The Transformer architecture has revolutionized deep learning on sequential data, becoming ubiquitous in state-of-the-art solutions for a wide variety of applications. Yet vanilla Transformers are notoriously resource-expensive, requiring $O(L^2)$ in serial time and memory as functions of input length $L$. Recent works proposed various linear self-attention mechanisms, scaling only as $O(L)$ for serial computation. We perform a thorough analysis of recent Transformer mechanisms with linear self-attention, Performers, in terms of overall computational complexity. We observe a remarkable computational flexibility: forward and backward propagation can be performed with no approximations using sublinear memory as a function of $L$ (in addition to negligible storage for the input sequence), at a cost of greater time complexity in the parallel setting. In the extreme case, a Performer consumes only $O(1)$ memory during training, and still requires $O(L)$ time. This discovered time-memory tradeoff can be used for training or, due to complete backward-compatibility, for fine-tuning on a low-memory device, e.g. a smartphone or an earlier-generation GPU, thus contributing towards decentralized and democratized deep learning.
Currently, Machine Learning (ML) is becoming ubiquitous in everyday life. Deep Learning (DL) is already present in many applications ranging from computer vision for medicine to autonomous driving of modern cars as well as other sectors in security, healthcare, and finance. However, to achieve impressive performance, these algorithms employ very deep networks, requiring a significant computational power, both during the training and inference time. A single inference of a DL model may require billions of multiply-and-accumulated operations, making the DL extremely compute- and energy-hungry. In a scenario where several sophisticated algorithms need to be executed with limited energy and low latency, the need for cost-effective hardware platforms capable of implementing energy-efficient DL execution arises. This paper first introduces the key properties of two brain-inspired models like Deep Neural Network (DNN), and Spiking Neural Network (SNN), and then analyzes techniques to produce efficient and high-performance designs. This work summarizes and compares the works for four leading platforms for the execution of algorithms such as CPU, GPU, FPGA and ASIC describing the main solutions of the state-of-the-art, giving much prominence to the last two solutions since they offer greater design flexibility and bear the potential of high energy-efficiency, especially for the inference process. In addition to hardware solutions, this paper discusses some of the important security issues that these DNN and SNN models may have during their execution, and offers a comprehensive section on benchmarking, explaining how to assess the quality of different networks and hardware systems designed for them.
Deep Reinforcement Learning reaches a superhuman level of play in many complete information games. The state of the art algorithm for learning with zero knowledge is AlphaZero. We take another approach, Ath\'enan, which uses a different, Minimax-based, search algorithm called Descent, as well as different learning targets and that does not use a policy. We show that for multiple games it is much more efficient than the reimplementation of AlphaZero: Polygames. It is even competitive with Polygames when Polygames uses 100 times more GPU (at least for some games). One of the keys to the superior performance is that the cost of generating state data for training is approximately 296 times lower with Ath\'enan. With the same reasonable ressources, Ath\'enan without reinforcement heuristic is at least 7 times faster than Polygames and much more than 30 times faster with reinforcement heuristic.
Video analytics applications use edge compute servers for the analytics of the videos (for bandwidth and privacy). Compressed models that are deployed on the edge servers for inference suffer from data drift, where the live video data diverges from the training data. Continuous learning handles data drift by periodically retraining the models on new data. Our work addresses the challenge of jointly supporting inference and retraining tasks on edge servers, which requires navigating the fundamental tradeoff between the retrained model's accuracy and the inference accuracy. Our solution Ekya balances this tradeoff across multiple models and uses a micro-profiler to identify the models that will benefit the most by retraining. Ekya's accuracy gain compared to a baseline scheduler is 29% higher, and the baseline requires 4x more GPU resources to achieve the same accuracy as Ekya.
Along with the development of AI democratization, the machine learning approach, in particular neural networks, has been applied to wide-range applications. In different application scenarios, the neural network will be accelerated on the tailored computing platform. The acceleration of neural networks on classical computing platforms, such as CPU, GPU, FPGA, ASIC, has been widely studied; however, when the scale of the application consistently grows up, the memory bottleneck becomes obvious, widely known as memory-wall. In response to such a challenge, advanced quantum computing, which can represent 2^N states with N quantum bits (qubits), is regarded as a promising solution. It is imminent to know how to design the quantum circuit for accelerating neural networks. Most recently, there are initial works studying how to map neural networks to actual quantum processors. To better understand the state-of-the-art design and inspire new design methodology, this paper carries out a case study to demonstrate an end-to-end implementation. On the neural network side, we employ the multilayer perceptron to complete image classification tasks using the standard and widely used MNIST dataset. On the quantum computing side, we target IBM Quantum processors, which can be programmed and simulated by using IBM Qiskit. This work targets the acceleration of the inference phase of a trained neural network on the quantum processor. Along with the case study, we will demonstrate the typical procedure for mapping neural networks to quantum circuits.
In this work, we present a lightweight, tightly-coupled deep depth network and visual-inertial odometry (VIO) system, which can provide accurate state estimates and dense depth maps of the immediate surroundings. Leveraging the proposed lightweight Conditional Variational Autoencoder (CVAE) for depth inference and encoding, we provide the network with previously marginalized sparse features from VIO to increase the accuracy of initial depth prediction and generalization capability. The compact encoded depth maps are then updated jointly with navigation states in a sliding window estimator in order to provide the dense local scene geometry. We additionally propose a novel method to obtain the CVAE's Jacobian which is shown to be more than an order of magnitude faster than previous works, and we additionally leverage First-Estimate Jacobian (FEJ) to avoid recalculation. As opposed to previous works relying on completely dense residuals, we propose to only provide sparse measurements to update the depth code and show through careful experimentation that our choice of sparse measurements and FEJs can still significantly improve the estimated depth maps. Our full system also exhibits state-of-the-art pose estimation accuracy, and we show that it can run in real-time with single-thread execution while utilizing GPU acceleration only for the network and code Jacobian.
In this paper, we aim at establishing accurate dense correspondences between a pair of images with overlapping field of view under challenging illumination variation, viewpoint changes, and style differences. Through an extensive ablation study of the state-of-the-art correspondence networks, we surprisingly discovered that the widely adopted 4D correlation tensor and its related learning and processing modules could be de-parameterised and removed from training with merely a minor impact over the final matching accuracy. Disabling these computational expensive modules dramatically speeds up the training procedure and allows to use 4 times bigger batch size, which in turn compensates for the accuracy drop. Together with a multi-GPU inference stage, our method facilitates the systematic investigation of the relationship between matching accuracy and up-sampling resolution of the native testing images from 1280 to 4K. This leads to discovery of the existence of an optimal resolution $\mathbb{X}$ that produces accurate matching performance surpassing the state-of-the-art methods particularly over the lower error band on public benchmarks for the proposed network.
Recent works on learned image compression perform encoding and decoding processes in a full-resolution manner, resulting in two problems when deployed for practical applications. First, parallel acceleration of the autoregressive entropy model cannot be achieved due to serial decoding. Second, full-resolution inference often causes the out-of-memory(OOM) problem with limited GPU resources, especially for high-resolution images. Block partition is a good design choice to handle the above issues, but it brings about new challenges in reducing the redundancy between blocks and eliminating block effects. To tackle the above challenges, this paper provides a learned block-based hybrid image compression (LBHIC) framework. Specifically, we introduce explicit intra prediction into a learned image compression framework to utilize the relation among adjacent blocks. Superior to context modeling by linear weighting of neighbor pixels in traditional codecs, we propose a contextual prediction module (CPM) to better capture long-range correlations by utilizing the strip pooling to extract the most relevant information in neighboring latent space, thus achieving effective information prediction. Moreover, to alleviate blocking artifacts, we further propose a boundary-aware postprocessing module (BPM) with the edge importance taken into account. Extensive experiments demonstrate that the proposed LBHIC codec outperforms the VVC, with a bit-rate conservation of 4.1%, and reduces the decoding time by approximately 86.7% compared with that of state-of-the-art learned image compression methods.
This work presents FG-Net, a general deep learning framework for large-scale point clouds understanding without voxelizations, which achieves accurate and real-time performance with a single NVIDIA GTX 1080 GPU. First, a novel noise and outlier filtering method is designed to facilitate subsequent high-level tasks. For effective understanding purpose, we propose a deep convolutional neural network leveraging correlated feature mining and deformable convolution based geometric-aware modelling, in which the local feature relationships and geometric patterns can be fully exploited. For the efficiency issue, we put forward an inverse density sampling operation and a feature pyramid based residual learning strategy to save the computational cost and memory consumption respectively. Extensive experiments on real-world challenging datasets demonstrated that our approaches outperform state-of-the-art approaches in terms of accuracy and efficiency. Moreover, weakly supervised transfer learning is also conducted to demonstrate the generalization capacity of our method.
The attention mechanism is becoming increasingly popular in Natural Language Processing (NLP) applications, showing superior performance than convolutional and recurrent architectures. However, attention becomes the compution bottleneck because of its quadratic computational complexity to input length, complicated data movement and low arithmetic intensity. Moreover, existing NN accelerators mainly focus on optimizing convolutional or recurrent models, and cannot efficiently support attention. In this paper, we present SpAtten, an efficient algorithm-architecture co-design that leverages token sparsity, head sparsity, and quantization opportunities to reduce the attention computation and memory access. Inspired by the high redundancy of human languages, we propose the novel cascade token pruning to prune away unimportant tokens in the sentence. We also propose cascade head pruning to remove unessential heads. Cascade pruning is fundamentally different from weight pruning since there is no trainable weight in the attention mechanism, and the pruned tokens and heads are selected on the fly. To efficiently support them on hardware, we design a novel top-k engine to rank token and head importance scores with high throughput. Furthermore, we propose progressive quantization that first fetches MSBs only and performs the computation; if the confidence is low, it fetches LSBs and recomputes the attention outputs, trading computation for memory reduction. Extensive experiments on 30 benchmarks show that, on average, SpAtten reduces DRAM access by 10.0x with no accuracy loss, and achieves 1.6x, 3.0x, 162x, 347x speedup, and 1,4x, 3.2x, 1193x, 4059x energy savings over A3 accelerator, MNNFast accelerator, TITAN Xp GPU, Xeon CPU, respectively.
Deep neural networks have recently drawn considerable attention to build and evaluate artificial learning models for perceptual tasks. Here, we present a study on the performance of the deep learning models to deal with global optimization problems. The proposed approach adopts the idea of the neural architecture search (NAS) to generate efficient neural networks for solving the problem at hand. The space of network architectures is represented using a directed acyclic graph and the goal is to find the best architecture to optimize the objective function for a new, previously unknown task. Different from proposing very large networks with GPU computational burden and long training time, we focus on searching for lightweight implementations to find the best architecture. The performance of NAS is first analyzed through empirical experiments on CEC 2017 benchmark suite. Thereafter, it is applied to a set of protein structure prediction (PSP) problems. The experiments reveal that the generated learning models can achieve competitive results when compared to hand-designed algorithms; given enough computational budget
Current state-of-the-art Neural Architecture Search (NAS) methods neither efficiently scale to multiple hardware platforms, nor handle diverse architectural search-spaces. To remedy this, we present DONNA (Distilling Optimal Neural Network Architectures), a novel pipeline for rapid, scalable and diverse NAS, that scales to many user scenarios. DONNA consists of three phases. First, an accuracy predictor is built using blockwise knowledge distillation from a reference model. This predictor enables searching across diverse networks with varying macro-architectural parameters such as layer types and attention mechanisms, as well as across micro-architectural parameters such as block repeats and expansion rates. Second, a rapid evolutionary search finds a set of pareto-optimal architectures for any scenario using the accuracy predictor and on-device measurements. Third, optimal models are quickly finetuned to training-from-scratch accuracy. DONNA is up to 100x faster than MNasNet in finding state-of-the-art architectures on-device. Classifying ImageNet, DONNA architectures are 20% faster than EfficientNet-B0 and MobileNetV2 on a Nvidia V100 GPU and 10% faster with 0.5% higher accuracy than MobileNetV2-1.4x on a Samsung S20 smartphone. In addition to NAS, DONNA is used for search-space extension and exploration, as well as hardware-aware model compression.
The spatially-varying field of the human visual system has recently received a resurgence of interest with the development of virtual reality (VR) and neural networks. The computational demands of high resolution rendering desired for VR can be offset by savings in the periphery, while neural networks trained with foveated input have shown perceptual gains in i.i.d and o.o.d generalization. In this paper, we present a technique that exploits the CUDA GPU architecture to efficiently generate Gaussian-based foveated images at high definition (1920x1080 px) in real-time (165 Hz), with a larger number of pooling regions than previous Gaussian-based foveation algorithms by several orders of magnitude, producing a smoothly foveated image that requires no further blending or stitching, and that can be well fit for any contrast sensitivity function. The approach described can be adapted from Gaussian blurring to any eccentricity-dependent image processing and our algorithm can meet demand for experimentation to evaluate the role of spatially-varying processing across biological and artificial agents, so that foveation can be added easily on top of existing systems rather than forcing their redesign (emulated foveated renderer). Altogether, this paper demonstrates how a GPU, with a CUDA block-wise architecture, can be employed for radially-variant rendering, with opportunities for more complex post-processing to ensure a metameric foveation scheme. Code is provided.
Leveraging spatial sparsity has become a popular approach to accelerate 3D computer graphics applications. Spatially sparse data structures and efficient sparse kernels (such as parallel stencil operations on active voxels), are key to achieve high performance. Existing work focuses on improving performance within a single sparse computational kernel. We show that a system that looks beyond a single kernel, plus additional domain-specific sparse data structure analysis, opens up exciting new space for optimizing sparse computations. Specifically, we propose a domain-specific data-flow graph model of imperative and sparse computation programs, which describes kernel relationships and enables easy analysis and optimization. Combined with an asynchronous execution engine that exposes a wide window of kernels, the inter-kernel optimizer can then perform effective sparse computation optimizations, such as eliminating unnecessary voxel list generations and removing voxel activation checks. These domain-specific optimizations further make way for classical general-purpose optimizations that are originally challenging to directly apply to computations with sparse data structures. Without any computational code modification, our new system leads to $4.02\times$ fewer kernel launches and $1.87\times$ speed up on our GPU benchmarks, including computations on Eulerian grids, Lagrangian particles, meshes, and automatic differentiation.
We present SimNet, an AI-driven multi-physics simulation framework, to accelerate simulations across a wide range of disciplines in science and engineering. Compared to traditional numerical solvers, SimNet addresses a wide range of use cases - coupled forward simulations without any training data, inverse and data assimilation problems. SimNet offers fast turnaround time by enabling parameterized system representation that solves for multiple configurations simultaneously, as opposed to the traditional solvers that solve for one configuration at a time. SimNet is integrated with parameterized constructive solid geometry as well as STL modules to generate point clouds. Furthermore, it is customizable with APIs that enable user extensions to geometry, physics and network architecture. It has advanced network architectures that are optimized for high-performance GPU computing, and offers scalable performance for multi-GPU and multi-Node implementation with accelerated linear algebra as well as FP32, FP64 and TF32 computations. In this paper we review the neural network solver methodology, the SimNet architecture, and the various features that are needed for effective solution of the PDEs. We present real-world use cases that range from challenging forward multi-physics simulations with turbulence and complex 3D geometries, to industrial design optimization and inverse problems that are not addressed efficiently by the traditional solvers. Extensive comparisons of SimNet results with open source and commercial solvers show good correlation.
Monte Carlo tree search (MCTS) has achieved state-of-the-art results in many domains such as Go and Atari games when combining with deep neural networks (DNNs). When more simulations are executed, MCTS can achieve higher performance but also requires enormous amounts of CPU and GPU resources. However, not all states require a long searching time to identify the best action that the agent can find. For example, in 19x19 Go and NoGo, we found that for more than half of the states, the best action predicted by DNN remains unchanged even after searching 2 minutes. This implies that a significant amount of resources can be saved if we are able to stop the searching earlier when we are confident with the current searching result. In this paper, we propose to achieve this goal by predicting the uncertainty of the current searching status and use the result to decide whether we should stop searching. With our algorithm, called Dynamic Simulation MCTS (DS-MCTS), we can speed up a NoGo agent trained by AlphaZero 2.5 times faster while maintaining a similar winning rate. Also, under the same average simulation count, our method can achieve a 61% winning rate against the original program.
We present PDFFlow, a new software for fast evaluation of parton distribution functions (PDFs) designed for platforms with hardware accelerators. PDFs are essential for the calculation of particle physics observables through Monte Carlo simulation techniques. The evaluation of a generic set of PDFs for quarks and gluons at a given momentum fraction and energy scale requires the implementation of interpolation algorithms as introduced for the first time by the LHAPDF project. PDFFlow extends and implements these interpolation algorithms using Google's TensorFlow library providing the possibility to perform PDF evaluations taking fully advantage of multi-threading CPU and GPU setups. We benchmark the performance of this library on multiple scenarios relevant for the particle physics community.
Recently, Neural Ordinary Differential Equations has emerged as a powerful framework for modeling physical simulations without explicitly defining the ODEs governing the system, but instead learning them via machine learning. However, the question: "Can Bayesian learning frameworks be integrated with Neural ODE's to robustly quantify the uncertainty in the weights of a Neural ODE?" remains unanswered. In an effort to address this question, we primarily evaluate the following categories of inference methods: (a) The No-U-Turn MCMC sampler (NUTS), (b) Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) and (c) Stochastic Langevin Gradient Descent (SGLD). We demonstrate the successful integration of Neural ODEs with the above Bayesian inference frameworks on classical physical systems, as well as on standard machine learning datasets like MNIST, using GPU acceleration. On the MNIST dataset, we achieve a posterior sample accuracy of 98.5% on the test ensemble of 10,000 images. Subsequently, for the first time, we demonstrate the successful integration of variational inference with normalizing flows and Neural ODEs, leading to a powerful Bayesian Neural ODE object. Finally, considering a predator-prey model and an epidemiological system, we demonstrate the probabilistic identification of model specification in partially-described dynamical systems using universal ordinary differential equations. Together, this gives a scientific machine learning tool for probabilistic estimation of epistemic uncertainties.
The analysis of source code through machine learning techniques is an increasingly explored research topic aiming at increasing smartness in the software toolchain to exploit modern architectures in the best possible way. In the case of low-power, parallel embedded architectures, this means finding the configuration, for instance in terms of the number of cores, leading to minimum energy consumption. Depending on the kernel to be executed, the energy optimal scaling configuration is not trivial. While recent work has focused on general-purpose systems to learn and predict the best execution target in terms of the execution time of a snippet of code or kernel (e.g. offload OpenCL kernel on multicore CPU or GPU), in this work we focus on static compile-time features to assess if they can be successfully used to predict the minimum energy configuration on PULP, an ultra-low-power architecture featuring an on-chip cluster of RISC-V processors. Experiments show that using machine learning models on the source code to select the best energy scaling configuration automatically is viable and has the potential to be used in the context of automatic system configuration for energy minimisation.
We introduce a real-time, high-resolution background replacement technique which operates at 30fps in 4K resolution, and 60fps for HD on a modern GPU. Our technique is based on background matting, where an additional frame of the background is captured and used in recovering the alpha matte and the foreground layer. The main challenge is to compute a high-quality alpha matte, preserving strand-level hair details, while processing high-resolution images in real-time. To achieve this goal, we employ two neural networks; a base network computes a low-resolution result which is refined by a second network operating at high-resolution on selective patches. We introduce two largescale video and image matting datasets: VideoMatte240K and PhotoMatte13K/85. Our approach yields higher quality results compared to the previous state-of-the-art in background matting, while simultaneously yielding a dramatic boost in both speed and resolution.