Neural architecture search (NAS) has achieved remarkable results in deep neural network design. Differentiable architecture search converts the search over discrete architectures into a hyperparameter optimization problem which can be solved by gradient descent. However, questions have been raised regarding the effectiveness and generalizability of gradient methods for solving non-convex architecture hyperparameter optimization problems. In this paper, we propose L$^{2}$NAS, which learns to intelligently optimize and update architecture hyperparameters via an actor neural network based on the distribution of high-performing architectures in the search history. We introduce a quantile-driven training procedure which efficiently trains L$^{2}$NAS in an actor-critic framework via continuous-action reinforcement learning. Experiments show that L$^{2}$NAS achieves state-of-the-art results on NAS-Bench-201 benchmark as well as DARTS search space and Once-for-All MobileNetV3 search space. We also show that search policies generated by L$^{2}$NAS are generalizable and transferable across different training datasets with minimal fine-tuning.
In real scenarios, state observations that an agent observes may contain measurement errors or adversarial noises, misleading the agent to take suboptimal actions or even collapse while training. In this paper, we study the training robustness of distributional Reinforcement Learning~(RL), a class of state-of-the-art methods that estimate the whole distribution, as opposed to only the expectation, of the total return. Firstly, we propose State-Noisy Markov Decision Process~(SN-MDP) in the tabular case to incorporate both random and adversarial state observation noises, in which the contraction of both expectation-based and distributional Bellman operators is derived. Beyond SN-MDP with the function approximation, we theoretically characterize the bounded gradient norm of histogram-based distributional loss, accounting for the better training robustness of distribution RL. We also provide stricter convergence conditions of the Temporal-Difference~(TD) learning under more flexible state noises, as well as the sensitivity analysis by the leverage of influence function. Finally, extensive experiments on the suite of games show that distributional RL enjoys better training robustness compared with its expectation-based counterpart across various state observation noises.
Domain adaptation (DA) aims to transfer the knowledge learned from a source domain to an unlabeled target domain. Some recent works tackle source-free domain adaptation (SFDA) where only a source pre-trained model is available for adaptation to the target domain. However, those methods do not consider keeping source performance which is of high practical value in real world applications. In this paper, we propose a new domain adaptation paradigm called Generalized Source-free Domain Adaptation (G-SFDA), where the learned model needs to perform well on both the target and source domains, with only access to current unlabeled target data during adaptation. First, we propose local structure clustering (LSC), aiming to cluster the target features with its semantically similar neighbors, which successfully adapts the model to the target domain in the absence of source data. Second, we propose sparse domain attention (SDA), it produces a binary domain specific attention to activate different feature channels for different domains, meanwhile the domain attention will be utilized to regularize the gradient during adaptation to keep source information. In the experiments, for target performance our method is on par with or better than existing DA and SFDA methods, specifically it achieves state-of-the-art performance (85.4%) on VisDA, and our method works well for all domains after adapting to single or multiple target domains. Code is available in https://github.com/Albert0147/G-SFDA.
Despite the empirical success of neural architecture search (NAS) in deep learning applications, the optimality, reproducibility and cost of NAS schemes remain hard to assess. In this paper, we propose Generative Adversarial NAS (GA-NAS) with theoretically provable convergence guarantees, promoting stability and reproducibility in neural architecture search. Inspired by importance sampling, GA-NAS iteratively fits a generator to previously discovered top architectures, thus increasingly focusing on important parts of a large search space. Furthermore, we propose an efficient adversarial learning approach, where the generator is trained by reinforcement learning based on rewards provided by a discriminator, thus being able to explore the search space without evaluating a large number of architectures. Extensive experiments show that GA-NAS beats the best published results under several cases on three public NAS benchmarks. In the meantime, GA-NAS can handle ad-hoc search constraints and search spaces. We show that GA-NAS can be used to improve already optimized baselines found by other NAS methods, including EfficientNet and ProxylessNAS, in terms of ImageNet accuracy or the number of parameters, in their original search space.
GANs largely increases the potential impact of generative models. Therefore, we propose a novel knowledge transfer method for generative models based on mining the knowledge that is most beneficial to a specific target domain, either from a single or multiple pretrained GANs. This is done using a miner network that identifies which part of the generative distribution of each pretrained GAN outputs samples closest to the target domain. Mining effectively steers GAN sampling towards suitable regions of the latent space, which facilitates the posterior finetuning and avoids pathologies of other methods, such as mode collapse and lack of flexibility. Furthermore, to prevent overfitting on small target domains, we introduce sparse subnetwork selection, that restricts the set of trainable neurons to those that are relevant for the target dataset. We perform comprehensive experiments on several challenging datasets using various GAN architectures (BigGAN, Progressive GAN, and StyleGAN) and show that the proposed method, called MineGAN, effectively transfers knowledge to domains with few target images, outperforming existing methods. In addition, MineGAN can successfully transfer knowledge from multiple pretrained GANs.
Unsupervised domain adaptation (UDA) aims to transfer the knowledge learned from labeled source domain to unlabeled target domain. Existing UDA methods require access to the data from the source domain, during adaptation to the target domain, which may not be feasible in some real-world situations. In this paper, we address Source-free Unsupervised Domain Adaptation (SFUDA), where the model has no access to any source data during the adaptation period. We propose a novel framework named BAIT to tackle SFUDA. Specifically, we first train the model on source domain. With the source-specific classifier head (referred to as anchor classifier) fixed, we further introduce a new learnable classifier head (referred to as bait classifier), which is initialized by the anchor classifier. When adapting the source model to the target domain, the source data are no more accessible and the bait classifier aims to push the target features towards the right side of the decision boundary of the anchor classifier, thus achieving the feature alignment. Experiment results show that proposed BAIT achieves state-of-the-art performance compared with existing normal UDA methods and several SFUDA methods.
Deep neural networks have recently become a popular solution to keyword spotting systems, which enable the control of smart devices via voice. In this paper, we apply neural architecture search to search for convolutional neural network models that can help boost the performance of keyword spotting based on features extracted from acoustic signals while maintaining an acceptable memory footprint. Specifically, we use differentiable architecture search techniques to search for operators and their connections in a predefined cell search space. The found cells are then scaled up in both depth and width to achieve competitive performance. We evaluated the proposed method on Google's Speech Commands Dataset and achieved a state-of-the-art accuracy of over 97% on the setting of 12-class utterance classification commonly reported in the literature.
Humans are capable of learning new tasks without forgetting previous ones, while neural networks fail due to catastrophic forgetting between new and previously-learned tasks. We consider a class-incremental setting which means that the task-ID is unknown at inference time. The imbalance between old and new classes typically results in a bias of the network towards the newest ones. This imbalance problem can either be addressed by storing exemplars from previous tasks, or by using image replay methods. However, the latter can only be applied to toy datasets since image generation for complex datasets is a hard problem. We propose a solution to the imbalance problem based on generative feature replay which does not require any exemplars. To do this, we split the network into two parts: a feature extractor and a classifier. To prevent forgetting, we combine generative feature replay in the classifier with feature distillation in the feature extractor. Through feature generation, our method reduces the complexity of generative replay and prevents the imbalance problem. Our approach is computationally efficient and scalable to large datasets. Experiments confirm that our approach achieves state-of-the-art results on CIFAR-100 and ImageNet, while requiring only a fraction of the storage needed for exemplar-based continual learning. Code available at \url{https://github.com/xialeiliu/GFR-IL}.
Class-incremental learning of deep networks sequentially increases the number of classes to be classified. During training, the network has only access to data of one task at a time, where each task contains several classes. In this setting, networks suffer from catastrophic forgetting which refers to the drastic drop in performance on previous tasks. The vast majority of methods have studied this scenario for classification networks, where for each new task the classification layer of the network must be augmented with additional weights to make room for the newly added classes. Embedding networks have the advantage that new classes can be naturally included into the network without adding new weights. Therefore, we study incremental learning for embedding networks. In addition, we propose a new method to estimate the drift, called semantic drift, of features and compensate for it without the need of any exemplars. We approximate the drift of previous tasks based on the drift that is experienced by current task data. We perform experiments on fine-grained datasets, CIFAR100 and ImageNet-Subset. We demonstrate that embedding networks suffer significantly less from catastrophic forgetting. We outperform existing methods which do not require exemplars and obtain competitive results compared to methods which store exemplars. Furthermore, we show that our proposed SDC when combined with existing methods to prevent forgetting consistently improves results.
Neural Architecture Search (NAS) is attractive for automatically producing deep networks with excellent performance and acceptable computational costs. The performance of intermediate networks in most of existing NAS algorithms are usually represented by the results evaluated on a small proxy dataset with insufficient training in order to save computational resources and time. Although these representations could help us to distinct some searched architectures, they are still far away from the exact performance or ranking orders of all networks sampled from the given search space. Therefore, we propose to learn a performance predictor for ranking different models in the searching period using few networks pre-trained on the entire dataset. We represent each neural architecture as a feature tensor and use the predictor to further refining the representations of networks in the search space. The resulting performance predictor can be utilized for searching desired architectures without additional evaluation. Experimental results illustrate that, we can only use $0.1\%$ (424 models) of the entire NASBench dataset to construct an accurate predictor for efficiently finding the architecture with $93.90\%$ accuracy ($0.04\%$ top performance in the whole search space), which is about $0.5\%$ higher than that of the state-of-the-art methods.