experiments. Third, the assignment must be consistent when users revisit or an experiment enrolls more users. We present a novel assignment algorithm and statistical tests to validate the randomized assignments. Our results demonstrate that not only is this algorithm computationally fast but also satisfies the statistical requirements -- unbiased and independent.
translated by 谷歌翻译