ACML 2021 Papers

Learning accurate models for monocular depth estimation requires precise depth annotation as e.g. gathered through LiDAR scanners. Because the data acquisition with sensors of this kind is costly and does not scale well in general, less advanced depth sources, such as time-of-flight cameras, are often used instead. However, these sensors provide less reliable signals, resulting in imprecise depth data for training regression models. As shown in idealized environments, the noise produced by commonly used RGB-D sensors violates standard statistical assumptions of regression methods, such as least squares estimation. In this paper, we investigate whether robust regression methods, which are more tolerant toward violations of statistical assumptions, can mitigate the effects of low-quality data. As a viable alternative to established approaches of that kind, we propose the use of so-called superset learning, where the original data is replaced by (less precise but more reliable) set-valued data. To evaluate and compare the methods, we provide an extensive empirical study on common benchmark data for monocular depth estimation. Our results clearly show the superiority of robust variants over conventional regression.

Due to the high efficiency and less weather dependency, autonomous greenhouses provide an ideal solution to meet the increasing demand for fresh food. However, managers are faced with some challenges in finding appropriate control strategies for crop growth, since the decision space of the greenhouse control problem is an astronomical number. Therefore, an intelligent closed-loop control framework is highly desired to generate an automatic control policy. As a powerful tool for optimal control, reinforcement learning (RL) algorithms can surpass human beings’ decision-making and can also be seamlessly integrated into the closed-loop control framework. However, in complex real-world scenarios such as agricultural automation control, where the interaction with the environment is time-consuming and expensive, the application of RL algorithms encounters two main challenges, i.e., sample efficiency and safety. Although model-based RL methods can greatly mitigate the efficiency problem of greenhouse control, the safety problem has not got too much attention. In this paper, we present a model-based robust RL framework for autonomous greenhouse control to meet the sample efficiency and safety challenges. Specifically, our framework introduces an ensemble of environment models to work as a simulator and assist in policy optimization, thereby addressing the low sample efficiency problem. As for the safety concern, we propose a sample dropout module to focus more on worst-case samples, which can help improve the adaptability of the greenhouse planting policy in extreme cases. Experimental results demonstrate that our approach can learn a more effective greenhouse planting policy with better robustness than existing methods.

In reinforcement learning, domain randomisation is a popular technique for learning general policies that are robust to new environments and domain-shifts at deployment. However, naively aggregating information from randomised domains may lead to high variances in gradient estimation and sub-optimal policies. To address this issue, we present a peer-to-peer online distillation strategy for reinforcement learning termed P2PDRL, where multiple learning agents are each assigned to a different environment, and then exchange knowledge through mutual regularisation based on Kullback–Leibler divergence. Our experiments on continuous control tasks show that P2PDRL enables robust learning across a wider randomisation distribution than baselines, and more robust generalisation performance to new environments at testing.

The proper {\em initialization of weights} is crucial for the effective training and fast convergence of {\em deep neural networks} (DNNs). Prior work in this area has mostly focused on the principle of {\em balancing the variance among weights per layer} to maintain stability of (i) the input data propagated forwards through the network, and (ii) the loss gradients propagated backwards, respectively. This prevalent heuristic is however agnostic of dependencies among gradients across the various layers and captures only first-order effects per layer. In this paper, we investigate a {\em unifying approach}, based on approximating and controlling the {\em norm of the layers’ Hessians}, which both generalizes and explains existing initialization schemes such as {\em smooth activation functions}, {\em Dropouts}, and {\em ReLU}.

Image captioning methods with attention mechanism are leading this field, especially models with global and local attention. But there are few conventional models to integrate the relationship information between various regions of the image. In this paper, this kind of relationship features are embedded into the fused attention mechanism to explore the internal visual and semantic relations between different object regions. Besides, to alleviate the exposure bias problem and make the training process more efficient, we combine Generative Adversarial Network with Reinforcement Learning and employ the greedy decoding method to generate a dynamic baseline reward for self-critical training. Finally, experiments on MSCOCO datasets show that the model can generate more accurate and vivid image captioning sentences and perform better in multiple prevailing metrics than the previous advanced models.

Federated Learning (FL) is a privacy-protected learning paradigm, which allows many clients to jointly train a model under the coordination of a server without the local data leakage. In real-world scenarios, data in different clients usually cannot satisfy the independent and identically distributed (i.i.d.) assumption adopted widely in machine learning. Traditionally training a single global model may cause performance degradation and difficulty in ensuring convergence in such a non-i.i.d. case. To handle this case, various models can be trained for each client to capture the personalization in each client. In this paper, we propose a new personalized FL framework, called Personalized Federated Mutual Learning (PFML), to use the non-i.i.d. characteristics to generate specific models for clients. Specifically, the PFML method integrates mutual learning into the local update process in each client to not only improve the performance of both the global and personalized models but also speed up the convergence compared with state-of-the-art methods. Moreover, the proposed PFML method can help maintain the heterogeneity of client models and protect the information of personalized models. Experiments on benchmark datasets show the effectiveness of the proposed PFML model.

Recently, graph neural networks (GNNs) have become an important and active research direction in deep learning. It is worth noting that most of the existing GNN-based methods learn graph representations within the Euclidean vector space. Beyond the Euclidean space, learning representation and embeddings in hyper-complex space have also shown to be a promising and effective approach. To this end, we propose Quaternion Graph Neural Networks (QGNN) to learn graph representations within the Quaternion space. As demonstrated, the Quaternion space, a hyper-complex vector space, provides highly meaningful computations and analogical calculus through Hamilton product compared to the Euclidean and complex vector spaces. Our QGNN obtains state-of-the-art results on a range of benchmark datasets for graph classification and node classification. Besides, regarding knowledge graphs, our QGNN-based embedding model achieves state-of-the-art results on three new and challenging benchmark datasets for knowledge graph completion. Our code is available at: \url{https://github.com/daiquocnguyen/QGNN}.

Noisy labeled data is more a norm than a rarity for self-generated content that is continuously published on the web and social media from non-experts. Active querying experts are conventionally adopted to provide labels for the informative samples which don’t have labels, instead of possibly incorrect labels. The new challenge that arises here is how to discern the informative and noisy labels which benefit from expert cleaning. In this paper, we aim to leverage the stringent oracle budget to robustly maximize learning accuracy. We propose a noise-aware active learning framework, QActor, and a novel measure \emph{CENT}, which considers both cross-entropy and entropy to select informative and noisy labels for an expert cleansing. QActor iteratively cleans samples via quality models and actively querying an expert on those noisy yet informative samples. To adapt to learning capacity per iteration, QActor dynamically adjusts the query limit according to the learning loss for each learning iteration. We extensively evaluate different image datasets with noise label ratios ranging between 30% and 60%. Our results show that QActor can nearly match the optimal accuracy achieved using only clean data at the cost of only an additional 10% of ground truth data from the oracle.

Most of the existing Siamese-based trackers treat tracking problem as a parallel task of classification and regression. However, some studies show that the sibling head structure could lead to suboptimal solutions during the network training. Through experiments we find that, without regression, the performance could be equally promising as long as we delicately design the network to suit the training objective. We introduce a novel voting-based classification-only tracking algorithm named Pyramid Correlation based Deep Hough Voting (short for PCDHV), to jointly locate the top-left and bottom-right corners of the target. Specifically we innovatively construct a Pyramid Correlation module to equip the embedded feature with fine-grained local structures and global spatial contexts; The elaborately designed Deep Hough Voting module further take over, integrating long-range dependencies of pixels to perceive corners; In addition, the prevalent discretization gap is simply yet effectively alleviated by increasing the spatial resolution of the feature maps while exploiting channel-space relationships. The algorithm is general, robust and simple. We demonstrate the effectiveness of the module through a series of ablation experiments. Without bells and whistles, our tracker achieves better or comparable performance to the SOTA algorithms on three challenging benchmarks (TrackingNet, GOT-10k and LaSOT) while running at a real-time speed of 80 FPS. Codes and models will be released.

Recently, improving the visual quality of underwater images using deep learning-based methods has drawn considerable attention. Unfortunately, diverse environmental factors (e.g., blue/green color distortion) severely limit their performance in real-world environments. Therefore, strengthening the superiority of the underwater image enhancement method is critical. In this paper, we devote ourselves to develop a new architecture with strong superiority and adaptability. Inspired by the underwater imaging principle, we establish a novel physics-inspired learning model that is easy to realize. A Structure-Aware Texture-Sensitive Network (SATS-Net) is further developed to portray the model. The structure-aware module is responsible for structural information, and the texture-sensitive module is responsible for textural information. Thus, SATS-Net successfully incorporates robust characterization absorbed from the physical principle to achieve strong robustness and adaptability. We conduct extensive experiments to demonstrate that SATS-Net outperforms existing advanced techniques in various real-world underwater environments.

Network structured data is ubiquitous in natural and social science applications. Graph Convolutional Neural Network (GCN) has attracted significant attention recently due to its success in representing, modeling, and predicting large-scale network data. Various types of graph convolutional filters were proposed to process graph signals to boost the performance of graph-based semi-supervised learning. This paper introduces a novel spectral learning technique called EigLearn, which uses residual learning to perturb the eigenvalues of the graph filter matrix to optimize its capability. EigLearn is relatively easy to implement, and yet thorough experimental studies reveal that it is more effective and efficient than the prior works on the specific issue, such as LanczosNet and FisherGCN. EigLearn only perturbs a small number of eigenvalues and does not require a complete eigendecomposition. Our investigation shows that EigLearn reaches the maximal performance improvement by perturbing about 30 to 40 eigenvalues, and the EigLearn-based GCN has comparable efficiency as the standard GCN. Furthermore, EigLearn bears a clear explanation in the spectral domain of the graph filter and shows aggregation effects in performance improvement when coupled with different graph filters. Hence, we anticipate that EigLearn may serve as a useful neural unit in various graph-involved neural net architectures.

Solving a bilevel optimization problem is at the core of several machine learning problems such as hyperparameter tuning, data denoising, meta- and few-shot learning, and trainingdata poisoning. Different from simultaneous or multi-objective optimization, the steepest descent direction for minimizing the upper-level cost in a bilevel problem requires the inverse of the Hessian of the lower-level cost. In this work, we propose a novel algorithm for solving bilevel optimization problems based on the classical penalty function approach. Our method avoids computing the Hessian inverse and can handle constrained bilevel problems easily. We prove the convergence of the method under mild conditions and show that the exact hypergradient is obtained asymptotically. Our method’s simplicity and small space and time complexities enable us to effectively solve large-scale bilevel problems involving deep neural networks. We present results on data denoising, few-shot learning, and training-data poisoning problems in a large-scale setting. Our results show that our approach outperforms or is comparable to previously proposed methods based on automatic differentiation and approximate inversion in terms of accuracy, run-time, and convergence speed

Urban planners and policy makers face the challenge of creating livable and enjoyable citiesfor larger populations in much denser urban conditions. While the urban microclimate holdsa key role in defining the quality of urban spaces today and in the future, the integrationof wind microclimate assessment in early urban design and planning processes remains achallenge due to the complexity and high computational expense of computational fluiddynamics (CFD) simulations. This work develops a data-driven workflow for real-timepedestrian wind comfort estimation in complex urban environments which may enabledesigners, policy makers and city residents to make informed decisions about mobility,health, and energy choices. We use a conditional generative adversarial network (cGAN)architecture to reduce the computational computation while maintaining high confidencelevels and interpretability, adequate representation of urban complexity, and suitability forpedestrian comfort estimation. We demonstrate high quality wind field approximationswhile reducing computation time from days to seconds.

In federated learning, heterogeneity among the clients’ local datasets results in large variations in the number of local updates performed by each client in a communication round. Simply aggregating such local models into a global model will confine the capacity of the system, that is, the single global model will be restricted from delivering good performance on each client’s task. This paper provides a general framework to analyze the convergence of personalized federated learning algorithms. It subsumes previously proposed methods and provides a principled understanding of the computational guarantees. Using insights from this analysis, we propose PFedAtt, a personalized federated learning method that incorporates attention-based grouping to facilitate similar clients’ collaborations. Theoretically, we provide the convergence guarantee for the algorithm, and empirical experiments corroborate the competitive performance of PFedAtt on heterogeneous clients.

A large scale human-labeled dataset plays an important role in creating high quality deep learning models. In this paper we present text annotation for Open Images V5 dataset. To our knowledge it is the largest among publicly available manually created text annotations. Having this annotation we trained a simple Mask-RCNN-based network, referred as Yet Another Mask Text Spotter (YAMTS), which achieves competitive performance or even outperforms current state-of-the-art approaches in some cases on ICDAR 2013, ICDAR 2015 and {Total-Text} datasets. Code for text spotting model available online at: \url{https://github.com/openvinotoolkit/training_extensions}. The model can be exported to OpenVINO{\texttrademark}-format and run on Intel{\textregistered} CPUs.

Determinantal point processes (DPPs) are attractive probabilistic models for expressing item quality and set diversity simultaneously. Although DPPs are widely-applicable to many subset selection tasks, there exist simple small-size probability distributions that any DPP cannot express. To overcome this drawback while keeping good properties of DPPs, in this paper we investigate the expressive power of \emph{convex combinations of DPPs}. We provide upper and lower bounds for the number of DPPs required for \emph{exactly} expressing any probability distribution. For the \emph{approximation} error, we give an upper bound on the Kullback–Leibler divergence $n-\lfloor \log t\rfloor +\epsilon$ for any $\epsilon >0$ of approximate distribution from a given joint probability distribution, where $t$ is the number of DPPs. Our numerical simulation on an online retail dataset empirically verifies that a convex combination of only two DPPs can outperform a nonsymmetric DPP in terms of the Kullback–Leibler divergence. By combining a polynomial number of DPPs, we can express probability distributions induced by bounded-degree pseudo-Boolean functions, which include weighted coverage functions of bounded occurrence.

With the rapid development of web services, various kinds of context data become available in recommender systems to handler the data sparsity problem, called context-aware recommendation (CAR). It is challenging to develop effective approaches to model and exploit these various and heterogeneous data. Recently, heterogeneous information network (HIN) has been adopted to model the context data due to its flexibility in modelling data heterogeneity. However, most of the HIN-based methods, which rely on meta paths or graph embedding to extract features from HINs, cannot fully mine the network structure and semantic features of users and items. Besides, these methods, utilizing the global dataset to learn personalized latent factors, usually suffer individuality loss problem. In this paper, we propose a neural graph filtering method for context-aware recommendation, called NGF. First, we use an unified HIN to model both the users’ feedback information and the context data. Then, we adopt graph filtering to predict aspect-level ratings on a series of independent subgraphs of the unified HIN and feed a deep neural network (DNN) to fuse the predictions for CAR. Concretely, graph filtering is a case-by-case algorithm for personalized recommendation on HINs, which predicts the further behavior by all its similar historical behaviors. We split the unified HIN into many single-aspect networks according to the semantic relations and utilize graph filtering to predict user’s behavior on each subgraphs. The following deep neural network is to fuse the personalized predictions in aspect-level. Extensive experiments on two real-world datasets demonstrate the effectiveness of our neural graph filtering for CAR.

The benchmark datasets for neural architecture search (NAS) have been developed to alleviate the computationally expensive evaluation process and ensure a fair comparison. Recent NAS benchmarks only focus on architecture optimization, although the training hyperparameters affect the obtained model performances. Building the benchmark dataset for joint optimization of architecture and training hyperparameters is essential to further NAS research. The existing NAS-HPO-Bench is a benchmark for joint optimization, but it does not consider the network connectivity design as done in modern NAS algorithms. This paper introduces the first benchmark dataset for joint optimization of network connections and training hyperparameters, which we call NAS-HPO-Bench-II. We collect the performance data of 4K cell-based convolutional neural network architectures trained on the CIFAR-10 dataset with different learning rate and batch size settings, resulting in the data of 192K configurations. The dataset includes the exact data for 12 epoch training. We further build the surrogate model predicting the accuracies after 200 epoch training to provide the performance data of longer training epoch. By analyzing NAS-HPO-Bench-II, we confirm the dependency between architecture and training hyperparameters and the necessity of joint optimization. Finally, we demonstrate the benchmarking of the baseline optimization algorithms using NAS-HPO-Bench-II.

Most existing multi-view clustering methods focus on the global structure or local structure among samples, and few methods focus on the two structures at the same time. In this paper, we propose a Multi-view Latent subspace Clustering based on both Global and Local structure (MLCGL). In this method, a latent embedding representation is learned by exploring the complementary information from different views. In the latent space, not only the global reconstruction relationship but also the local geometric structure among the latent variables are discovered. In this way, a unified affinity graph matrix is constructed in the latent space for different views, which indicates a clear between-class relationship. Meanwhile, a rank constraint is introduced on the Laplacian graph to facilitate the division of samples into the required clusters. In MLCGL, the affinity graph also provides positive feedback to optimize the learned latent representation and contribute to divided it into reasonable clusters. Moreover, we present an alternating iterative optimization scheme to optimize objective functions. Compared with the state-of-art algorithms, MLCGL has achieved excellent experimental performance on several real-world datasets.