Wen-mei Hwu

Arpandeep Khatua, Vikram Sharma Mailthody, Bhagyashree Taleka et al.

Graph neural networks (GNNs) have shown high potential for a variety of real-world, challenging applications, but one of the major obstacles in GNN research is the lack of large-scale flexible datasets. Most existing public datasets for GNNs are relatively small, which limits the ability of GNNs to generalize to unseen data. The few existing large-scale graph datasets provide very limited labeled data. This makes it difficult to determine if the GNN model's low accuracy for unseen data is inherently due to insufficient training data or if the model failed to generalize. Additionally, datasets used to train GNNs need to offer flexibility to enable a thorough study of the impact of various factors while training GNN models. In this work, we introduce the Illinois Graph Benchmark (IGB), a research dataset tool that the developers can use to train, scrutinize and systematically evaluate GNN models with high fidelity. IGB includes both homogeneous and heterogeneous academic graphs of enormous sizes, with more than 40% of their nodes labeled. Compared to the largest graph datasets publicly available, the IGB provides over 162X more labeled data for deep learning practitioners and developers to create and evaluate models with higher accuracy. The IGB dataset is a collection of academic graphs designed to be flexible, enabling the study of various GNN architectures, embedding generation techniques, and analyzing system performance issues for node classification tasks. IGB is open-sourced, supports DGL and PyG frameworks, and comes with releases of the raw text that we believe foster emerging language models and GNN research projects. An early public version of IGB is available at https://github.com/IllinoisGraphBenchmark/IGB-Datasets.

Jeongmin Brian Park, Vikram Sharma Mailthody, Zaid Qureshi et al.

Graph Neural Networks (GNNs) are emerging as a powerful tool for learning from graph-structured data and performing sophisticated inference tasks in various application domains. Although GNNs have been shown to be effective on modest-sized graphs, training them on large-scale graphs remains a significant challenge due to lack of efficient data access and data movement methods. Existing frameworks for training GNNs use CPUs for graph sampling and feature aggregation, while the training and updating of model weights are executed on GPUs. However, our in-depth profiling shows the CPUs cannot achieve the throughput required to saturate GNN model training throughput, causing gross under-utilization of expensive GPU resources. Furthermore, when the graph and its embeddings do not fit in the CPU memory, the overhead introduced by the operating system, say for handling page-faults, comes in the critical path of execution. To address these issues, we propose the GPU Initiated Direct Storage Access (GIDS) dataloader, to enable GPU-oriented GNN training for large-scale graphs while efficiently utilizing all hardware resources, such as CPU memory, storage, and GPU memory with a hybrid data placement strategy. By enabling GPU threads to fetch feature vectors directly from storage, GIDS dataloader solves the memory capacity problem for GPU-oriented GNN training. Moreover, GIDS dataloader leverages GPU parallelism to tolerate storage latency and eliminates expensive page-fault overhead. Doing so enables us to design novel optimizations for exploiting locality and increasing effective bandwidth for GNN training. Our evaluation using a single GPU on terabyte-scale GNN datasets shows that GIDS dataloader accelerates the overall DGL GNN training pipeline by up to 392X when compared to the current, state-of-the-art DGL dataloader.

Jie Huang, Kerui Zhu, Kevin Chen-Chuan Chang et al.

We propose DEER (Descriptive Knowledge Graph for Explaining Entity Relationships) - an open and informative form of modeling entity relationships. In DEER, relationships between entities are represented by free-text relation descriptions. For instance, the relationship between entities of machine learning and algorithm can be represented as ``Machine learning explores the study and construction of algorithms that can learn from and make predictions on data.'' To construct DEER, we propose a self-supervised learning method to extract relation descriptions with the analysis of dependency patterns and generate relation descriptions with a transformer-based relation description synthesizing model, where no human labeling is required. Experiments demonstrate that our system can extract and generate high-quality relation descriptions for explaining entity relationships. The results suggest that we can build an open and informative knowledge graph without human annotation.

Jie Huang, Kevin Chen-Chuan Chang, Jinjun Xiong et al.

"He is a person", "Paris is located on the earth". Both statements are correct but meaningless - due to lack of specificity. In this paper, we propose to measure how specific the language of pre-trained language models (PLMs) is. To achieve this, we introduce a novel approach to build a benchmark for specificity testing by forming masked token prediction tasks with prompts. For instance, given "Toronto is located in [MASK].", we want to test whether a more specific answer will be better filled in by PLMs, e.g., Ontario instead of Canada. From our evaluations, we show that existing PLMs have only a slight preference for more specific answers. We identify underlying factors affecting the specificity and design two prompt-based methods to improve the specificity. Results show that the specificity of the models can be improved by the proposed methods without additional training. We hope this work can bring to awareness the notion of specificity of language models and encourage the research community to further explore this important but understudied problem.

Tong Zhang, Vikram Sharma Mailthody, Fei Sun et al.

In 1987, Jim Gray and Gianfranco Putzolu introduced the five-minute rule, a simple, storage-memory-economics-based heuristic for deciding when data should live in DRAM rather than on storage. Subsequent revisits to the rule largely retained that economics-only view, leaving host costs, feasibility limits, and workload behavior out of scope. This paper revisits the rule from first principles, integrating host costs, DRAM bandwidth/capacity, and physics-grounded models of SSD performance and cost, and then embedding these elements in a constraint- and workload-aware framework that yields actionable provisioning guidance. We show that, for modern AI platforms, especially GPU-centric hosts paired with ultra-high-IOPS SSDs engineered for fine-grained random access, the DRAM$\leftrightarrow$flash caching threshold collapses from minutes to a few seconds. This shift reframes NAND flash memory as an \emph{active data tier} and exposes a broad research space across the hardware-software stack. We further introduce MQSim-Next, a calibrated SSD simulator that supports validation and sensitivity analysis and facilitates future architectural and system research. Finally, we present two concrete case studies that showcase the software system design space opened by such memory hierarchy paradigm shift. Overall, we turn a classical heuristic into an actionable, feasibility-aware analysis and provisioning framework and set the stage for further research on AI-era memory hierarchy.

Omer Anjum, Alok Kamatar, Toby Liang et al.

Assigning qualified, unbiased and interested reviewers to paper submissions is vital for maintaining the integrity and quality of the academic publishing system and providing valuable reviews to authors. However, matching thousands of submissions with thousands of potential reviewers within a limited time is a daunting challenge for a conference program committee. Prior efforts based on topic modeling have suffered from losing the specific context that help define the topics in a publication or submission abstract. Moreover, in some cases, topics identified are difficult to interpret. We propose an approach that learns from each abstract published by a potential reviewer the topics studied and the explicit context in which the reviewer studied the topics. Furthermore, we contribute a new dataset for evaluating reviewer matching systems. Our experiments show a significant, consistent improvement in precision when compared with the existing methods. We also use examples to demonstrate why our recommendations are more explainable. The new approach has been deployed successfully at top-tier conferences in the last two years.

Jeongmin Brian Park, Kun Wu, Vikram Sharma Mailthody et al.

Graph Neural Networks (GNNs) are widely used today in recommendation systems, fraud detection, and node/link classification tasks. Real world GNNs continue to scale in size and require a large memory footprint for storing graphs and embeddings that often exceed the memory capacities of the target GPUs used for training. To address limited memory capacities, traditional GNN training approaches use graph partitioning and sharding techniques to scale up across multiple GPUs within a node and/or scale out across multiple nodes. However, this approach suffers from the high computational costs of graph partitioning algorithms and inefficient communication across GPUs. To address these overheads, we propose Large-scale Storage-based Multi-GPU GNN framework (LSM-GNN), a storagebased approach to train GNN models that utilizes a novel communication layer enabling GPU software caches to function as a system-wide shared cache with low overheads.LSM-GNN incorporates a hybrid eviction policy that intelligently manages cache space by using both static and dynamic node information to significantly enhance cache performance. Furthermore, we introduce the Preemptive Victim-buffer Prefetcher (PVP), a mechanism for prefetching node feature data from a Victim Buffer located in CPU pinned-memory to further reduce the pressure on the storage devices. Experimental results show that despite the lower compute capabilities and memory capacities, LSM-GNN in a single node with two GPUs offers superior performance over two-node-four-GPU Dist-DGL baseline and provides up to 3.75x speed up on end-to-end epoch time while running large-scale GNN training

Ali Hassani, Fengzhe Zhou, Aditya Kane et al. · gatech

Many sparse attention mechanisms such as Neighborhood Attention have typically failed to consistently deliver speedup over the self attention baseline. This is largely due to the level of complexity in attention infrastructure, and the rapid evolution of AI hardware architecture. At the same time, many state-of-the-art foundational models, particularly in computer vision, are heavily bound by attention, and need reliable sparsity to escape the O(n^2) complexity. In this paper, we study a class of promising sparse attention mechanisms that focus on locality, and aim to develop a better analytical model of their performance improvements. We first introduce Generalized Neighborhood Attention (GNA), which can describe sliding window, strided sliding window, and blocked attention. We then consider possible design choices in implementing these approaches, and create a simulator that can provide much more realistic speedup upper bounds for any given setting. Finally, we implement GNA on top of a state-of-the-art fused multi-headed attention (FMHA) kernel designed for the NVIDIA Blackwell architecture in CUTLASS. Our implementation can fully realize the maximum speedup theoretically possible in many perfectly block-sparse cases, and achieves an effective utilization of 1.3 petaFLOPs/second in FP16. In addition, we plug various GNA configurations into off-the-shelf generative models, such as Cosmos-7B, HunyuanVideo, and FLUX, and show that it can deliver 28% to 46% end-to-end speedup on B200 without any fine-tuning. We will open source our simulator and Blackwell kernels directly through the NATTEN project.

Yen-Hsiang Chang, Jianhao Pu, Wen-mei Hwu et al.

With the society's growing adoption of machine learning (ML) and deep learning (DL) for various intelligent solutions, it becomes increasingly imperative to standardize a common set of measures for ML/DL models with large scale open datasets under common development practices and resources so that people can benchmark and compare models quality and performance on a common ground. MLCommons has emerged recently as a driving force from both industry and academia to orchestrate such an effort. Despite its wide adoption as standardized benchmarks, MLCommons Inference has only included a limited number of ML/DL models (in fact seven models in total). This significantly limits the generality of MLCommons Inference's benchmarking results because there are many more novel ML/DL models from the research community, solving a wide range of problems with different inputs and outputs modalities. To address such a limitation, we propose MLHarness, a scalable benchmarking harness system for MLCommons Inference with three distinctive features: (1) it codifies the standard benchmark process as defined by MLCommons Inference including the models, datasets, DL frameworks, and software and hardware systems; (2) it provides an easy and declarative approach for model developers to contribute their models and datasets to MLCommons Inference; and (3) it includes the support of a wide range of models with varying inputs/outputs modalities so that we can scalably benchmark these models across different datasets, frameworks, and hardware systems. This harness system is developed on top of the MLModelScope system, and will be open sourced to the community. Our experimental results demonstrate the superior flexibility and scalability of this harness system for MLCommons Inference benchmarking.

Jiachen Li, Bowen Cheng, Rogerio Feris et al.

Current anchor-free object detectors are quite simple and effective yet lack accurate label assignment methods, which limits their potential in competing with classic anchor-based models that are supported by well-designed assignment methods based on the Intersection-over-Union~(IoU) metric. In this paper, we present \textbf{Pseudo-Intersection-over-Union~(Pseudo-IoU)}: a simple metric that brings more standardized and accurate assignment rule into anchor-free object detection frameworks without any additional computational cost or extra parameters for training and testing, making it possible to further improve anchor-free object detection by utilizing training samples of good quality under effective assignment rules that have been previously applied in anchor-based methods. By incorporating Pseudo-IoU metric into an end-to-end single-stage anchor-free object detection framework, we observe consistent improvements in their performance on general object detection benchmarks such as PASCAL VOC and MSCOCO. Our method (single-model and single-scale) also achieves comparable performance to other recent state-of-the-art anchor-free methods without bells and whistles. Our code is based on mmdetection toolbox and will be made publicly available at https://github.com/SHI-Labs/Pseudo-IoU-for-Anchor-Free-Object-Detection.

Abdul Dakkak, Cheng Li, Jinjun Xiong et al.

Machine Learning (ML) and Deep Learning (DL) innovations are being introduced at such a rapid pace that researchers are hard-pressed to analyze and study them. The complicated procedures for evaluating innovations, along with the lack of standard and efficient ways of specifying and provisioning ML/DL evaluation, is a major "pain point" for the community. This paper proposes MLModelScope, an open-source, framework/hardware agnostic, extensible and customizable design that enables repeatable, fair, and scalable model evaluation and benchmarking. We implement the distributed design with support for all major frameworks and hardware, and equip it with web, command-line, and library interfaces. To demonstrate MLModelScope's capabilities we perform parallel evaluation and show how subtle changes to model evaluation pipeline affects the accuracy and HW/SW stack choices affect performance.

Cheng Li, Abdul Dakkak, Jinjun Xiong et al.

The current Deep Learning (DL) landscape is fast-paced and is rife with non-uniform models, hardware/software (HW/SW) stacks, but lacks a DL benchmarking platform to facilitate evaluation and comparison of DL innovations, be it models, frameworks, libraries, or hardware. Due to the lack of a benchmarking platform, the current practice of evaluating the benefits of proposed DL innovations is both arduous and error-prone - stifling the adoption of the innovations. In this work, we first identify $10$ design features which are desirable within a DL benchmarking platform. These features include: performing the evaluation in a consistent, reproducible, and scalable manner, being framework and hardware agnostic, supporting real-world benchmarking workloads, providing in-depth model execution inspection across the HW/SW stack levels, etc. We then propose MLModelScope, a DL benchmarking platform design that realizes the $10$ objectives. MLModelScope proposes a specification to define DL model evaluations and techniques to provision the evaluation workflow using the user-specified HW/SW stack. MLModelScope defines abstractions for frameworks and supports board range of DL models and evaluation scenarios. We implement MLModelScope as an open-source project with support for all major frameworks and hardware architectures. Through MLModelScope's evaluation and automated analysis workflows, we performed case-study analyses of $37$ models across $4$ systems and show how model, hardware, and framework selection affects model accuracy and performance under different benchmarking scenarios. We further demonstrated how MLModelScope's tracing capability gives a holistic view of model execution and helps pinpoint bottlenecks.

Bowen Cheng, Yunchao Wei, Rogerio Feris et al.

In this paper, we analyze failure cases of state-of-the-art detectors and observe that most hard false positives result from classification instead of localization and they have a large negative impact on the performance of object detectors. We conjecture there are three factors: (1) Shared feature representation is not optimal due to the mismatched goals of feature learning for classification and localization; (2) multi-task learning helps, yet optimization of the multi-task loss may result in sub-optimal for individual tasks; (3) large receptive field for different scales leads to redundant context information for small objects. We demonstrate the potential of detector classification power by a simple, effective, and widely-applicable Decoupled Classification Refinement (DCR) network. In particular, DCR places a separate classification network in parallel with the localization network (base detector). With ROI Pooling placed on the early stage of the classification network, we enforce an adaptive receptive field in DCR. During training, DCR samples hard false positives from the base detector and trains a strong classifier to refine classification results. During testing, DCR refines all boxes from the base detector. Experiments show competitive results on PASCAL VOC and COCO without any bells and whistles. Our codes are available at: https://github.com/bowenc0221/Decoupled-Classification-Refinement.

Ali Hassani, Wen-Mei Hwu, Humphrey Shi · gatech

Neighborhood attention reduces the cost of self attention by restricting each token's attention span to its nearest neighbors. This restriction, parameterized by a window size and dilation factor, draws a spectrum of possible attention patterns between linear projection and self attention. Neighborhood attention, and more generally sliding window attention patterns, have long been bounded by infrastructure, particularly in higher-rank spaces (2-D and 3-D), calling for the development of custom kernels, which have been limited in either functionality, or performance, if not both. In this work, we aim to massively improve upon existing infrastructure by providing two new methods for implementing neighborhood attention. We first show that neighborhood attention can be represented as a batched GEMM problem, similar to standard attention, and implement it for 1-D and 2-D neighborhood attention. These kernels on average provide 895% and 272% improvement in full precision runtime compared to existing naive CUDA kernels for 1-D and 2-D neighborhood attention respectively. We find that aside from being heavily bound by memory bandwidth, certain inherent inefficiencies exist in all unfused implementations of neighborhood attention, which in most cases undo their theoretical efficiency gain. Motivated by the progress made into fused dot-product attention kernels, we developed fused neighborhood attention; an adaptation of fused dot-product attention kernels that allow fine-grained control over attention across different spatial axes. Known for reducing the quadratic time complexity of self attention to a linear complexity, neighborhood attention can now enjoy a reduced and constant memory footprint, and record-breaking half precision runtime. We observe that our fused implementation successfully circumvents some of the unavoidable inefficiencies in unfused implementations...

Jitesh Jain, Shubham Maheshwari, Ning Yu et al. · gatech

Riding on the success of LLMs with retrieval-augmented generation (RAG), there has been a growing interest in augmenting agent systems with external memory databases. However, the existing systems focus on storing text information in their memory, ignoring the importance of multimodal signals. Motivated by the multimodal nature of human memory, we present AUGUSTUS, a multimodal agent system aligned with the ideas of human memory in cognitive science. Technically, our system consists of 4 stages connected in a loop: (i) encode: understanding the inputs; (ii) store in memory: saving important information; (iii) retrieve: searching for relevant context from memory; and (iv) act: perform the task. Unlike existing systems that use vector databases, we propose conceptualizing information into semantic tags and associating the tags with their context to store them in a graph-structured multimodal contextual memory for efficient concept-driven retrieval. Our system outperforms the traditional multimodal RAG approach while being 3.5 times faster for ImageNet classification and outperforming MemGPT on the MSC benchmark.

Jie Huang, Hanyin Shao, Kevin Chen-Chuan Chang et al.

Can machines know what twin prime is? From the composition of this phrase, machines may guess twin prime is a certain kind of prime, but it is still difficult to deduce exactly what twin stands for without additional knowledge. Here, twin prime is a jargon - a specialized term used by experts in a particular field. Explaining jargon is challenging since it usually requires domain knowledge to understand. Recently, there is an increasing interest in extracting and generating definitions of words automatically. However, existing approaches, either extraction or generation, perform poorly on jargon. In this paper, we propose to combine extraction and generation for jargon definition modeling: first extract self- and correlative definitional information of target jargon from the Web and then generate the final definitions by incorporating the extracted definitional information. Our framework is remarkably simple but effective: experiments demonstrate our method can generate high-quality definitions for jargon and outperform state-of-the-art models significantly, e.g., BLEU score from 8.76 to 22.66 and human-annotated score from 2.34 to 4.04.

Seung Won Min, Kun Wu, Mert Hidayetoğlu et al.

Graph Neural Networks (GNNs) have shown success in learning from graph-structured data, with applications to fraud detection, recommendation, and knowledge graph reasoning. However, training GNN efficiently is challenging because: 1) GPU memory capacity is limited and can be insufficient for large datasets, and 2) the graph-based data structure causes irregular data access patterns. In this work, we provide a method to statistical analyze and identify more frequently accessed data ahead of GNN training. Our data tiering method not only utilizes the structure of input graph, but also an insight gained from actual GNN training process to achieve a higher prediction result. With our data tiering method, we additionally provide a new data placement and access strategy to further minimize the CPU-GPU communication overhead. We also take into account of multi-GPU GNN training as well and we demonstrate the effectiveness of our strategy in a multi-GPU system. The evaluation results show that our work reduces CPU-GPU traffic by 87-95% and improves the training speed of GNN over the existing solutions by 1.6-2.1x on graphs with hundreds of millions of nodes and billions of edges.

Jie Huang, Kevin Chen-Chuan Chang, Jinjun Xiong et al.

Relations between entities can be represented by different instances, e.g., a sentence containing both entities or a fact in a Knowledge Graph (KG). However, these instances may not well capture the general relations between entities, may be difficult to understand by humans, even may not be found due to the incompleteness of the knowledge source. In this paper, we introduce the Open Relation Modeling problem - given two entities, generate a coherent sentence describing the relation between them. To solve this problem, we propose to teach machines to generate definition-like relation descriptions by letting them learn from defining entities. Specifically, we fine-tune Pre-trained Language Models (PLMs) to produce definitions conditioned on extracted entity pairs. To help PLMs reason between entities and provide additional relational knowledge to PLMs for open relation modeling, we incorporate reasoning paths in KGs and include a reasoning path selection mechanism. Experimental results show that our model can generate concise but informative relation descriptions that capture the representative characteristics of entities.

Jie Huang, Kevin Chen-Chuan Chang, Jinjun Xiong et al.

We propose to measure fine-grained domain relevance - the degree that a term is relevant to a broad (e.g., computer science) or narrow (e.g., deep learning) domain. Such measurement is crucial for many downstream tasks in natural language processing. To handle long-tail terms, we build a core-anchored semantic graph, which uses core terms with rich description information to bridge the vast remaining fringe terms semantically. To support a fine-grained domain without relying on a matching corpus for supervision, we develop hierarchical core-fringe learning, which learns core and fringe terms jointly in a semi-supervised manner contextualized in the hierarchy of the domain. To reduce expensive human efforts, we employ automatic annotation and hierarchical positive-unlabeled learning. Our approach applies to big or small domains, covers head or tail terms, and requires little human effort. Extensive experiments demonstrate that our methods outperform strong baselines and even surpass professional human performance.

Seung Won Min, Kun Wu, Sitao Huang et al.

Graph Convolutional Networks (GCNs) are increasingly adopted in large-scale graph-based recommender systems. Training GCN requires the minibatch generator traversing graphs and sampling the sparsely located neighboring nodes to obtain their features. Since real-world graphs often exceed the capacity of GPU memory, current GCN training systems keep the feature table in host memory and rely on the CPU to collect sparse features before sending them to the GPUs. This approach, however, puts tremendous pressure on host memory bandwidth and the CPU. This is because the CPU needs to (1) read sparse features from memory, (2) write features into memory as a dense format, and (3) transfer the features from memory to the GPUs. In this work, we propose a novel GPU-oriented data communication approach for GCN training, where GPU threads directly access sparse features in host memory through zero-copy accesses without much CPU help. By removing the CPU gathering stage, our method significantly reduces the consumption of the host resources and data access latency. We further present two important techniques to achieve high host memory access efficiency by the GPU: (1) automatic data access address alignment to maximize PCIe packet efficiency, and (2) asynchronous zero-copy access and kernel execution to fully overlap data transfer with training. We incorporate our method into PyTorch and evaluate its effectiveness using several graphs with sizes up to 111 million nodes and 1.6 billion edges. In a multi-GPU training setup, our method is 65-92% faster than the conventional data transfer method, and can even match the performance of all-in-GPU-memory training for some graphs that fit in GPU memory.

Seung Won Min, Kun Wu, Sitao Huang et al.

With the increasing adoption of graph neural networks (GNNs) in the machine learning community, GPUs have become an essential tool to accelerate GNN training. However, training GNNs on very large graphs that do not fit in GPU memory is still a challenging task. Unlike conventional neural networks, mini-batching input samples in GNNs requires complicated tasks such as traversing neighboring nodes and gathering their feature values. While this process accounts for a significant portion of the training time, we find existing GNN implementations using popular deep neural network (DNN) libraries such as PyTorch are limited to a CPU-centric approach for the entire data preparation step. This "all-in-CPU" approach has negative impact on the overall GNN training performance as it over-utilizes CPU resources and hinders GPU acceleration of GNN training. To overcome such limitations, we introduce PyTorch-Direct, which enables a GPU-centric data accessing paradigm for GNN training. In PyTorch-Direct, GPUs are capable of efficiently accessing complicated data structures in host memory directly without CPU intervention. Our microbenchmark and end-to-end GNN training results show that PyTorch-Direct reduces data transfer time by 47.1% on average and speeds up GNN training by up to 1.6x. Furthermore, by reducing CPU utilization, PyTorch-Direct also saves system power by 12.4% to 17.5% during training. To minimize programmer effort, we introduce a new "unified tensor" type along with necessary changes to the PyTorch memory allocator, dispatch logic, and placement rules. As a result, users need to change at most two lines of their PyTorch GNN training code for each tensor object to take advantage of PyTorch-Direct.

Vikram Sharma Mailthody, James Wei, Nicholas Chen et al.

COVID-19 has fundamentally disrupted the way we live. Government bodies, universities, and companies worldwide are rapidly developing technologies to combat the COVID-19 pandemic and safely reopen society. Essential analytics tools such as contact tracing, super-spreader event detection, and exposure mapping require collecting and analyzing sensitive user information. The increasing use of such powerful data-driven applications necessitates a secure, privacy-preserving infrastructure for computation on personal data. In this paper, we analyze two such computing infrastructures under development at the University of Illinois at Urbana-Champaign to track and mitigate the spread of COVID-19. First, we present Safer Illinois, a system for decentralized health analytics supporting two applications currently deployed with widespread adoption: digital contact tracing and COVID-19 status cards. Second, we introduce the RokWall architecture for privacy-preserving centralized data analytics on sensitive user data. We discuss the architecture of these systems, design choices, threat models considered, and the challenges we experienced in developing production-ready systems for sensitive data analysis.

Zhonghao Wang, Kai Wang, Mo Yu et al.

We study the problem of concept induction in visual reasoning, i.e., identifying concepts and their hierarchical relationships from question-answer pairs associated with images; and achieve an interpretable model via working on the induced symbolic concept space. To this end, we first design a new framework named object-centric compositional attention model (OCCAM) to perform the visual reasoning task with object-level visual features. Then, we come up with a method to induce concepts of objects and relations using clues from the attention patterns between objects' visual features and question words. Finally, we achieve a higher level of interpretability by imposing OCCAM on the objects represented in the induced symbolic concept space. Our model design makes this an easy adaption via first predicting the concepts of objects and relations and then projecting the predicted concepts back to the visual feature space so the compositional reasoning module can process normally. Experiments on the CLEVR and GQA datasets demonstrate: 1) our OCCAM achieves a new state of the art without human-annotated functional programs; 2) our induced concepts are both accurate and sufficient as OCCAM achieves an on-par performance on objects represented either in visual features or in the induced symbolic concept space.

Cong Hao, Yao Chen, Xiaofan Zhang et al.

High quality AI solutions require joint optimization of AI algorithms, such as deep neural networks (DNNs), and their hardware accelerators. To improve the overall solution quality as well as to boost the design productivity, efficient algorithm and accelerator co-design methodologies are indispensable. In this paper, we first discuss the motivations and challenges for the Algorithm/Accelerator co-design problem and then provide several effective solutions. Especially, we highlight three leading works of effective co-design methodologies: 1) the first simultaneous DNN/FPGA co-design method; 2) a bi-directional lightweight DNN and accelerator co-design method; 3) a differentiable and efficient DNN and accelerator co-search method. We demonstrate the effectiveness of the proposed co-design approaches using extensive experiments on both FPGAs and GPUs, with comparisons to existing works. This paper emphasizes the importance and efficacy of algorithm-accelerator co-design and calls for more research breakthroughs in this interesting and demanding area.

Jie Huang, Zilong Wang, Kevin Chen-Chuan Chang et al.

We introduce and study semantic capacity of terms. For example, the semantic capacity of artificial intelligence is higher than that of linear regression since artificial intelligence possesses a broader meaning scope. Understanding semantic capacity of terms will help many downstream tasks in natural language processing. For this purpose, we propose a two-step model to investigate semantic capacity of terms, which takes a large text corpus as input and can evaluate semantic capacity of terms if the text corpus can provide enough co-occurrence information of terms. Extensive experiments in three fields demonstrate the effectiveness and rationality of our model compared with well-designed baselines and human-level evaluations.

Mert Hidayetoglu, Carl Pearson, Vikram Sharma Mailthody et al.

This paper presents GPU performance optimization and scaling results for inference models of the Sparse Deep Neural Network Challenge 2020. Demands for network quality have increased rapidly, pushing the size and thus the memory requirements of many neural networks beyond the capacity of available accelerators. Sparse deep neural networks (SpDNN) have shown promise for reining in the memory footprint of large neural networks. However, there is room for improvement in implementing SpDNN operations on GPUs. This work presents optimized sparse matrix multiplication kernels fused with the ReLU function. The optimized kernels reuse input feature maps from the shared memory and sparse weights from registers. For multi-GPU parallelism, our SpDNN implementation duplicates weights and statically partition the feature maps across GPUs. Results for the challenge benchmarks show that the proposed kernel design and multi-GPU parallelization achieve up to 180 tera-edges per second inference throughput. These results are up to 4.3x faster for a single GPU and an order of magnitude faster at full scale than those of the champion of the 2019 Sparse Deep Neural Network Graph Challenge for the same generation of NVIDIA V100 GPUs. Using the same implementation, we also show single-GPU throughput on NVIDIA A100 is 2.37$\times$ faster than V100.

Yuhong Li, Cong Hao, Xiaofan Zhang et al.

High quality AI solutions require joint optimization of AI algorithms and their hardware implementations. In this work, we are the first to propose a fully simultaneous, efficient differentiable DNN architecture and implementation co-search (EDD) methodology. We formulate the co-search problem by fusing DNN search variables and hardware implementation variables into one solution space, and maximize both algorithm accuracy and hardware implementation quality. The formulation is differentiable with respect to the fused variables, so that gradient descent algorithm can be applied to greatly reduce the search time. The formulation is also applicable for various devices with different objectives. In the experiments, we demonstrate the effectiveness of our EDD methodology by searching for three representative DNNs, targeting low-latency GPU implementation and FPGA implementations with both recursive and pipelined architectures. Each model produced by EDD achieves similar accuracy as the best existing DNN models searched by neural architecture search (NAS) methods on ImageNet, but with superior performance obtained within 12 GPU-hour searches. Our DNN targeting GPU is 1.40x faster than the state-of-the-art solution reported in Proxyless, and our DNN targeting FPGA delivers 1.45x higher throughput than the state-of-the-art solution reported in DNNBuilder.

Zhonghao Wang, Yunchao Wei, Rogerior Feris et al.

Learning segmentation from synthetic data and adapting to real data can significantly relieve human efforts in labelling pixel-level masks. A key challenge of this task is how to alleviate the data distribution discrepancy between the source and target domains, i.e. reducing domain shift. The common approach to this problem is to minimize the discrepancy between feature distributions from different domains through adversarial training. However, directly aligning the feature distribution globally cannot guarantee consistency from a local view (i.e. semantic-level), which prevents certain semantic knowledge learned on the source domain from being applied to the target domain. To tackle this issue, we propose a semi-supervised approach named Alleviating Semantic-level Shift (ASS), which can successfully promote the distribution consistency from both global and local views. Specifically, leveraging a small number of labeled data from the target domain, we directly extract semantic-level feature representations from both the source and the target domains by averaging the features corresponding to same categories advised by pixel-level masks. We then feed the produced features to the discriminator to conduct semantic-level adversarial learning, which collaborates with the adversarial learning from the global view to better alleviate the domain shift. We apply our ASS to two domain adaptation tasks, from GTA5 to Cityscapes and from Synthia to Cityscapes. Extensive experiments demonstrate that: (1) ASS can significantly outperform the current unsupervised state-of-the-arts by employing a small number of annotated samples from the target domain; (2) ASS can beat the oracle model trained on the whole target dataset by over 3 points by augmenting the synthetic source data with annotated samples from the target domain without suffering from the prevalent problem of overfitting to the source domain.

Zhonghao Wang, Mo Yu, Yunchao Wei et al.

We consider the problem of unsupervised domain adaptation for semantic segmentation by easing the domain shift between the source domain (synthetic data) and the target domain (real data) in this work. State-of-the-art approaches prove that performing semantic-level alignment is helpful in tackling the domain shift issue. Based on the observation that stuff categories usually share similar appearances across images of different domains while things (i.e. object instances) have much larger differences, we propose to improve the semantic-level alignment with different strategies for stuff regions and for things: 1) for the stuff categories, we generate feature representation for each class and conduct the alignment operation from the target domain to the source domain; 2) for the thing categories, we generate feature representation for each individual instance and encourage the instance in the target domain to align with the most similar one in the source domain. In this way, the individual differences within thing categories will also be considered to alleviate over-alignment. In addition to our proposed method, we further reveal the reason why the current adversarial loss is often unstable in minimizing the distribution discrepancy and show that our method can help ease this issue by minimizing the most similar stuff and instance features between the source and the target domains. We conduct extensive experiments in two unsupervised domain adaptation tasks, i.e. GTA5 to Cityscapes and SYNTHIA to Cityscapes, and achieve the new state-of-the-art segmentation accuracy.

Abdul Dakkak, Cheng Li, Jinjun Xiong et al.

Deep Learning (DL) innovations are being introduced at a rapid pace. However, the current lack of standard specification of DL tasks makes sharing, running, reproducing, and comparing these innovations difficult. To address this problem, we propose DLSpec, a model-, dataset-, software-, and hardware-agnostic DL specification that captures the different aspects of DL tasks. DLSpec has been tested by specifying and running hundreds of DL tasks.

Cheng Li, Abdul Dakkak, Jinjun Xiong et al.

The past few years have seen a surge of applying Deep Learning (DL) models for a wide array of tasks such as image classification, object detection, machine translation, etc. While DL models provide an opportunity to solve otherwise intractable tasks, their adoption relies on them being optimized to meet latency and resource requirements. Benchmarking is a key step in this process but has been hampered in part due to the lack of representative and up-to-date benchmarking suites. This is exacerbated by the fast-evolving pace of DL models. This paper proposes DLBricks, a composable benchmark generation design that reduces the effort of developing, maintaining, and running DL benchmarks on CPUs. DLBricks decomposes DL models into a set of unique runnable networks and constructs the original model's performance using the performance of the generated benchmarks. DLBricks leverages two key observations: DL layers are the performance building blocks of DL models and layers are extensively repeated within and across DL models. Since benchmarks are generated automatically and the benchmarking time is minimized, DLBricks can keep up-to-date with the latest proposed models, relieving the pressure of selecting representative DL models. Moreover, DLBricks allows users to represent proprietary models within benchmark suites. We evaluate DLBricks using $50$ MXNet models spanning $5$ DL tasks on $4$ representative CPU systems. We show that DLBricks provides an accurate performance estimate for the DL models and reduces the benchmarking time across systems (e.g. within $95\%$ accuracy and up to $4.4\times$ benchmarking time speedup on Amazon EC2 c5.xlarge).

Cong Hao, Yao Chen, Xinheng Liu et al.

The rapidly growing demands for powerful AI algorithms in many application domains have motivated massive investment in both high-quality deep neural network (DNN) models and high-efficiency implementations. In this position paper, we argue that a simultaneous DNN/implementation co-design methodology, named Neural Architecture and Implementation Search (NAIS), deserves more research attention to boost the development productivity and efficiency of both DNN models and implementation optimization. We propose a stylized design methodology that can drastically cut down the search cost while preserving the quality of the end solution.As an illustration, we discuss this DNN/implementation methodology in the context of both FPGAs and GPUs. We take autonomous driving as a key use case as it is one of the most demanding areas for high quality AI algorithms and accelerators. We discuss how such a co-design methodology can impact the autonomous driving industry significantly. We identify several research opportunities in this exciting domain.

Cheng Li, Abdul Dakkak, Jinjun Xiong et al.

As Deep Learning (DL) models have been increasingly used in latency-sensitive applications, there has been a growing interest in improving their response time. An important venue for such improvement is to profile the execution of these models and characterize their performance to identify possible optimization opportunities. However, the current profiling tools lack the highly desired abilities to characterize ideal performance, identify sources of inefficiency, and quantify the benefits of potential optimizations. Such deficiencies have led to slow characterization/optimization cycles that cannot keep up with the fast pace at which new DL models are introduced. We propose Benanza, a sustainable and extensible benchmarking and analysis design that speeds up the characterization/optimization cycle of DL models on GPUs. Benanza consists of four major components: a model processor that parses models into an internal representation, a configurable benchmark generator that automatically generates micro-benchmarks given a set of models, a database of benchmark results, and an analyzer that computes the "lower-bound" latency of DL models using the benchmark data and informs optimizations of model execution. The "lower-bound" latency metric estimates the ideal model execution on a GPU system and serves as the basis for identifying optimization opportunities in frameworks or system libraries. We used Benanza to evaluate 30 ONNX models in MXNet, ONNX Runtime, and PyTorch on 7 GPUs ranging from Kepler to the latest Turing, and identified optimizations in parallel layer execution, cuDNN convolution algorithm selection, framework inefficiency, layer fusion, and using Tensor Cores.

Omer Anjum, Hongyu Gong, Suma Bhat et al.

Finding the right reviewers to assess the quality of conference submissions is a time consuming process for conference organizers. Given the importance of this step, various automated reviewer-paper matching solutions have been proposed to alleviate the burden. Prior approaches, including bag-of-words models and probabilistic topic models have been inadequate to deal with the vocabulary mismatch and partial topic overlap between a paper submission and the reviewer's expertise. Our approach, the common topic model, jointly models the topics common to the submission and the reviewer's profile while relying on abstract topic vectors. Experiments and insightful evaluations on two datasets demonstrate that the proposed method achieves consistent improvements compared to available state-of-the-art implementations of paper-reviewer matching.

Xiaofan Zhang, Haoming Lu, Cong Hao et al.

Object detection and tracking are challenging tasks for resource-constrained embedded systems. While these tasks are among the most compute-intensive tasks from the artificial intelligence domain, they are only allowed to use limited computation and memory resources on embedded devices. In the meanwhile, such resource-constrained implementations are often required to satisfy additional demanding requirements such as real-time response, high-throughput performance, and reliable inference accuracy. To overcome these challenges, we propose SkyNet, a hardware-efficient neural network to deliver the state-of-the-art detection accuracy and speed for embedded systems. Instead of following the common top-down flow for compact DNN (Deep Neural Network) design, SkyNet provides a bottom-up DNN design approach with comprehensive understanding of the hardware constraints at the very beginning to deliver hardware-efficient DNNs. The effectiveness of SkyNet is demonstrated by winning the competitive System Design Contest for low power object detection in the 56th IEEE/ACM Design Automation Conference (DAC-SDC), where our SkyNet significantly outperforms all other 100+ competitors: it delivers 0.731 Intersection over Union (IoU) and 67.33 frames per second (FPS) on a TX2 embedded GPU; and 0.716 IoU and 25.05 FPS on an Ultra96 embedded FPGA. The evaluation of SkyNet is also extended to GOT-10K, a recent large-scale high-diversity benchmark for generic object tracking in the wild. For state-of-the-art object trackers SiamRPN++ and SiamMask, where ResNet-50 is employed as the backbone, implementations using our SkyNet as the backbone DNN are 1.60X and 1.73X faster with better or similar accuracy when running on a 1080Ti GPU, and 37.20X smaller in terms of parameter size for significantly better memory and storage footprint.

Bowen Cheng, Liang-Chieh Chen, Yunchao Wei et al.

Multi-scale context module and single-stage encoder-decoder structure are commonly employed for semantic segmentation. The multi-scale context module refers to the operations to aggregate feature responses from a large spatial extent, while the single-stage encoder-decoder structure encodes the high-level semantic information in the encoder path and recovers the boundary information in the decoder path. In contrast, multi-stage encoder-decoder networks have been widely used in human pose estimation and show superior performance than their single-stage counterpart. However, few efforts have been attempted to bring this effective design to semantic segmentation. In this work, we propose a Semantic Prediction Guidance (SPG) module which learns to re-weight the local features through the guidance from pixel-wise semantic prediction. We find that by carefully re-weighting features across stages, a two-stage encoder-decoder network coupled with our proposed SPG module can significantly outperform its one-stage counterpart with similar parameters and computations. Finally, we report experimental results on the semantic segmentation benchmark Cityscapes, in which our SPGNet attains 81.1% on the test set using only 'fine' annotations.

Cheng Li, Abdul Dakkak, Jinjun Xiong et al.

There has been a rapid proliferation of machine learning/deep learning (ML) models and wide adoption of them in many application domains. This has made profiling and characterization of ML model performance an increasingly pressing task for both hardware designers and system providers, as they would like to offer the best possible system to serve ML models with the target latency, throughput, cost, and energy requirements while maximizing resource utilization. Such an endeavor is challenging as the characteristics of an ML model depend on the interplay between the model, framework, system libraries, and the hardware (or the HW/SW stack). Existing profiling tools are disjoint, however, and only focus on profiling within a particular level of the stack, which limits the thoroughness and usefulness of the profiling results. This paper proposes XSP - an across-stack profiling design that gives a holistic and hierarchical view of ML model execution. XSP leverages distributed tracing to aggregate and correlates profile data from different sources. XSP introduces a leveled and iterative measurement approach that accurately captures the latencies at all levels of the HW/SW stack in spite of the profiling overhead. We couple the profiling design with an automated analysis pipeline to systematically analyze 65 state-of-the-art ML models. We demonstrate that XSP provides insights which would be difficult to discern otherwise.

Xiaofan Zhang, Cong Hao, Haoming Lu et al.

Developing artificial intelligence (AI) at the edge is always challenging, since edge devices have limited computation capability and memory resources but need to meet demanding requirements, such as real-time processing, high throughput performance, and high inference accuracy. To overcome these challenges, we propose SkyNet, an extremely lightweight DNN with 12 convolutional (Conv) layers and only 1.82 megabyte (MB) of parameters following a bottom-up DNN design approach. SkyNet is demonstrated in the 56th IEEE/ACM Design Automation Conference System Design Contest (DAC-SDC), a low power object detection challenge in images captured by unmanned aerial vehicles (UAVs). SkyNet won the first place award for both the GPU and FPGA tracks of the contest: we deliver 0.731 Intersection over Union (IoU) and 67.33 frames per second (FPS) on a TX2 GPU and deliver 0.716 IoU and 25.05 FPS on an Ultra96 FPGA.

Omer Anjum, Wen-Mei Hwu, Jinjun Xiong

This study began with a research project, called DISCvR, conducted at the IBM-ILLINOIS Center for Cognitive Computing Systems Reseach. The goal of DISCvR was to build a practical NLP based AI pipeline for document understanding which will help us better understand the computation patterns and requirements of modern computing systems. While building such a prototype, an early use case came to us thanks to the 2017 IEEE/ACM International Symposium on Microarchitecture (MICRO-50) Program Co-chairs, Drs. Hillery Hunter and Jaime Moreno. They asked us if we can perform some data-driven analysis of the past 50 years of MICRO papers and show some interesting historical perspectives on MICRO's 50 years of publication. We learned two important lessons from that experience: (1) building an AI solution to truly understand unstructured data is hard in spite of the many claimed successes in natural language understanding; and (2) providing a data-driven perspective on computer architecture research is a very interesting and fun project. Recently we decided to conduct a more thorough study based on all past papers of International Symposium on Computer Architecture (ISCA) from 1973 to 2018, which resulted this article. We recognize that we have just scratched the surface of natural language understanding of unstructured data, and there are many more aspects that we can improve. But even with our current study, we felt there were enough interesting findings that may be worthwhile to share with the community. Hence we decided to write this article to summarize our findings so far based only on ISCA publications. Our hope is to generate further interests from the community in this topic, and we welcome collaboration from the community to deepen our understanding both of the computer architecture research and of the challenges of NLP-based AI solutions.

Xiaofan Zhang, Cong Hao, Yuhong Li et al.

Developing deep learning models for resource-constrained Internet-of-Things (IoT) devices is challenging, as it is difficult to achieve both good quality of results (QoR), such as DNN model inference accuracy, and quality of service (QoS), such as inference latency, throughput, and power consumption. Existing approaches typically separate the DNN model development step from its deployment on IoT devices, resulting in suboptimal solutions. In this paper, we first introduce a few interesting but counterintuitive observations about such a separate design approach, and empirically show why it may lead to suboptimal designs. Motivated by these observations, we then propose a novel and practical bi-directional co-design approach: a bottom-up DNN model design strategy together with a top-down flow for DNN accelerator design. It enables a joint optimization of both DNN models and their deployment configurations on IoT devices as represented as FPGAs. We demonstrate the effectiveness of the proposed co-design approach on a real-life object detection application using Pynq-Z1 embedded FPGA. Our method obtains the state-of-the-art results on both QoR with high accuracy (IoU) and QoS with high throughput (FPS) and high energy efficiency.

Cheng Li, Abdul Dakkak, Jinjun Xiong et al.

An increasingly complex and diverse collection of Machine Learning (ML) models as well as hardware/software stacks, collectively referred to as "ML artifacts", are being proposed - leading to a diverse landscape of ML. These ML innovations proposed have outpaced researchers' ability to analyze, study and adapt them. This is exacerbated by the complicated and sometimes non-reproducible procedures for ML evaluation. A common practice of sharing ML artifacts is through repositories where artifact authors post ad-hoc code and some documentation, but often fail to reveal critical information for others to reproduce their results. This results in users' inability to compare with artifact authors' claims or adapt the model to his/her own use. This paper discusses common challenges and pitfalls of ML evaluation and benchmarking, which can be used as a guideline for ML model authors when sharing ML artifacts, and for system developers when benchmarking or designing ML systems.

Cong Hao, Xiaofan Zhang, Yuhong Li et al.

While embedded FPGAs are attractive platforms for DNN acceleration on edge-devices due to their low latency and high energy efficiency, the scarcity of resources of edge-scale FPGA devices also makes it challenging for DNN deployment. In this paper, we propose a simultaneous FPGA/DNN co-design methodology with both bottom-up and top-down approaches: a bottom-up hardware-oriented DNN model search for high accuracy, and a top-down FPGA accelerator design considering DNN-specific characteristics. We also build an automatic co-design flow, including an Auto-DNN engine to perform hardware-oriented DNN model search, as well as an Auto-HLS engine to generate synthesizable C code of the FPGA accelerator for explored DNNs. We demonstrate our co-design approach on an object detection task using PYNQ-Z1 FPGA. Results show that our proposed DNN model and accelerator outperform the state-of-the-art FPGA designs in all aspects including Intersection-over-Union (IoU) (6.2% higher), frames per second (FPS) (2.48X higher), power consumption (40% lower), and energy efficiency (2.5X higher). Compared to GPU-based solutions, our designs deliver similar accuracy but consume far less energy.

Hongyu Gong, Suma Bhat, Lingfei Wu et al.

Text style transfer rephrases a text from a source style (e.g., informal) to a target style (e.g., formal) while keeping its original meaning. Despite the success existing works have achieved using a parallel corpus for the two styles, transferring text style has proven significantly more challenging when there is no parallel training corpus. In this paper, we address this challenge by using a reinforcement-learning-based generator-evaluator architecture. Our generator employs an attention-based encoder-decoder to transfer a sentence from the source style to the target style. Our evaluator is an adversarially trained style discriminator with semantic and syntactic constraints that score the generated sentence for style, meaning preservation, and fluency. Experimental results on two different style transfer tasks (sentiment transfer and formality transfer) show that our model outperforms state-of-the-art approaches. Furthermore, we perform a manual evaluation that demonstrates the effectiveness of the proposed method using subjective metrics of generated text quality.

Abdul Dakkak, Cheng Li, Jinjun Xiong et al.

Machine Learning (ML) and Deep Learning (DL) innovations are being introduced at such a rapid pace that model owners and evaluators are hard-pressed analyzing and studying them. This is exacerbated by the complicated procedures for evaluation. The lack of standard systems and efficient techniques for specifying and provisioning ML/DL evaluation is the main cause of this "pain point". This work discusses common pitfalls for replicating DL model evaluation, and shows that these subtle pitfalls can affect both accuracy and performance. It then proposes a solution to remedy these pitfalls called MLModelScope, a specification for repeatable model evaluation and a runtime to provision and measure experiments. We show that by easing the model specification and evaluation process, MLModelScope facilitates rapid adoption of ML/DL innovations.

Abdul Dakkak, Cheng Li, Simon Garcia de Gonzalo et al.

Deep neural networks (DNNs) have become core computation components within low latency Function as a Service (FaaS) prediction pipelines: including image recognition, object detection, natural language processing, speech synthesis, and personalized recommendation pipelines. Cloud computing, as the de-facto backbone of modern computing infrastructure for both enterprise and consumer applications, has to be able to handle user-defined pipelines of diverse DNN inference workloads while maintaining isolation and latency guarantees, and minimizing resource waste. The current solution for guaranteeing isolation within FaaS is suboptimal -- suffering from "cold start" latency. A major cause of such inefficiency is the need to move large amount of model data within and across servers. We propose TrIMS as a novel solution to address these issues. Our proposed solution consists of a persistent model store across the GPU, CPU, local storage, and cloud storage hierarchy, an efficient resource management layer that provides isolation, and a succinct set of application APIs and container technologies for easy and transparent integration with FaaS, Deep Learning (DL) frameworks, and user code. We demonstrate our solution by interfacing TrIMS with the Apache MXNet framework and demonstrate up to 24x speedup in latency for image classification models and up to 210x speedup for large models. We achieve up to 8x system throughput improvement.

Bowen Cheng, Yunchao Wei, Jiahui Yu et al.

While training on samples drawn from independent and identical distribution has been a de facto paradigm for optimizing image classification networks, humans learn new concepts in an easy-to-hard manner and on the selected examples progressively. Driven by this fact, we investigate the training paradigms where the samples are not drawn from independent and identical distribution. We propose a data sampling strategy, named Drop-and-Refresh (DaR), motivated by the learning behaviors of humans that selectively drop easy samples and refresh them only periodically. We show in our experiments that the proposed DaR strategy can maintain (and in many cases improve) the predictive accuracy even when the training cost is reduced by 15% on various datasets (CIFAR 10, CIFAR 100 and ImageNet) and with different backbone architectures (ResNets, DenseNets and MobileNets). Furthermore and perhaps more importantly, we find the ImageNet pre-trained models using our DaR sampling strategy achieves better transferability for the downstream tasks including object detection (+0.3 AP), instance segmentation (+0.3 AP), scene parsing (+0.5 mIoU) and human pose estimation (+0.6 AP). Our investigation encourages people to rethink the connections between the sampling strategy for training and the transferability of its learned features for pre-training ImageNet models.