Current approaches for clinical information extraction are inefficient in terms of computational costs and memory consumption, hindering their application to process large-scale electronic health records (EHRs). We propose an efficient end-to-end model, the Joint-NER-RE-Fourier (JNRF), to jointly learn the tasks of named entity recognition and relation extraction for documents of variable length. The architecture uses positional encoding and unitary batch sizes to process variable length documents and uses a weight-shared Fourier network layer for low-complexity token mixing. Finally, we reach the theoretical computational complexity lower bound for relation extraction using a selective pooling strategy and distance-aware attention weights with trainable polynomial distance functions. We evaluated the JNRF architecture using the 2018 N2C2 ADE benchmark to jointly extract medication-related entities and relations in variable-length EHR summaries. JNRF outperforms rolling window BERT with selective pooling by 0.42%, while being twice as fast to train. Compared to state-of-the-art BiLSTM-CRF architectures on the N2C2 ADE benchmark, results show that the proposed approach trains 22 times faster and reduces GPU memory consumption by 1.75 folds, with a reasonable performance tradeoff of 90%, without the use of external tools, hand-crafted rules or post-processing. Given the significant carbon footprint of deep learning models and the current energy crises, these methods could support efficient and cleaner information extraction in EHRs and other types of large-scale document databases.
Batching has a fundamental influence on the efficiency of deep neural network (DNN) execution. However, for dynamic DNNs, efficient batching is particularly challenging as the dataflow graph varies per input instance. As a result, state-of-the-art frameworks use heuristics that result in suboptimal batching decisions. Further, batching puts strict restrictions on memory adjacency and can lead to high data movement costs. In this paper, we provide an approach for batching dynamic DNNs based on finite state machines, which enables the automatic discovery of batching policies specialized for each DNN via reinforcement learning. Moreover, we find that memory planning that is aware of the batching policy can save significant data movement overheads, which is automated by a PQ tree-based algorithm we introduce. Experimental results show that our framework speeds up state-of-the-art frameworks by on average 1.15x, 1.39x, and 2.45x for chain-based, tree-based, and lattice-based DNNs across CPU and GPU.
VAEs, or variational autoencoders, are autoencoders that explicitly learn the distribution of the input image space rather than assuming no prior information about the distribution. This allows it to classify similar samples close to each other in the latent space's distribution. VAEs classically assume the latent space is normally distributed, though many distribution priors work, and they encode this assumption through a K-L divergence term in the loss function. While VAEs learn the distribution of the latent space and naturally make each dimension in the latent space as disjoint from the others as possible, they do not group together similar features -- the image space feature represented by one unit of the representation layer does not necessarily have high correlation with the feature represented by a neighboring unit of the representation layer. This makes it difficult to interpret VAEs since the representation layer is not structured in a way that is easy for humans to parse. We aim to make a more interpretable VAE by partitioning the representation layer into disjoint sets of units. Partitioning the representation layer into disjoint sets of interconnected units yields a prior that features of the input space to this new VAE, which we call a partition VAE or PVAE, are grouped together by correlation -- for example, if our image space were the space of all ping ping game images (a somewhat complex image space we use to test our architecture) then we would hope the partitions in the representation layer each learned some large feature of the image like the characteristics of the ping pong table or the characteristics and position of the players or the ball. We also add to the PVAE a cost-saving measure: subresolution. Because we do not have access to GPU training environments for long periods of time and Google Colab Pro costs money, we attempt to decrease the complexity of the PVAE by outputting an image with dimensions scaled down from the input image by a constant factor, thus forcing the model to output a smaller version of the image. We then increase the resolution to calculate loss and train by interpolating through neighboring pixels. We train a tuned PVAE on MNIST and Sports10 to test its effectiveness.
We present an end-to-end binaural audio rendering approach (Listen2Scene) for virtual reality (VR) and augmented reality (AR) applications. We propose a novel neural-network-based binaural sound propagation method to generate acoustic effects for indoor 3D models of real environments. Any clean audio or dry audio can be convolved with the generated acoustic effects to render audio corresponding to the real environment. We propose a graph neural network that uses both the material and the topology information of the 3D scenes and generates a scene latent vector. Moreover, we use a conditional generative adversarial network (CGAN) to generate acoustic effects from the scene latent vector. Our network can handle holes or other artifacts in the reconstructed 3D mesh model. We present an efficient cost function for the generator network to incorporate spatial audio effects. Given the source and the listener position, our learning-based binaural sound propagation approach can generate an acoustic effect in 0.1 milliseconds on an NVIDIA GeForce RTX 2080 Ti GPU. We have evaluated the accuracy of our approach with binaural acoustic effects generated using an interactive geometric sound propagation algorithm and captured real acoustic effects / real-world recordings. We also performed a perceptual evaluation and observed that the audio rendered by our approach is more plausible than audio rendered using prior learning-based and geometric-based sound propagation algorithms. We quantitatively evaluated the accuracy of our approach using statistical acoustic parameters, and energy decay curves. The demo videos, code and dataset are available online (https://anton-jeran.github.io/Listen2Scene/).
The top-k operator returns a sparse vector, where the non-zero values correspond to the k largest values of the input. Unfortunately, because it is a discontinuous function, it is difficult to incorporate in neural networks trained end-to-end with backpropagation. Recent works have considered differentiable relaxations, based either on regularization or perturbation techniques. However, to date, no approach is fully differentiable and sparse. In this paper, we propose new differentiable and sparse top-k operators. We view the top-k operator as a linear program over the permutahedron, the convex hull of permutations. We then introduce a p-norm regularization term to smooth out the operator, and show that its computation can be reduced to isotonic optimization. Our framework is significantly more general than the existing one and allows for example to express top-k operators that select values in magnitude. On the algorithmic side, in addition to pool adjacent violator (PAV) algorithms, we propose a new GPU/TPU-friendly Dykstra algorithm to solve isotonic optimization problems. We successfully use our operators to prune weights in neural networks, to fine-tune vision transformers, and as a router in sparse mixture of experts.
In this work, we aim at an important but less explored problem of a simple yet effective backbone specific for cross-view geo-localization task. Existing methods for cross-view geo-localization tasks are frequently characterized by 1) complicated methodologies, 2) GPU-consuming computations, and 3) a stringent assumption that aerial and ground images are centrally or orientation aligned. To address the above three challenges for cross-view image matching, we propose a new backbone network, named Simple Attention-based Image Geo-localization network (SAIG). The proposed SAIG effectively represents long-range interactions among patches as well as cross-view correspondence with multi-head self-attention layers. The "narrow-deep" architecture of our SAIG improves the feature richness without degradation in performance, while its shallow and effective convolutional stem preserves the locality, eliminating the loss of patchify boundary information. Our SAIG achieves state-of-the-art results on cross-view geo-localization, while being far simpler than previous works. Furthermore, with only 15.9% of the model parameters and half of the output dimension compared to the state-of-the-art, the SAIG adapts well across multiple cross-view datasets without employing any well-designed feature aggregation modules or feature alignment algorithms. In addition, our SAIG attains competitive scores on image retrieval benchmarks, further demonstrating its generalizability. As a backbone network, our SAIG is both easy to follow and computationally lightweight, which is meaningful in practical scenario. Moreover, we propose a simple Spatial-Mixed feature aggregation moDule (SMD) that can mix and project spatial information into a low-dimensional space to generate feature descriptors... (The code is available at https://github.com/yanghongji2007/SAIG)
We present a hardware-efficient architecture of convolutional neural network, which has a repvgg-like architecture. Flops or parameters are traditional metrics to evaluate the efficiency of networks which are not sensitive to hardware including computing ability and memory bandwidth. Thus, how to design a neural network to efficiently use the computing ability and memory bandwidth of hardware is a critical problem. This paper proposes a method how to design hardware-aware neural network. Based on this method, we designed EfficientRep series convolutional networks, which are high-computation hardware(e.g. GPU) friendly and applied in YOLOv6 object detection framework. YOLOv6 has published YOLOv6N/YOLOv6S/YOLOv6M/YOLOv6L models in v1 and v2 versions.
The detection of small and medium-sized objects in three dimensions has always been a frontier exploration problem. This technology has a very wide application in sports analysis, games, virtual reality, human animation and other fields. The traditional three-dimensional small target detection technology has the disadvantages of high cost, low precision and inconvenience, so it is difficult to apply in practice. With the development of machine learning and deep learning, the technology of computer vision algorithms is becoming more mature. Creating an immersive media experience is considered to be a very important research work in sports. The main work is to explore and solve the problem of football detection under the multiple cameras, aiming at the research and implementation of the live broadcast system of football matches. Using multi cameras detects a target ball and determines its position in three dimension with the occlusion, motion, low illumination of the target object. This paper designed and implemented football detection system under multiple cameras for the detection and capture of targets in real-time matches. The main work mainly consists of three parts, football detector, single camera detection, and multi-cameras detection. The system used bundle adjustment to obtain the three-dimensional position of the target, and the GPU to accelerates data pre-processing and achieve accurate real-time capture of the target. By testing the system, it shows that the system can accurately detect and capture the moving targets in 3D. In addition, the solution in this paper is reusable for large-scale competitions, like basketball and soccer. The system framework can be well transplanted into other similar engineering project systems. It has been put into the market.
Siamese networks are one of the most trending methods to achieve self-supervised visual representation learning (SSL). Since hand labeling is costly, SSL can play a crucial part by allowing deep learning to train on large unlabeled datasets. Meanwhile, Neural Architecture Search (NAS) is becoming increasingly important as a technique to discover novel deep learning architectures. However, early NAS methods based on reinforcement learning or evolutionary algorithms suffered from ludicrous computational and memory costs. In contrast, differentiable NAS, a gradient-based approach, has the advantage of being much more efficient and has thus retained most of the attention in the past few years. In this article, we present NASiam, a novel approach that uses for the first time differentiable NAS to improve the multilayer perceptron projector and predictor (encoder/predictor pair) architectures inside siamese-networks-based contrastive learning frameworks (e.g., SimCLR, SimSiam, and MoCo) while preserving the simplicity of previous baselines. We crafted a search space designed explicitly for multilayer perceptrons, inside which we explored several alternatives to the standard ReLU activation function. We show that these new architectures allow ResNet backbone convolutional models to learn strong representations efficiently. NASiam reaches competitive performance in both small-scale (i.e., CIFAR-10/CIFAR-100) and large-scale (i.e., ImageNet) image classification datasets while costing only a few GPU hours. We discuss the composition of the NAS-discovered architectures and emit hypotheses on why they manage to prevent collapsing behavior. Our code is available at https://github.com/aheuillet/NASiam.
Challenging the Nvidia monopoly, dedicated AI-accelerator chips have begun emerging for tackling the computational challenge that the inference and, especially, the training of modern deep neural networks (DNNs) poses to modern computers. The field has been ridden with studies assessing the performance of these contestants across various DNN model types. However, AI-experts are aware of the limitations of current DNNs and have been working towards the fourth AI wave which will, arguably, rely on more biologically inspired models, predominantly on spiking neural networks (SNNs). At the same time, GPUs have been heavily used for simulating such models in the field of computational neuroscience, yet AI-chips have not been tested on such workloads. The current paper aims at filling this important gap by evaluating multiple, cutting-edge AI-chips (Graphcore IPU, GroqChip, Nvidia GPU with Tensor Cores and Google TPU) on simulating a highly biologically detailed model of a brain region, the inferior olive (IO). This IO application stress-tests the different AI-platforms for highlighting architectural tradeoffs by varying its compute density, memory requirements and floating-point numerical accuracy. Our performance analysis reveals that the simulation problem maps extremely well onto the GPU and TPU architectures, which for networks of 125,000 cells leads to a 28x respectively 1,208x speedup over CPU runtimes. At this speed, the TPU sets a new record for largest real-time IO simulation. The GroqChip outperforms both platforms for small networks but, due to implementing some floating-point operations at reduced accuracy, is found not yet usable for brain simulation.
Classical machine learning (CML) occupies nearly half of machine learning pipelines in production applications. Unfortunately, it fails to utilize the state-of-the-practice devices fully and performs poorly. Without a unified framework, the hybrid deployments of deep learning (DL) and CML also suffer from severe performance and portability issues. This paper presents the design of a unified compiler, called CMLCompiler, for CML inference. We propose two unified abstractions: operator representations and extended computational graphs. The CMLCompiler framework performs the conversion and graph optimization based on two unified abstractions, then outputs an optimized computational graph to DL compilers or frameworks. We implement CMLCompiler on TVM. The evaluation shows CMLCompiler's portability and superior performance. It achieves up to 4.38$\times$ speedup on CPU, 3.31$\times$ speedup on GPU, and 5.09$\times$ speedup on IoT devices, compared to the state-of-the-art solutions -- scikit-learn, intel sklearn, and hummingbird. Our performance of CML and DL mixed pipelines achieves up to 3.04x speedup compared with cross-framework implementations. The project documents and source code are available at https://www.computercouncil.org/cmlcompiler.
Data distillation aims to generate a small data set that closely mimics the performance of a given learning algorithm on the original data set. The distilled dataset is hence useful to simplify the training process thanks to its small data size. However, distilled data samples are not necessarily privacy-preserving, even if they are generally humanly indiscernible. To address this limitation, we introduce differentially private kernel inducing points (DP-KIP) for privacy-preserving data distillation. Unlike our original intention to simply apply DP-SGD to the framework of KIP, we find that KIP using infinitely-wide convolutional neural tangent kernels (conv-NTKs) performs better compared to KIP using fully-connected NTKs. However, KIP with conv-NTKs, due to its convolutional and pooling operations, introduces an unbearable computational complexity, requiring hundreds of V100 GPUs in parallel to train, which is impractical and more importantly, such computational resources are inaccessible to many. To overcome this issue, we propose an alternative that does not require pre-training (to avoid a privacy loss) and can well capture complex information on images, as those features from conv-NKTs do, while the computational cost is manageable by a single V100 GPU. To this end, we propose DP-KIP-ScatterNet, which uses the wavelet features from Scattering networks (ScatterNet) instead of those from conv-NTKs, to perform DP-KIP at a reasonable computational cost. We implement DP-KIP-ScatterNet in -- computationally efficient -- JAX and test on several popular image datasets to show its efficacy and its superior performance compared to state-of-the art methods in image data distillation with differential privacy guarantees.
The recent surge in popularity of diffusion models for image generation has brought new attention to the potential of these models in other areas of media generation. One area that has yet to be fully explored is the application of diffusion models to audio generation. Audio generation requires an understanding of multiple aspects, such as the temporal dimension, long term structure, multiple layers of overlapping sounds, and the nuances that only trained listeners can detect. In this work, we investigate the potential of diffusion models for audio generation. We propose a set of models to tackle multiple aspects, including a new method for text-conditional latent audio diffusion with stacked 1D U-Nets, that can generate multiple minutes of music from a textual description. For each model, we make an effort to maintain reasonable inference speed, targeting real-time on a single consumer GPU. In addition to trained models, we provide a collection of open source libraries with the hope of simplifying future work in the field. Samples can be found at https://bit.ly/audio-diffusion. Codes are at https://github.com/archinetai/audio-diffusion-pytorch.
Despite the popularity of policy gradient methods, they are known to suffer from large variance and high sample complexity. To mitigate this, we introduce SoftTreeMax -- a generalization of softmax that takes planning into account. In SoftTreeMax, we extend the traditional logits with the multi-step discounted cumulative reward, topped with the logits of future states. We consider two variants of SoftTreeMax, one for cumulative reward and one for exponentiated reward. For both, we analyze the gradient variance and reveal for the first time the role of a tree expansion policy in mitigating this variance. We prove that the resulting variance decays exponentially with the planning horizon as a function of the expansion policy. Specifically, we show that the closer the resulting state transitions are to uniform, the faster the decay. In a practical implementation, we utilize a parallelized GPU-based simulator for fast and efficient tree search. Our differentiable tree-based policy leverages all gradients at the tree leaves in each environment step instead of the traditional single-sample-based gradient. We then show in simulation how the variance of the gradient is reduced by three orders of magnitude, leading to better sample complexity compared to the standard policy gradient. On Atari, SoftTreeMax demonstrates up to 5x better performance in a faster run time compared to distributed PPO. Lastly, we demonstrate that high reward correlates with lower variance.
Inspired by natural evolutionary processes, Evolutionary Computation (EC) has established itself as a cornerstone of Artificial Intelligence. Recently, with the surge in data-intensive applications and large-scale complex systems, the demand for scalable EC solutions has grown significantly. However, most existing EC infrastructures fall short of catering to the heightened demands of large-scale problem solving. While the advent of some pioneering GPU-accelerated EC libraries is a step forward, they also grapple with some limitations, particularly in terms of flexibility and architectural robustness. In response, we introduce EvoX: a computing framework tailored for automated, distributed, and heterogeneous execution of EC algorithms. At the core of EvoX lies a unique programming model to streamline the development of parallelizable EC algorithms, complemented by a computation model specifically optimized for distributed GPU acceleration. Building upon this foundation, we have crafted an extensive library comprising a wide spectrum of 50+ EC algorithms for both single- and multi-objective optimization. Furthermore, the library offers comprehensive support for a diverse set of benchmark problems, ranging from dozens of numerical test functions to hundreds of reinforcement learning tasks. Through extensive experiments across a range of problem scenarios and hardware configurations, EvoX demonstrates robust system and model performances. EvoX is open-source and accessible at: https://github.com/EMI-Group/EvoX.
Pitch is a foundational aspect of our perception of audio signals. Pitch contours are commonly used to analyze speech and music signals and as input features for many audio tasks, including music transcription, singing voice synthesis, and prosody editing. In this paper, we describe a set of techniques for improving the accuracy of widely-used neural pitch and periodicity estimators to achieve state-of-the-art performance on both speech and music. We also introduce a novel entropy-based method for extracting periodicity and per-frame voiced-unvoiced classifications from statistical inference-based pitch estimators (e.g., neural networks), and show how to train a neural pitch estimator to simultaneously handle both speech and music data (i.e., cross-domain estimation) without performance degradation. Our estimator implementations run 11.2x faster than real-time on a Intel i9-9820X 10-core 3.30 GHz CPU$\unicode{x2014}$approaching the speed of state-of-the-art DSP-based pitch estimators$\unicode{x2014}$or 408x faster than real-time on a NVIDIA GeForce RTX 3090 GPU. We release all of our code and models as Pitch-Estimating Neural Networks (penn), an open-source, pip-installable Python module for training, evaluating, and performing inference with pitch- and periodicity-estimating neural networks. The code for penn is available at https://github.com/interactiveaudiolab/penn.
Training large deep neural network models is highly challenging due to their tremendous computational and memory requirements. Blockwise distillation provides one promising method towards faster convergence by splitting a large model into multiple smaller models. In state-of-the-art blockwise distillation methods, training is performed block-by-block in a data-parallel manner using multiple GPUs. To produce inputs for the student blocks, the teacher model is executed from the beginning until the current block under training. However, this results in a high overhead of redundant teacher execution, low GPU utilization, and extra data loading. To address these problems, we propose Pipe-BD, a novel parallelization method for blockwise distillation. Pipe-BD aggressively utilizes pipeline parallelism for blockwise distillation, eliminating redundant teacher block execution and increasing per-device batch size for better resource utilization. We also extend to hybrid parallelism for efficient workload balancing. As a result, Pipe-BD achieves significant acceleration without modifying the mathematical formulation of blockwise distillation. We implement Pipe-BD on PyTorch, and experiments reveal that Pipe-BD is effective on multiple scenarios, models, and datasets.
In up-to-date machine learning (ML) applications on cloud or edge computing platforms, batching is an important technique for providing efficient and economical services at scale. In particular, parallel computing resources on the platforms, such as graphics processing units (GPUs), have higher computational and energy efficiency with larger batch sizes. However, larger batch sizes may also result in longer response time, and thus it requires a judicious design. This paper aims to provide a dynamic batching policy that strikes a balance between efficiency and latency. The GPU-based inference service is modeled as a batch service queue with batch-size dependent processing time. Then, the design of dynamic batching is a continuous-time average-cost problem, and is formulated as a semi-Markov decision process (SMDP) with the objective of minimizing the weighted sum of average response time and average power consumption. The optimal policy is acquired by solving an associated discrete-time Markov decision process (MDP) problem with finite state approximation and "discretization". By introducing an abstract cost to reflect the impact of "tail" states, the space complexity and the time complexity of the procedure can decrease by 63.5% and 98%, respectively. Our results show that the optimal policies potentially possess a control limit structure. Numerical results also show that SMDP-based batching policies can adapt to different traffic intensities and outperform other benchmark policies. Furthermore, the proposed solution has notable flexibility in balancing power consumption and latency.
Micro-drones can be integrated into various industrial applications but are constrained by their computing power and expert pilots, a secondary challenge. This study presents a computationally-efficient deep convolutional neural network that utilizes Gabor filters and spatial separable convolutions with low computational complexities. An attention module is integrated with the model to complement the performance. Further, perception-based action space and trajectory generators are integrated with the model's predictions for intuitive navigation. The computationally-efficient model aids a human operator in controlling a micro-drone via gestures. Nearly 18% of computational resources are conserved using the NVIDIA GPU profiler during training. Using a low-cost DJI Tello drone for experiment verification, the computationally-efficient model shows promising results compared to a state-of-the-art and conventional computer vision-based technique.
Text-to-audio (TTA) system has recently gained attention for its ability to synthesize general audio based on text descriptions. However, previous studies in TTA have limited generation quality with high computational costs. In this study, we propose AudioLDM, a TTA system that is built on a latent space to learn the continuous audio representations from contrastive language-audio pretraining (CLAP) latents. The pretrained CLAP models enable us to train LDMs with audio embedding while providing text embedding as a condition during sampling. By learning the latent representations of audio signals and their compositions without modeling the cross-modal relationship, AudioLDM is advantageous in both generation quality and computational efficiency. Trained on AudioCaps with a single GPU, AudioLDM achieves state-of-the-art TTA performance measured by both objective and subjective metrics (e.g., frechet distance). Moreover, AudioLDM is the first TTA system that enables various text-guided audio manipulations (e.g., style transfer) in a zero-shot fashion. Our implementation and demos are available at https://audioldm.github.io.