State-of-the-art transformer models use pairwise dot-product based self-attention, which comes at a computational cost quadratic in the input sequence length. In this paper, we investigate the global structure of attention scores computed using this
dot product mechanism on a typical distribution of inputs, and study the principal components of their variation. Through eigen analysis of full attention score matrices, as well as of their individual rows, we find that most of the variation among attention scores lie in a low-dimensional eigenspace. Moreover, we find significant overlap between these eigenspaces for different layers and even different transformer models. Based on this, we propose to compute scores only for a partial subset of token pairs, and use them to estimate scores for the remaining pairs. Beyond investigating the accuracy of reconstructing attention scores themselves, we investigate training transformer models that employ these approximations, and analyze the effect on overall accuracy. Our analysis and the proposed method provide insights into how to balance the benefits of exact pair-wise attention and its significant computational expense.
Deep Convolutional Neural Networks (CNNs) have long been the architecture of choice for computer vision tasks. Recently, Transformer-based architectures like Vision Transformer (ViT) have matched or even surpassed ResNets for image classification. Ho
wever, details of the Transformer architecture -- such as the use of non-overlapping patches -- lead one to wonder whether these networks are as robust. In this paper, we perform an extensive study of a variety of different measures of robustness of ViT models and compare the findings to ResNet baselines. We investigate robustness to input perturbations as well as robustness to model perturbations. We find that when pre-trained with a sufficient amount of data, ViT models are at least as robust as the ResNet counterparts on a broad range of perturbations. We also find that Transformers are robust to the removal of almost any single layer, and that while activations from later layers are highly correlated with each other, they nevertheless play an important role in classification.
To deploy machine learning-based algorithms for real-time applications with strict latency constraints, we consider an edge-computing setting where a subset of inputs are offloaded to the edge for processing by an accurate but resource-intensive mode
l, and the rest are processed only by a less-accurate model on the device itself. Both models have computational costs that match available compute resources, and process inputs with low-latency. But offloading incurs network delays, and to manage these delays to meet application deadlines, we use a token bucket to constrain the average rate and burst length of transmissions from the device. We introduce a Markov Decision Process-based framework to make offload decisions under these constraints, based on the local models confidence and the token bucket state, with the goal of minimizing a specified error measure for the application. Beyond isolated decisions for individual devices, we also propose approaches to allow multiple devices connected to the same access switch to share their bursting allocation. We evaluate and analyze the policies derived using our framework on the standard ImageNet image classification benchmark.
Recent works propose neural network- (NN-) inspired analog-to-digital converters (NNADCs) and demonstrate their great potentials in many emerging applications. These NNADCs often rely on resistive random-access memory (RRAM) devices to realize the NN
operations and require high-precision RRAM cells (6~12-bit) to achieve a moderate quantization resolution (4~8-bit). Such optimistic assumption of RRAM resolution, however, is not supported by fabrication data of RRAM arrays in large-scale production process. In this paper, we propose an NN-inspired super-resolution ADC based on low-precision RRAM devices by taking the advantage of a co-design methodology that combines a pipelined hardware architecture with a custom NN training framework. Results obtained from SPICE simulations demonstrate that our method leads to robust design of a 14-bit super-resolution ADC using 3-bit RRAM devices with improved power and speed performance and competitive figure-of-merits (FoMs). In addition to the linear uniform quantization, the proposed ADC can also support configurable high-resolution nonlinear quantization with high conversion speed and low conversion energy, enabling future intelligent analog-to-information interfaces for near-sensor analytics and processing.
Deep neural networks have been very successful in image estimation applications such as compressive-sensing and image restoration, as a means to estimate images from partial, blurry, or otherwise degraded measurements. These networks are trained on a
large number of corresponding pairs of measurements and ground-truth images, and thus implicitly learn to exploit domain-specific image statistics. But unlike measurement data, it is often expensive or impractical to collect a large training set of ground-truth images in many application settings. In this paper, we introduce an unsupervised framework for training image estimation networks, from a training set that contains only measurements---with two varied measurements per image---but no ground-truth for the full images desired as output. We demonstrate that our framework can be applied for both regular and blind image estimation tasks, where in the latter case parameters of the measurement model (e.g., the blur kernel) are unknown: during inference, and potentially, also during training. We evaluate our method for training networks for compressive-sensing and blind deconvolution, considering both non-blind and blind training for the latter. Our unsupervised framework yields models that are nearly as accurate as those from fully supervised training, despite not having access to any ground-truth images.
Beyond depth estimation from a single image, the monocular cue is useful in a broader range of depth inference applications and settings---such as when one can leverage other available depth cues for improved accuracy. Currently, different applicatio
ns, with different inference tasks and combinations of depth cues, are solved via different specialized networks---trained separately for each application. Instead, we propose a versatile task-agnostic monocular model that outputs a probability distribution over scene depth given an input color image, as a sample approximation of outputs from a patch-wise conditional VAE. We show that this distributional output can be used to enable a variety of inference tasks in different settings, without needing to retrain for each application. Across a diverse set of applications (depth completion, user guided estimation, etc.), our common model yields results with high accuracy---comparable to or surpassing that of state-of-the-art methods dependent on application-specific networks.
Modern neural network-based algorithms are able to produce highly accurate depth estimates from stereo image pairs, nearly matching the reliability of measurements from more expensive depth sensors. However, this accuracy comes with a higher computat
ional cost since these methods use network architectures designed to compute and process matching scores across all candidate matches at all locations, with floating point computations repeated across a match volume with dimensions corresponding to both space and disparity. This leads to longer running times to process each image pair, making them impractical for real-time use in robots and autonomous vehicles. We propose a new stereo algorithm that employs a significantly more efficient network architecture. Our method builds an initial match cost volume using traditional matching costs that are fast to compute, and trains a network to estimate disparity from this volume. Crucially, our network only employs per-pixel and two-dimensional convolution operations: to summarize the match information at each location as a low-dimensional feature vector, and to spatially process these `cost-signature features to produce a dense disparity map. Experimental results on the KITTI benchmark show that our method delivers competitive accuracy at significantly higher speeds---running at 48 frames per second on a modern GPU.
Training convolutional neural network models is memory intensive since back-propagation requires storing activations of all intermediate layers. This presents a practical concern when seeking to deploy very deep architectures in production, especiall
y when models need to be frequently re-trained on updated datasets. In this paper, we propose a new implementation for back-propagation that significantly reduces memory usage, by enabling the use of approximations with negligible computational cost and minimal effect on training performance. The algorithm reuses common buffers to temporarily store full activations and compute the forward pass exactly. It also stores approximate per-layer copies of activations, at significant memory savings, that are used in the backward pass. Compared to simply approximating activations within standard back-propagation, our method limits accumulation of errors across layers. This allows the use of much lower-precision approximations without affecting training accuracy. Experiments on CIFAR-10, CIFAR-100, and ImageNet show that our method yields performance close to exact training, while storing activations compactly with as low as 4-bit precision.
Effective theory of fluctuations based on underlying symmetry plays very important role in understanding the low energy phenomena. Using this powerful technique we study the fluctuation dynamics keeping in mind the following central question: does th
e effective theory of black hole provide any information about the possible existence of hair? Assuming the symmetry of the hair being that of the underlying black hole space-time, we start by writing down the most general action for the background and the fluctuation in the effective field theory framework. Considering the asymptotically flat and de Sitter black hole background with a spherically symmetric hair we derived the most general equation of motion for the fluctuation. For a particular choice of theory parameters, quasinormal modes corresponding to those fluctuations appeared to have distinct features compared to that of the usual black hole quasinormal modes. The background equations from the effective theory Lagrangian, on the other hand, seemed to suggest that the underlying theory of the hair under consideration should be higher derivative in nature. Therefore as a concrete example we construct a class of higher derivative scalar field theory which gives rise to spherically symmetric hair through background cosmological constant. We also calculate the quasinormal modes whose behaviour turned out to be similar to the one discussed from the effective theory.
Image denoising methods must effectively model, implicitly or explicitly, the vast diversity of patterns and textures that occur in natural images. This is challenging, even for modern methods that leverage deep neural networks trained to regress to
clean images from noisy inputs. One recourse is to rely on internal image statistics, by searching for similar patterns within the input image itself. In this work, we propose a new method for natural image denoising that trains a deep neural network to determine whether patches in a noisy image input share common underlying patterns. Given a pair of noisy patches, our network predicts whether different sub-band coefficients of the original noise-free patches are similar. The denoising algorithm then aggregates matched coefficients to obtain an initial estimate of the clean image. Finally, this estimate is provided as input, along with the original noisy image, to a standard regression-based denoising network. Experiments show that our method achieves state-of-the-art color image denoising performance, including with a blind version that trains a common model for a range of noise levels, and does not require knowledge of level of noise in an input image. Our approach also has a distinct advantage when training with limited amounts of training data.
