The success of neural networks (NNs) in a wide range of applications has led to increased interest in understanding the underlying learning dynamics of these models. In this paper, we go beyond mere descriptions of the learning dynamics by taking a graph perspective and investigating the relationship between the graph structure of NNs and their performance. Specifically, we propose (1) representing the neural network learning process as a time-evolving graph (i.e., a series of static graph snapshots over epochs), (2) capturing the structural changes of the NN during the training phase in a simple temporal summary, and (3) leveraging the structural summary to predict the accuracy of the underlying NN in a classification or regression task. For the dynamic graph representation of NNs, we explore structural representations for fully-connected and convolutional layers, which are key components of powerful NN models. Our analysis shows that a simple summary of graph statistics, such as weighted degree and eigenvector centrality, over just a few epochs can be used to accurately predict the performance of NNs. For example, a weighted degree-based summary of the time-evolving graph that is constructed based on 5 training epochs of the LeNet architecture achieves classification accuracy of over 93%. Our findings are consistent for different NN architectures, including LeNet, VGG, AlexNet and ResNet.

In this paper, we propose a novel pipeline for the 3D reconstruction of the full body from egocentric viewpoints. 3-D reconstruction of the human body from egocentric viewpoints is a challenging task as the view is skewed and the body parts farther from the cameras are occluded. One such example is the view from cameras installed below VR headsets. To achieve this task, we first make use of conditional GANs to translate the egocentric views to full body third-person views. This increases the comprehensibility of the image and caters to occlusions. The generated third-person view is further sent through the 3D reconstruction module that generates a 3D mesh of the body. We also train a network that can take the third person full-body view of the subject and generate the texture maps for applying on the mesh. The generated mesh has fairly realistic body proportions and is fully rigged allowing for further applications such as real-time animation and pose transfer in games. This approach can be key to a new domain of mobile human telepresence.

Artificial intelligence (AI) systems have become increasingly popular in many areas. Nevertheless, AI technologies are still in their developing stages, and many issues need to be addressed. Among those, the reliability of AI systems needs to be demonstrated so that the AI systems can be used with confidence by the general public. In this paper, we provide statistical perspectives on the reliability of AI systems. Different from other considerations, the reliability of AI systems focuses on the time dimension. That is, the system can perform its designed functionality for the intended period. We introduce a so-called SMART statistical framework for AI reliability research, which includes five components: Structure of the system, Metrics of reliability, Analysis of failure causes, Reliability assessment, and Test planning. We review traditional methods in reliability data analysis and software reliability, and discuss how those existing methods can be transformed for reliability modeling and assessment of AI systems. We also describe recent developments in modeling and analysis of AI reliability and outline statistical research challenges in this area, including out-of-distribution detection, the effect of the training set, adversarial attacks, model accuracy, and uncertainty quantification, and discuss how those topics can be related to AI reliability, with illustrative examples. Finally, we discuss data collection and test planning for AI reliability assessment and how to improve system designs for higher AI reliability. The paper closes with some concluding remarks.

We present a neat yet effective recursive operation on vision transformers that can improve parameter utilization without involving additional parameters. This is achieved by sharing weights across depth of transformer networks. The proposed method can obtain a substantial gain ( 2%) simply using naïve recursive operation, requires no special or sophisticated knowledge for designing principles of networks, and introduces minimum computational overhead to the training procedure. To reduce the additional computation caused by recursive operation while maintaining the superior accuracy, we propose an approximating method through multiple sliced group self-attentions across recursive layers which can reduce the cost consumption by 10 30% with minimal performance loss. We call our model Sliced Recursive Transformer (SReT), which is compatible with a broad range of other designs for efficient vision transformers. The flexible scalability has shown great potential for scaling up and constructing extremely deep and large dimensionality vision transformers. The architectures of transformer have achieved substantively breakthroughs recently in the fields of natural language processing (NLP) [Vaswani et al., 2017], computer vision (CV) [Dosovitskiy et al., 2021] and speech [Dong et al., 2018, Wang et al., 2021b]. In the vision area, Dosovitskiy et al. [Dosovitskiy et al., 2021] introduced the vision transformer (ViT) method that split a raw image to a patch sequence as input and directly applied transformer model [Vaswani et al., 2017] for the image classification task. ViT achieved impressive results and has inspired many follow-up works. However, the benefits of a transformer often come with a large computational cost and it is always of great challenge to achieve the optimal trade-off between the accuracy and model complexity. In this work, we are motivated by the following question: How can we improve the parameter utilization of a vision transformer, i.e., the representation ability without increasing the model size? We observe recursive operation as shown in Figure 1 is a simple while effective way to achieve this purpose.

Deep learning-based detection networks have made remarkable progress in autonomous driving systems (ADS). ADS should have reliable performance across a variety of ambient lighting and adverse weather conditions. However, luminance degradation and visual obstructions (such as glare, fog) result in poor quality images by the visual camera which leads to performance decline. To overcome these challenges, we explore the idea of leveraging a different data modality that is disparate yet complementary to the visual data. We propose a comprehensive detection system based on a multimodal-collaborative framework that learns from both RGB (from visual cameras) and thermal (from Infrared cameras) data. This framework trains two networks collaboratively and provides flexibility in learning optimal features of its own modality while also incorporating the complementary knowledge of the other. Our extensive empirical results show that while the improvement in accuracy is nominal, the value lies in challenging and extremely difficult edge cases which is crucial in safety-critical applications such as AD. We provide a holistic view of both merits and limitations of using a thermal imaging system in detection.

Deep neural networks are susceptible to adversarially crafted, small and imperceptible changes in the natural inputs. The most effective defense mechanism against these examples is adversarial training which constructs adversarial examples during training by iterative maximization of loss. The model is then trained to minimize the loss on these constructed examples. This min-max optimization requires more data, larger capacity models, and additional computing resources. It also degrades the standard generalization performance of a model. Can we achieve robustness more efficiently? In this work, we explore this question from the perspective of knowledge transfer. First, we theoretically show the transferability of robustness from an adversarially trained teacher model to a student model with the help of mixup augmentation. Second, we propose a novel robustness transfer method called Mixup-Based Activated Channel Maps (MixACM) Transfer. MixACM transfers robustness from a robust teacher to a student by matching activated channel maps generated without expensive adversarial perturbations. Finally, extensive experiments on multiple datasets and different learning scenarios show our method can transfer robustness while also improving generalization on natural images.

Sparse convolutional neural networks (CNNs) have gained significant traction over the past few years as sparse CNNs can drastically decrease the model size and computations, if exploited befittingly, as compared to their dense counterparts. Sparse CNNs often introduce variations in the layer shapes and sizes, which can prevent dense accelerators from performing well on sparse CNN models. Recently proposed sparse accelerators like SCNN, Eyeriss v2, and SparTen, actively exploit the two-sided or full sparsity, that is, sparsity in both weights and activations, for performance gains. These accelerators, however, either have inefficient micro-architecture, which limits their performance, have no support for non-unit stride convolutions and fully-connected (FC) layers, or suffer massively from systematic load imbalance. To circumvent these issues and support both sparse and dense models, we propose Phantom, a multi-threaded, dynamic, and flexible neural computational core. Phantom uses sparse binary mask representation to actively lookahead into sparse computations, and dynamically schedule its computational threads to maximize the thread utilization and throughput. We also generate a two-dimensional (2D) mesh architecture of Phantom neural computational cores, which we refer to as Phantom-2D accelerator, and propose a novel dataflow that supports all layers of a CNN, including unit and non-unit stride convolutions, and FC layers. In addition, Phantom-2D uses a two-level load balancing strategy to minimize the computational idling, thereby, further improving the hardware utilization. To show support for different types of layers, we evaluate the performance of the Phantom architecture on VGG16 and MobileNet. Our simulations show that the Phantom-2D accelerator attains a performance gain of 12x, 4.1x, 1.98x, and 2.36x, over dense architectures, SCNN, SparTen, and Eyeriss v2, respectively.

Keyword spotting (KWS) has become an indispensable part of many intelligent devices surrounding us, as audio is one of the most efficient ways of interacting with these devices. The accuracy and performance of KWS solutions have been the main focus of the researchers, and thanks to deep learning, substantial progress has been made in this domain. However, as the use of KWS spreads into IoT devices, energy efficiency becomes a very critical requirement besides the performance. We believe KWS solutions that would seek power optimization both in the hardware and the neural network (NN) model architecture are advantageous over many solutions in the literature where mostly the architecture side of the problem is considered. In this work, we designed an optimized KWS CNN model by considering end-to-end energy efficiency for the deployment at MAX78000, an ultra-low-power CNN accelerator. With the combined hardware and model optimization approach, we achieve 96.3\% accuracy for 12 classes while only consuming 251 uJ per inference. We compare our results with other small-footprint neural network-based KWS solutions in the literature. Additionally, we share the energy consumption of our model in power-optimized ARM Cortex-M4F to depict the effectiveness of the chosen hardware for the sake of clarity.

Learning to capture feature relations effectively and efficiently is essential in clickthrough rate (CTR) prediction of modern recommendation systems. Most existing CTR prediction methods model such relations either through tedious manuallydesigned low-order interactions or through inflexible and inefficient high-order interactions, which both require extra DNN modules for implicit interaction modeling. In this paper, we proposed a novel plug-in operation, Dynamic Parameterized Operation (DPO), to learn both explicit and implicit interaction instance-wisely. We showed that the introduction of DPO into DNN modules and Attention modules can respectively benefit two main tasks in CTR prediction, enhancing the adaptiveness of feature-based modeling and improving user behavior modeling with the instance-wise locality. Our Dynamic Parameterized Networks significantly outperforms state-of-the-art methods in the offline experiments on the public dataset and real-world production dataset, together with an online A/B test. Furthermore, the proposed Dynamic Parameterized Networks has been deployed in the ranking system of one of the world's largest e-commerce companies, serving the main traffic of hundreds of millions of active users. Click-through rate (CTR) prediction, which aims to estimate the probability of a user clicking an item, is of great importance in recommendation systems and online advertising systems (Cheng et al., 2016; Guo et al., 2017; Rendle, 2010; Zhou et al., 2018b). Effective feature modeling and user behavior modeling are two critical parts of CTR prediction. Deep neural networks (DNNs) have achieved tremendous success on a variety of CTR prediction methods for feature modeling (Cheng et al., 2016; Guo et al., 2017; Wang et al., 2017). Under the hood, its core component is a linear transformation followed by a nonlinear function, which models weighted interaction between the flattened inputs and contexts by fixed kernels, regardless of the intrinsic decoupling relations from specific contexts (Rendle et al., 2020). This property makes DNN learn interaction in an implicit manner, while limiting its ability to model explicit relation, which is often captured by feature crossing component (Rendle, 2010; Song et al., 2019). Most existing solutions exploit a combinatorial framework (feature crossing component DNN component) to leverage both implicit and explicit feature interactions, which is suboptimal and inefficient (Cheng et al., 2016; Wang et al., 2017). For instance, wide & deep combines a linear module in the wide part for explicit low-order interaction and a DNN module to learn high-order feature interactions. Follow-up works such as Deep & Cross Network (DCN) follows a similar manner by replacing the wide part with more sophistic networks, however, posits restriction to input size which is inflexible.

Researchers at the University of Waterloo have developed a deep learning network that can detect disease biomarkers with a high degree of accuracy. It achieves 98 percent detection of peptide features in a dataset, which means scientists and medical experts would have a greater chance at discovering possible diseases through tissue sample analysis.