Papers

Neural network quantization has become increasingly popular due to efficient memory consumption and faster computation resulting from bitwise operations on the quantized networks. Even though they exhibit excellent generalization capabilities, their robustness properties are not well-understood. In this work, we systematically study the robustness of quantized networks against gradient based adversarial attacks and demonstrate that these quantized models suffer from gradient vanishing issues and show a fake sense of robustness. By attributing gradient vanishing to poor forward-backward signal propagation in the trained network, we introduce a simple temperature scaling approach to mitigate this issue while preserving the decision boundary. Despite being a simple modification to existing gradient based adversarial attacks, experiments on multiple image classification datasets with multiple network architectures demonstrate that our temperature scaled attacks obtain near-perfect success rate on quantized networks while outperforming original attacks on adversarially trained models as well as floating-point networks.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Deep learning techniques for separating audio into different sound sources face several challenges. Standard architectures require training separate models for different types of audio sources. Although some universal separators employ a single model to target multiple sources, they have difficulty generalizing to unseen sources. In this paper, we propose a three-component pipeline to train a universal audio source separator from a large, but weakly-labeled dataset: AudioSet. First, we propose a transformer-based sound event detection system for processing weakly-labeled training data. Second, we devise a query-based audio separation model that leverages this data for model training. Third, we design a latent embedding processor to encode queries that specify audio targets for separation, allowing for zero-shot generalization. Our approach uses a single model for source separation of multiple sound types, and relies solely on weakly-labeled data for training. In addition, the proposed audio separator can be used in a zero-shot setting, learning to separate types of audio sources that were never seen in training. To evaluate the separation performance, we test our model on MUSDB18, while training on the disjoint AudioSet. We further verify the zero-shot performance by conducting another experiment on audio source types that are held-out from training. The model achieves comparable Source-to-Distortion Ratio (SDR) performance to current supervised models in both cases.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 3 @ blue 3, poster session 3, poster session 8, oral session 3

Unsupervised domain adaptation (UDA) has been highly successful in transferring knowledge acquired from a label-rich source domain to a label-scarce target domain. Open-set domain adaptation (open-set DA) and universal domain adaptation (UniDA) have been proposed as solutions to the problem concerning the presence of additional novel categories in the target domain. Existing open-set DA and UniDA approaches treat all novel categories as one unified unknown class and attempt to detect this unknown class during the training process. However, the features of the novel categories learned by these methods are not discriminative. This limits the applicability of UDA in the further classification of these novel categories into their original categories, rather than assigning them to a single unified class. In this paper, we propose a self-labeling framework to cluster all target samples, including those in the ``unknown'' categories. We train the network to learn the representations of target samples via self-supervised learning (SSL) and to identify the seen and unseen (novel) target-sample categories simultaneously by maximizing the mutual information between labels and input data. We evaluated our approach under different DA settings and concluded that our method generally outperformed existing ones by a wide margin.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

In federated learning (FL), a server determines a global learning model by aggregating the local learning models of clients, and the determined global model is broadcast to all the clients. However, the global learning model can significantly deteriorate if a Byzantine attacker transmits malicious learning models trained with incorrectly labeled data. We propose a Byzantine-robust FL algorithm that, by employing a consensus confirmation method, can reduce the success probability of Byzantine attacks. After aggregating the local models from clients, the proposed FL server validates the global model candidate by sending the global model candidate to a set of randomly selected FL clients and asking them to perform local validation with their local data. If most of the validation is positive, the global model is confirmed and broadcast to all the clients. We compare the performance of the proposed FL against Byzantine attacks with that of existing FL algorithms analytically and empirically.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Federated learning (FL) enables distribution of machine learning workloads from the cloud to resource-limited edge devices. Unfortunately, current deep networks remain not only too compute-heavy for inference and training on edge devices, but also too large for communicating updates over bandwidth-constrained networks. In this paper, we develop, implement, and experimentally validate a novel FL framework termed Federated Dynamic Sparse Training (FedDST) by which complex neural networks can be deployed and trained with substantially improved efficiency in both on-device computation and in-network communication. At the core of FedDST is a dynamic process that extracts and trains sparse sub-networks from the target full network. With this scheme, "two birds are killed with one stone:'' instead of full models, each client performs efficient training of its own sparse networks, and only sparse networks are transmitted between devices and the cloud. Furthermore, our results reveal that the dynamic sparsity during FL training more flexibly accommodates local heterogeneity in FL agents than the fixed, shared sparse masks. Moreover, dynamic sparsity naturally introduces an "in-time self-ensembling effect'' into the training dynamics, and improves the FL performance even over dense training. In a realistic and challenging non i.i.d. FL setting, FedDST consistently outperforms competing algorithms in our experiments: for instance, at any fixed upload data cap on non-iid CIFAR-10, it gains an impressive accuracy advantage of 10% over FedAvgM when given the same upload data cap; the accuracy gap remains 3% even when FedAvgM is given 2 times the upload data cap, further demonstrating efficacy of FedDST. Code is available at: https://github.com/bibikar/feddst.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Modern Convolutional Neural Network (CNN) architectures, despite their superiority in solving various problems, are generally too large to be deployed on resource constrained edge devices. In this paper, we reduce memory usage and floating-point operations required by convolutional layers in CNNs. We compress these layers by generalizing the Kronecker Product Decomposition to apply to multidimensional tensors, leading to the Generalized Kronecker Product Decomposition} (GKPD). Our approach yields a plug-and-play module that can be used as a drop-in replacement for any convolutional layer. Experimental results for image classification on CIFAR-10 and ImageNet datasets using ResNet, MobileNetv2 and SeNet architectures substantiate the effectiveness of our proposed approach. We find that GKPD outperforms state-of-the-art decomposition methods including Tensor-Train and Tensor-Ring as well as other relevant compression methods such as pruning and knowledge distillation.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

When crowdsourced workers perform annotation tasks in an unfamiliar domain, their accuracy will dramatically decline due to the lack of expertise. Transferring knowledge from relevant domains can form a better representation for instances, which benefits the estimation of workers' expertise in truth inference models. However, existing knowledge transfer processes for crowdsourcing require a considerable number of well-collected instances in source domains. This paper proposes a novel truth inference model for crowdsourcing, where (meta-)knowledge is transferred by meta-learning and used in the estimation of workers' expertise. Our preliminary experiments demonstrate that the meta-knowledge transfer significantly reduces instances in source domains and increases the accuracy of truth inference.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

The 2:4 structured sparsity pattern released by NVIDIA Ampere architecture, requiring four consecutive values containing at least two zeros, enables doubling math throughput for matrix multiplications. Recent works mainly focus on inference speedup via 2:4 sparsity while training acceleration has been largely overwhelmed where backpropagation consumes around 70% of the training time. However, unlike inference, training speedup with structured pruning is nontrivial due to the need to maintain the fidelity of gradients and reduce the additional overhead of performing 2:4 sparsity online. For the first time, this article proposes fully sparse training (FST) where `fully' indicates that ALL matrix multiplications in forward/backward propagation are structurally pruned while maintaining accuracy. To this end, we begin with saliency analysis, investigating the sensitivity of different sparse objects to structured pruning. Based on the observation of spatial similarity among activations, we propose pruning activations with fixed 2:4 masks. Moreover, an Information Restoration block is proposed to retrieve the lost information, which can be implemented by efficient gradient-shift operation. Evaluation of accuracy and efficiency shows that we can achieve 2× training acceleration with negligible accuracy degradation on challenging large-scale classification and detection tasks.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Federated learning (FL) and split learning (SL) are two popular distributed machine learning approaches. Both follow a model-to-data scenario; clients train and test machine learning models without sharing raw data. SL provides better model privacy than FL due to the machine learning model architecture split between clients and the server. Moreover, the split model makes SL a better option for resource-constrained environments. However, SL performs slower than FL due to the relay-based training across multiple clients. In this regard, this paper presents a novel approach, named splitfed learning (SFL), that amalgamates the two approaches eliminating their inherent drawbacks, along with a refined architectural configuration incorporating differential privacy and PixelDP to enhance data privacy and model robustness. Our analysis and empirical results demonstrate that (pure) SFL provides similar communication efficiency and test accuracy as SL while significantly decreasing its computation time per global epoch than in SL for multiple clients. Furthermore, as in SL, its communication efficiency over FL improves with the number of clients. Besides, the performance of SFL with privacy and robustness measures is further evaluated under extended experimental settings.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 8 @ blue 3, poster session 3, poster session 8, oral session 8

Data-free knowledge distillation (DFKD) has recently been attracting increasing attention from research communities, attributed to its capability of compressing a model only using synthetic data. Despite the encouraging results achieved, state-of-the-art DFKD methods still suffer from the inefficiency of data synthesis, making the data-free training process extremely time-consuming and thus inapplicable for large-scale tasks. In this work, we introduce an efficacious scheme, termed as FastDFKD, that allows us to accelerate DFKD by a factor of orders of magnitude. At the heart of our approach is a novel strategy to reuse the shared common features in training data so as to synthesize different data instances. Unlike prior methods that optimize a set of data independently, we propose to learn a meta-synthesizer that seeks common features as the initialization for the fast data synthesis. As a result, FastDFKD achieves data synthesis within only a few steps, significantly enhancing the efficiency of data-free training. Experiments over CIFAR, NYUv2 and ImageNet demonstrate that the proposed FastDFKD achieves 10$\times$ and even 100$\times$ acceleration while preserving performances on par with the state of the art.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Batch Normalization (BN) as an important component assists Deep Neural Networks achieving promising performance for extensive learning tasks by scaling distribution of feature representations within mini-batches. However, the application of BN suffers from performance degradation under the scenario of Unsupervised Domain Adaptation (UDA), since the estimated statistics fail to concurrently describe two different domains. In this paper, we develop a novel normalization technique, named Collaborative Normalization (CoN), for eliminating domain discrepancy and accelerating the model training of neural networks for UDA. Unlike typical strategies only exploiting domain-specific statistics during normalization, our CoN excavates cross-domain knowledge and simultaneously scales features from various domains by mimicking the merits of collaborative representation. Our CoN can be easily plugged into popular neural network backbones for cross-domain learning. On the one hand, \textbf{theoretical analysis} guarantees that models with CoN promote discriminability of feature representations and accelerate convergence rate; on the other hand, \textbf{empirical study} verifies that replacing BN with CoN in popular network backbones effectively improves classification accuracy in most learning tasks across three cross-domain visual benchmarks.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Federated learning (FL) is a popular technique to train machine learning (ML) models with decentralized data. Extensive works have studied the performance of the global model; however, it is still unclear how the training process affects the final test accuracy. Exacerbating this problem is the fact that FL executions differ significantly from traditional ML with heterogeneous data characteristics across clients, involving more hyperparameters. In this work, we show that the final test accuracy of FL is dramatically affected by the early phase of the training process, i.e., FL exhibits critical learning periods, in which small gradient errors can have irrecoverable impact on the final test accuracy. To further explain this phenomenon, we generalize the trace of the Fisher Information Matrix (FIM) to FL and define a new notation called FedFIM, a quantity reflecting the local curvature of each clients from the beginning of the training in FL. Our findings suggest that the initial learning phase plays a critical role in understanding the FL performance. This is in contrast to many existing works which generally do not connect the final accuracy of FL to the early phase training. Finally, seizing critical learning periods in FL is of independent interest and could be useful for other problems such as the choices of hyperparameters such as the number of client selected per round, batch size, and more, so as to improve the performance of FL training and testing.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Decentralized stochastic gradient descent methods have attracted increasing interest in recent years. Numerous methods have been proposed for the nonconvex finite-sum optimization problem. However, existing methods have a large sample complexity, which slows down the empirical convergence speed. To address this issue, in this paper, we proposed a novel decentralized stochastic gradient descent method for the nonconvex finite-sum optimization problem, which enjoys a better sample and communication complexity than existing methods. Importantly, these complexities are independent of the network topology of the decentralized training system. To the best of our knowledge, our work is the first one achieving such favorable sample and communication complexities. Finally, we have conducted extensive experiments and the experimental results have confirmed the superior performance of our proposed method.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

We introduce NaturalInversion, a novel model inversion-based method to synthesize images that agrees well with the original data distribution without using real data. In NaturalInversion, we propose: (1) a Feature Transfer Pyramid which uses enhanced image prior of the original data by combining the multi-scale feature maps extracted from the pre-trained classifier, (2) a one-to-one approach generative model where only one batch of images are synthesized by one generator to bring the non-linearity to optimization and to ease the overall optimizing process, (3) learnable Adaptive Channel Scaling parameters which are end-to-end trained to scale the output image channel to utilize the original image prior further. With our NaturalInversion, we synthesize images from classifiers trained on CIFAR-10/100 and show that our images are more consistent with original data distribution than prior works by visualization and additional analysis. Furthermore, our synthesized images outperform prior works on various applications such as knowledge distillation and pruning, demonstrating the effectiveness of our proposed method.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Optimizing the combustion efficiency of a thermal power generating unit (TPGU) is a highly challenging and critical task in the energy industry. We develop a new data-driven AI system, namely DeepThermal, to optimize the combustion control strategy for TPGUs. At its core, is a new model-based offline reinforcement learning (RL) framework, called MORE, which leverages historical operational data of a TGPU to solve a highly complex constrained Markov decision process problem via purely offline training. In DeepThermal, we first learn a data-driven combustion process simulator from the offline dataset. The RL agent of MORE is then trained by combining real historical data as well as carefully filtered and processed simulation data through a novel restrictive exploration scheme. DeepThermal has been successfully deployed in four large coal-fired thermal power plants in China. Real-world experiments show that DeepThermal effectively improves the combustion efficiency of TPGUs. We also report the superior performance of MORE by comparing with the state-of-the-art algorithms on the standard offline RL benchmarks.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Heterogeneity across clients in federated learning (FL) usually hinders the optimization convergence and generalization performance when the aggregation of clients' knowledge occurs in the gradient space. For example, clients may differ in terms of data distribution, network latency, input/output space, and/or model architecture, which can easily lead to the misalignment of their local gradients. To improve the tolerance to heterogeneity, we propose a novel federated prototype learning (FedProto) framework in which the clients and server communicate the abstract class prototypes instead of the gradients. FedProto aggregates the local prototypes collected from different clients, and then sends the global prototypes back to all clients to regularize the training of local models. The training on each client aims to minimize the classification error on the local data while keeping the resulting local prototypes sufficiently close to the corresponding global ones. Moreover, we provide a theoretical analysis to the convergence rate of FedProto under non-convex objectives. In experiments, we propose a benchmark setting tailored for heterogeneous FL, with FedProto outperforming several recent FL approaches on multiple datasets.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Training deep neural networks (DNNs) for meaningful differential privacy (DP) guarantees severely degrades model utility. In this paper, we demonstrate that the architecture of DNNs has a significant impact on model utility in the context of private deep learning, whereas its effect is largely unexplored in previous studies. In light of this missing, we propose the very first framework that employs neural architecture search to automatic model design for private deep learning, dubbed as DPNAS. To integrate private learning with architecture search, a DP-aware approach is introduced for training candidate models composed on a delicately defined novel search space. We empirically certify the effectiveness of the proposed framework. The searched model DPNASNet achieves state-of-the-art privacy/utility trade-offs, e.g., for the privacy budget of $(\epsilon, \delta)=(3, 1\times10^{-5})$, our model obtains test accuracy of $98.57\%$ on MNIST, $88.09\%$ on FashionMNIST, and $68.33\%$ on CIFAR-10. Furthermore, by studying the generated architectures, we provide several intriguing findings of designing private-learning-friendly DNNs, which can shed new light on model design for deep learning with differential privacy.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

Scene graph in a video conveys a wealth of information about objects and their relationships in the scene, thus benefiting many downstream tasks such as video captioning and visual question answering. Existing methods of scene graph generation require large-scale training videos annotated with objects and relationships in each frame to learn a powerful model. However, such comprehensive annotation is time-consuming and labor-intensive. On the other hand, it is much easier and less cost to annotate images with scene graphs, so we investigate leveraging annotated images to facilitate training a scene graph generation model for unannotated videos, namely image-to-video scene graph generation. This task presents two challenges: 1) infer unseen dynamic relationships in videos from static relationships in images due to the absence of motion information in images; 2) adapt objects and static relationships from images to video frames due to the domain shift between them. To address the first challenge, we exploit external commonsense knowledge to infer the unseen dynamic relationship from the temporal evolution of static relationships. We tackle the second challenge by hierarchical adversarial learning to reduce the data distribution discrepancy between images and video frames. Extensive experiment results on two benchmark video datasets demonstrate the effectiveness of our method.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 3 @ blue 3, poster session 3, poster session 8, oral session 3

Knowledge Distillation (KD), which is an effective model compression and acceleration technique, has been successfully applied to graph neural networks (GNNs) recently. Existing approaches utilize a single GNN model as the teacher to distill knowledge. However, we notice that GNN models with different number of layers demonstrate different classification abilities on nodes with different degrees. On the one hand, for nodes with high degrees, their local structures are dense and complex, hence more message passing is needed. Therefore, GNN models with more layers perform better. On the other hand, for nodes with low degrees, whose local structures are relatively sparse and simple, the repeated message passing can easily lead to over-smoothing. Thus, GNN models with less layers are more suitable. However, existing single-teacher GNN knowledge distillation approaches which are based on a single GNN model, are sub-optimal. To this end, we propose a novel approach to distill multi-scale knowledge, which learns from multiple GNN teacher models with different number of layers to capture the topological semantic at different scales. Instead of learning from the teacher models equally, the proposed method automatically assigns proper weights for each teacher model via an attention mechanism which enables the student to select teachers for different local structures. Extensive experiments are conducted to evaluate the proposed method on four public datasets. The experimental results demonstrate the superiority of our proposed method over state-of-the-art methods.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 8 @ blue 3, poster session 3, poster session 8, oral session 8

Deep representation learning has succeeded in several fields. However, pre-trained deep representations are usually biased and make downstream models sensitive to different attributes. In this work, we propose a post-processing unsupervised deep representation debiasing algorithm, DeepMinMax, which can obtain unbiased representations directly from pre-trained representations without re-training or fine-tuning the entire model. The experimental results on synthetic and real-world datasets indicate that DeepMinMax outperforms the existing state-of-the-art algorithms on downstream tasks.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

We consider a distributed learning setting where strategic users are incentivized by a fusion center, to train a learning model based on local data. The users are not obliged to provide their true gradient updates and the fusion center is not capable of validating the authenticity of reported updates. Thus motivated, we formulate the interactions between the fusion center and the users as repeated games, manifesting an under-explored interplay between machine learning and game theory. We then develop an incentive mechanism for the fusion center based on a joint gradient estimation and user action classification scheme, and study its impact on the convergence performance of distributed learning. Further, we devise adaptive zero-determinant (ZD) strategies, thereby generalizing the classical ZD strategies to the repeated games with time-varying stochastic errors. Theoretical and empirical analysis show that the fusion center can incentivize the strategic users to cooperate and report informative gradient updates, thus ensuring the convergence.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 8 @ blue 3, poster session 3, poster session 8, oral session 8

This paper studies a weakly supervised domain adaptation (WSDA) problem, where we only have access to the source domain with noisy labels, from which we need to transfer useful information to the unlabeled target domain. Although there have been a few studies on this problem, most of them only exploit unidirectional relationships from the source domain to the target domain. In this paper, we propose a universal paradigm called GearNet to exploit bilateral relationships between the two domains. Specifically, we take the two domains as different inputs to train two models alternately, and a symmetrical Kullback-Leibler loss is used for selectively matching the predictions of the two models in the same domain. This interactive learning schema enables implicit label noise canceling and exploit correlations between the source and target domains. Therefore, our GearNet has the great potential to boost the performance of a wide range of existing WSDA methods. Comprehensive experimental results show that the performance of existing methods can be significantly improved by equipping with our GearNet.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8

There has been significant recent interest in quantum neural networks (QNNs), along with their applications in diverse domains. Current solutions for QNNs pose significant challenges concerning their scalability, ensuring that the postulates of quantum mechanics are satisfied and that the networks are physically realizable. The exponential state space of QNNs poses significant challenges for the scalability of training procedures. The no-cloning principle prohibits making multiple copies of training samples, and the measurement postulates lead to non-deterministic loss functions. Consequently, the physical realizability and efficiency of existing approaches that rely on repeated measurement of several copies of each sample for training QNNs are unclear. This paper presents a new model for QNNs that relies on band-limited Fourier expansions of transfer functions of quantum perceptrons (QPs) to design scalable training procedures. We augment this training procedure with techniques that eliminate the need for sample replication. We show that our novel stochastic gradient descent (SGD) procedure converges to the true minima in expectation, even in the presence of non-determinism due to quantum measurement. Our solution has a number of important benefits: (i) using QPs with concentrated Fourier power spectrum, we show that the training procedure for QNNs can be made scalable; (ii) eliminates the need for resampling, thus staying consistent with the no-cloning rule; and (iii) enhanced data efficiency for the overall training process since each data sample is processed once per epoch. We present a detailed theoretical foundation for our models and methods' scalability, accuracy, and data effectiveness. Finally, we demonstrate the utility of our approach through a series of numerical experiments.

poster session 3 @ blue 3, poster session 8 @ blue 3, oral session 3 @ blue 3, poster session 3, poster session 8, oral session 3

Top-k sparsification method is popular and powerful forreducing the communication cost in Federated Learning(FL). However, according to our experimental observation, it spends most of the total communication cost on the index of the selected parameters (i.e., their position informa-tion), which is inefficient for FL training. To solve this problem, we propose a FL compression algorithm for convolution neural networks (CNNs), called SmartIdx, by extending the traditional Top-k largest variation selection strategy intothe convolution-kernel-based selection, to reduce the proportion of the index in the overall communication cost and thusachieve a high compression ratio. The basic idea of SmartIdx is to improve the 1:1 proportion relationship betweenthe value and index of the parameters to n:1, by regarding the convolution kernel as the basic selecting unit in parameter selection, which can potentially deliver more informationto the parameter server under the limited network traffic. Tothis end, a set of rules are designed for judging which kernel should be selected and the corresponding packaging strategies are also proposed for further improving the compressionratio. Experiments on mainstream CNNs and datasets showthat our proposed SmartIdx performs2.5×−69.2×higher compression ratio than the state-of-the-art FL compression algorithms without degrading model performance.

poster session 3 @ blue 3, poster session 8 @ blue 3, poster session 3, poster session 8