Transformers have been successful for many natural language processing tasks. However, applying transformers to the video domain for tasks such as long-term video generation and scene understanding has remained elusive due to the high computational complexity and the lack of natural tokenization. In this paper, we propose the Object-Centric Video Transformer (OCVT) which utilizes an object-centric approach for decomposing scenes into tokens suitable for use in a generative video transformer. By factoring the video into objects, our fully unsupervised model is able to learn complex spatio-temporal dynamics of multiple interacting objects in a scene and generate future frames of the video. Our model is also significantly more memory-efficient than pixel-based models and thus able to train on videos of length up to 70 frames with a single 48GB GPU. We compare our model with previous RNN-based approaches as well as other possible video transformer baselines. We demonstrate OCVT performs well when compared to baselines in generating future frames. OCVT also develops useful representations for video reasoning, achieving start-of-the-art performance on the CATER task.
The Deep Neural Networks are vulnerable toadversarial exam-ples(Figure 1), making the DNNs-based systems collapsed byadding the inconspicuous perturbations to the images. Most of the existing works for adversarial attack are gradient-based and suf-fer from the latency efficiencies and the load on GPU memory. Thegenerative-based adversarial attacks can get rid of this limitation,and some relative works propose the approaches based on GAN.However, suffering from the difficulty of the convergence of train-ing a GAN, the adversarial examples have either bad attack abilityor bad visual quality. In this work, we find that the discriminatorcould be not necessary for generative-based adversarial attack, andpropose theSymmetric Saliency-based Auto-Encoder (SSAE)to generate the perturbations, which is composed of the saliencymap module and the angle-norm disentanglement of the featuresmodule. The advantage of our proposed method lies in that it is notdepending on discriminator, and uses the generative saliency map to pay more attention to label-relevant regions. The extensive exper-iments among the various tasks, datasets, and models demonstratethat the adversarial examples generated by SSAE not only make thewidely-used models collapse, but also achieves good visual quality.The code is available at https://github.com/BravoLu/SSAE.
The Transformer architecture is widely used for machine translation tasks. However, its resource-intensive nature makes it challenging to implement on constrained embedded devices, particularly where available hardware resources can vary at run-time. We propose a dynamic machine translation model that scales the Transformer architecture based on the available resources at any particular time. The proposed approach, 'Dynamic-HAT', uses a HAT SuperTransformer as the backbone to search for SubTransformers with different accuracy-latency trade-offs at design time. The optimal SubTransformers are sampled from the SuperTransformer at run-time, depending on latency constraints. The Dynamic-HAT is tested on the Jetson Nano and the approach uses inherited SubTransformers sampled directly from the SuperTransformer with a switching time of <1s. Using inherited SubTransformers results in a BLEU score loss of <1.5% because the SubTransformer configuration is not retrained from scratch after sampling. However, to recover this loss in performance, the dimensions of the design space can be reduced to tailor it to a family of target hardware. The new reduced design space results in a BLEU score increase of approximately 1% for sub-optimal models from the original design space, with a wide range for performance scaling between 0.356s - 1.526s for the GPU and 2.9s - 7.31s for the CPU.
We present Megaverse, a new 3D simulation platform for reinforcement learning and embodied AI research. The efficient design of our engine enables physics-based simulation with high-dimensional egocentric observations at more than 1,000,000 actions per second on a single 8-GPU node. Megaverse is up to 70x faster than DeepMind Lab in fully-shaded 3D scenes with interactive objects. We achieve this high simulation performance by leveraging batched simulation, thereby taking full advantage of the massive parallelism of modern GPUs. We use Megaverse to build a new benchmark that consists of several single-agent and multi-agent tasks covering a variety of cognitive challenges. We evaluate model-free RL on this benchmark to provide baselines and facilitate future research. The source code is available at https://www.megaverse.info
In this paper, we develop face.evoLVe -- a comprehensive library that collects and implements a wide range of popular deep learning-based methods for face recognition. First of all, face.evoLVe is composed of key components that cover the full process of face analytics, including face alignment, data processing, various backbones, losses, and alternatives with bags of tricks for improving performance. Later, face.evoLVe supports multi-GPU training on top of different deep learning platforms, such as PyTorch and PaddlePaddle, which facilitates researchers to work on both large-scale datasets with millions of images and low-shot counterparts with limited well-annotated data. More importantly, along with face.evoLVe, images before & after alignment in the common benchmark datasets are released with source codes and trained models provided. All these efforts lower the technical burdens in reproducing the existing methods for comparison, while users of our library could focus on developing advanced approaches more efficiently. Last but not least, face.evoLVe is well designed and vibrantly evolving, so that new face recognition approaches can be easily plugged into our framework. Note that we have used face.evoLVe to participate in a number of face recognition competitions and secured the first place. The version that supports PyTorch is publicly available at https://github.com/ZhaoJ9014/face.evoLVe.PyTorch and the PaddlePaddle version is available at https://github.com/ZhaoJ9014/face.evoLVe.PyTorch/tree/master/paddle. Face.evoLVe has been widely used for face analytics, receiving 2.4K stars and 622 forks.
Evolution-based neural architecture search requires high computational resources, resulting in long search time. In this work, we propose a framework of applying the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to the neural architecture search problem called CMANAS, which achieves better results than previous evolution-based methods while reducing the search time significantly. The architectures are modelled using a normal distribution, which is updated using CMA-ES based on the fitness of the sampled population. We used the accuracy of a trained one shot model (OSM) on the validation data as a prediction of the fitness of an individual architecture to reduce the search time. We also used an architecture-fitness table (AF table) for keeping record of the already evaluated architecture, thus further reducing the search time. CMANAS finished the architecture search on CIFAR-10 with the top-1 test accuracy of 97.44% in 0.45 GPU day and on CIFAR-100 with the top-1 test accuracy of 83.24% for 0.6 GPU day on a single GPU. The top architectures from the searches on CIFAR-10 and CIFAR-100 were then transferred to ImageNet, achieving the top-5 accuracy of 92.6% and 92.1%, respectively.
Training large deep learning models at scale is very challenging. This paper proposes Chimera, a novel pipeline parallelism scheme which combines bidirectional pipelines for efficiently training large-scale models. Chimera is a synchronous approach and therefore no loss of accuracy, which is more convergence-friendly than asynchronous approaches. Compared with the latest synchronous pipeline approach, Chimera reduces the number of bubbles by up to 50%; benefiting from the sophisticated scheduling of bidirectional pipelines, Chimera has a more balanced activation memory consumption. Evaluations are conducted on Transformer based language models. For a GPT-2 model with 1.3 billion parameters running on 2,048 GPU nodes of the Piz Daint supercomputer, Chimera improves the training throughput by 1.16x-2.34x over the state-of-the-art synchronous and asynchronous pipeline approaches.
The search for specific objects or motifs is essential to art history as both assist in decoding the meaning of artworks. Digitization has produced large art collections, but manual methods prove to be insufficient to analyze them. In the following, we introduce an algorithm that allows users to search for image regions containing specific motifs or objects and find similar regions in an extensive dataset, helping art historians to analyze large digitized art collections. Computer vision has presented efficient methods for visual instance retrieval across photographs. However, applied to art collections, they reveal severe deficiencies because of diverse motifs and massive domain shifts induced by differences in techniques, materials, and styles. In this paper, we present a multi-style feature fusion approach that successfully reduces the domain gap and improves retrieval results without labelled data or curated image collections. Our region-based voting with GPU-accelerated approximate nearest-neighbour search allows us to find and localize even small motifs within an extensive dataset in a few seconds. We obtain state-of-the-art results on the Brueghel dataset and demonstrate its generalization to inhomogeneous collections with a large number of distractors.
Neural architecture search (NAS), an important branch of automatic machine learning, has become an effective approach to automate the design of deep learning models. However, the major issue in NAS is how to reduce the large search time imposed by the heavy computational burden. While most recent approaches focus on pruning redundant sets or developing new search methodologies, this paper attempts to formulate the problem based on the data curation manner. Our key strategy is to search the architecture using summarized data distribution, i.e., core-set. Typically, many NAS algorithms separate searching and training stages, and the proposed core-set methodology is only used in search stage, thus their performance degradation can be minimized. In our experiments, we were able to save overall computational time from 30.8 hours to 3.5 hours, 8.8x reduction, on a single RTX 3090 GPU without sacrificing accuracy.
Distributed training with synchronous stochastic gradient descent (SGD) on GPU clusters has been widely used to accelerate the training process of deep models. However, SGD only utilizes the first-order gradient in model parameter updates, which may take days or weeks. Recent studies have successfully exploited approximate second-order information to speed up the training process, in which the Kronecker-Factored Approximate Curvature (KFAC) emerges as one of the most efficient approximation algorithms for training deep models. Yet, when leveraging GPU clusters to train models with distributed KFAC (D-KFAC), it incurs extensive computation as well as introduces extra communications during each iteration. In this work, we propose D-KFAC (SPD-KFAC) with smart parallelism of computing and communication tasks to reduce the iteration time. Specifically, 1) we first characterize the performance bottlenecks of D-KFAC, 2) we design and implement a pipelining mechanism for Kronecker factors computation and communication with dynamic tensor fusion, and 3) we develop a load balancing placement for inverting multiple matrices on GPU clusters. We conduct real-world experiments on a 64-GPU cluster with 100Gb/s InfiniBand interconnect. Experimental results show that our proposed SPD-KFAC training scheme can achieve 10%-35% improvement over state-of-the-art algorithms.
Accurate capacitance extraction is becoming more important for designing integrated circuits under advanced process technology. The pattern matching based full-chip extraction methodology delivers fast computational speed, but suffers from large error, and tedious efforts on building capacitance models of the increasing structure patterns. In this work, we propose an effective method for building convolutional neural network (CNN) based capacitance models (called CNN-Cap) for two-dimensional (2-D) structures in full-chip capacitance extraction. With a novel grid-based data representation, the proposed method is able to model the pattern with a variable number of conductors, so that largely reduce the number of patterns. Based on the ability of ResNet architecture on capturing spatial information and the proposed training skills, the obtained CNN-Cap exhibits much better performance over the multilayer perception neural network based capacitance model while being more versatile. Extensive experiments on a 55nm and a 15nm process technologies have demonstrated that the error of total capacitance produced with CNN-Cap is always within 1.3% and the error of produced coupling capacitance is less than 10% in over 99.5% probability. CNN-Cap runs more than 4000X faster than 2-D field solver on a GPU server, while it consumes negligible memory compared to the look-up table based capacitance model.
We describe several additions to the ENIGMA system that guides clause selection in the E automated theorem prover. First, we significantly speed up its neural guidance by adding server-based GPU evaluation. The second addition is motivated by fast weight-based rejection filters that are currently used in systems like E and Prover9. Such systems can be made more intelligent by instead training fast versions of ENIGMA that implement more intelligent pre-filtering. This results in combinations of trainable fast and slow thinking that improves over both the fast-only and slow-only methods. The third addition is based on "judging the children by their parents", i.e., possibly rejecting an inference before it produces a clause. This is motivated by standard evolutionary mechanisms, where there is always a cost to producing all possible offsprings in the current population. This saves time by not evaluating all clauses by more expensive methods and provides a complementary view of the generated clauses. The methods are evaluated on a large benchmark coming from the Mizar Mathematical Library, showing good improvements over the state of the art.
This paper is both an introduction and an invitation. It is an introduction to CARLE, a Life-like cellular automata simulator and reinforcement learning environment. It is also an invitation to Carle's Game, a challenge in open-ended machine exploration and creativity. Inducing machine agents to excel at creating interesting patterns across multiple cellular automata universes is a substantial challenge, and approaching this challenge is likely to require contributions from the fields of artificial life, AI, machine learning, and complexity, at multiple levels of interest. Carle's Game is based on machine agent interaction with CARLE, a Cellular Automata Reinforcement Learning Environment. CARLE is flexible, capable of simulating any of the 262,144 different rules defining Life-like cellular automaton universes. CARLE is also fast and can simulate automata universes at a rate of tens of thousands of steps per second through a combination of vectorization and GPU acceleration. Finally, CARLE is simple. Compared to high-fidelity physics simulators and video games designed for human players, CARLE's two-dimensional grid world offers a discrete, deterministic, and atomic universal playground, despite its complexity. In combination with CARLE, Carle's Game offers an initial set of agent policies, learning and meta-learning algorithms, and reward wrappers that can be tailored to encourage exploration or specific tasks.
BayesSim is a statistical technique for domain randomization in reinforcement learning based on likelihood-free inference of simulation parameters. This paper outlines BayesSimIG: a library that provides an implementation of BayesSim integrated with the recently released NVIDIA IsaacGym. This combination allows large-scale parameter inference with end-to-end GPU acceleration. Both inference and simulation get GPU speedup, with support for running more than 10K parallel simulation environments for complex robotics tasks that can have more than 100 simulation parameters to estimate. BayesSimIG provides an integration with TensorBoard to easily visualize slices of high-dimensional posteriors. The library is built in a modular way to support research experiments with novel ways to collect and process the trajectories from the parallel IsaacGym environments.
Recently, artificial neural networks have been gaining momentum in the field of gravitational wave astronomy, for example in surrogate modelling of computationally expensive waveform models for binary black hole inspiral and merger. Surrogate modelling yields fast and accurate approximations of gravitational waves and neural networks have been used in the final step of interpolating the coefficients of the surrogate model for arbitrary waveforms outside the training sample. We investigate the existence of underlying structures in the empirical interpolation coefficients using autoencoders. We demonstrate that when the coefficient space is compressed to only two dimensions, a spiral structure appears, wherein the spiral angle is linearly related to the mass ratio. Based on this finding, we design a spiral module with learnable parameters, that is used as the first layer in a neural network, which learns to map the input space to the coefficients. The spiral module is evaluated on multiple neural network architectures and consistently achieves better speed-accuracy trade-off than baseline models. A thorough experimental study is conducted and the final result is a surrogate model which can evaluate millions of input parameters in a single forward pass in under 1ms on a desktop GPU, while the mismatch between the corresponding generated waveforms and the ground-truth waveforms is better than the compared baseline methods. We anticipate the existence of analogous underlying structures and corresponding computational gains also in the case of spinning black hole binaries.
Convolutional neural networks (CNNs) have been used quite successfully for semantic segmentation of brain tumors. However, current CNNs and attention mechanisms are stochastic in nature and neglect the morphological indicators used by radiologists to manually annotate regions of interest. In this paper, we introduce a channel and spatial wise asymmetric attention (CASPIAN) by leveraging the inherent structure of tumors to detect regions of saliency. To demonstrate the efficacy of our proposed layer, we integrate this into a well-established convolutional neural network (CNN) architecture to achieve higher Dice scores, with less GPU resources. Also, we investigate the inclusion of auxiliary multiscale and multiplanar attention branches to increase the spatial context crucial in semantic segmentation tasks. The resulting architecture is the new CASPIANET++, which achieves Dice Scores of 91.19% whole tumor, 87.6% for tumor core and 81.03% for enhancing tumor. Furthermore, driven by the scarcity of brain tumor data, we investigate the Noisy Student method for segmentation tasks. Our new Noisy Student Curriculum Learning paradigm, which infuses noise incrementally to increase the complexity of the training images exposed to the network, further boosts the enhancing tumor region to 81.53%. Additional validation performed on the BraTS2020 data shows that the Noisy Student Curriculum Learning method works well without any additional training or finetuning.
A two-step concept for 3D segmentation on 5 abdominal organs inside volumetric CT images is presented. First each relevant organ's volume of interest is extracted as bounding box. The extracted volume acts as input for a second stage, wherein two compared U-Nets with different architectural dimensions re-construct an organ segmentation as label mask. In this work, we focus on comparing 2D U-Nets vs. 3D U-Net counterparts. Our initial results indicate Dice improvements of about 6\% at maximum. In this study to our surprise, liver and kidneys for instance were tackled significantly better using the faster and GPU-memory saving 2D U-Nets. For other abdominal key organs, there were no significant differences, but we observe highly significant advantages for the 2D U-Net in terms of GPU computational efforts for all organs under study.
The fine-grained relationship between form and function with respect to deep neural network architecture design and hardware-specific acceleration is one area that is not well studied in the research literature, with form often dictated by accuracy as opposed to hardware function. In this study, a comprehensive empirical exploration is conducted to investigate the impact of deep neural network architecture design on the degree of inference speedup that can be achieved via hardware-specific acceleration. More specifically, we empirically study the impact of a variety of commonly used macro-architecture design patterns across different architectural depths through the lens of OpenVINO microprocessor-specific and GPU-specific acceleration. Experimental results showed that while leveraging hardware-specific acceleration achieved an average inference speed-up of 380%, the degree of inference speed-up varied drastically depending on the macro-architecture design pattern, with the greatest speedup achieved on the depthwise bottleneck convolution design pattern at 550%. Furthermore, we conduct an in-depth exploration of the correlation between FLOPs requirement, level 3 cache efficacy, and network latency with increasing architectural depth and width. Finally, we analyze the inference time reductions using hardware-specific acceleration when compared to native deep learning frameworks across a wide variety of hand-crafted deep convolutional neural network architecture designs as well as ones found via neural architecture search strategies. We found that the DARTS-derived architecture to benefit from the greatest improvement from hardware-specific software acceleration (1200%) while the depthwise bottleneck convolution-based MobileNet-V2 to have the lowest overall inference time of around 2.4 ms.
Gravitational waves are ripples in the fabric of space-time that travel at the speed of light. The detection of gravitational waves by LIGO is a major breakthrough in the field of astronomy. Deep Learning has revolutionized many industries including health care, finance and education. Deep Learning techniques have also been explored for detection of gravitational waves to overcome the drawbacks of traditional matched filtering method. However, in several researches, the training phase of neural network is very time consuming and hardware devices with large memory are required for the task. In order to reduce the extensive amount of hardware resources and time required in training a neural network for detecting gravitational waves, we made SpecGrav. We use 2D Convolutional Neural Network and spectrograms of gravitational waves embedded in noise to detect gravitational waves from binary black hole merger and binary neutron star merger. The training phase of our neural network was of about just 19 minutes on a 2GB GPU.
The performance of human pose estimation depends on the spatial accuracy of keypoint localization. Most existing methods pursue the spatial accuracy through learning the high-resolution (HR) representation from input images. By the experimental analysis, we find that the HR representation leads to a sharp increase of computational cost, while the accuracy improvement remains marginal compared with the low-resolution (LR) representation. In this paper, we propose a design paradigm for cost-effective network with LR representation for efficient pose estimation, named FasterPose. Whereas the LR design largely shrinks the model complexity, yet how to effectively train the network with respect to the spatial accuracy is a concomitant challenge. We study the training behavior of FasterPose, and formulate a novel regressive cross-entropy (RCE) loss function for accelerating the convergence and promoting the accuracy. The RCE loss generalizes the ordinary cross-entropy loss from the binary supervision to a continuous range, thus the training of pose estimation network is able to benefit from the sigmoid function. By doing so, the output heatmap can be inferred from the LR features without loss of spatial accuracy, while the computational cost and model size has been significantly reduced. Compared with the previously dominant network of pose estimation, our method reduces 58% of the FLOPs and simultaneously gains 1.3% improvement of accuracy. Extensive experiments show that FasterPose yields promising results on the common benchmarks, i.e., COCO and MPII, consistently validating the effectiveness and efficiency for practical utilization, especially the low-latency and low-energy-budget applications in the non-GPU scenarios.