Bilevel optimization problems are receiving increasing attention in machine learning as they provide a natural framework for hyperparameter optimization and meta-learning. A key step to tackle these problems is the efficient computation of the gradient of the upper-level objective (hypergradient). In this work, we study stochastic approximation schemes for the hypergradient, which are important when the lower-level problem is empirical risk minimization on a large dataset. The method that we propose is a stochastic variant of the approximate implicit differentiation approach in (Pedregosa, 2016). We provide bounds for the mean square error of the hypergradient approximation, under the assumption that the lower-level problem is accessible only through a stochastic mapping which is a contraction in expectation. In particular, our main bound is agnostic to the choice of the two stochastic solvers employed by the procedure. We provide numerical experiments to support our theoretical analysis and to show the advantage of using stochastic hypergradients in practice.
Federated Learning (FL) is a new machine learning framework, which enables millions of participants to collaboratively train machine learning model without compromising data privacy and security. Due to the independence and confidentiality of each client, FL does not guarantee that all clients are honest by design, which makes it vulnerable to adversarial attack naturally. In this paper, we focus on dynamic backdoor attacks under FL setting, where the goal of the adversary is to reduce the performance of the model on targeted tasks while maintaining a good performance on the main task, current existing studies are mainly focused on static backdoor attacks, that is the poison pattern injected is unchanged, however, FL is an online learning framework, and adversarial targets can be changed dynamically by attacker, traditional algorithms require learning a new targeted task from scratch, which could be computationally expensive and require a large number of adversarial training examples, to avoid this, we bridge meta-learning and backdoor attacks under FL setting, in which case we can learn a versatile model from previous experiences, and fast adapting to new adversarial tasks with a few of examples. We evaluate our algorithm on different datasets, and demonstrate that our algorithm can achieve good results with respect to dynamic backdoor attacks. To the best of our knowledge, this is the first paper that focus on dynamic backdoor attacks research under FL setting.
Model-Agnostic Meta-Learning (MAML), a popular gradient-based meta-learning framework, assumes that the contribution of each task or instance to the meta-learner is equal. Hence, it fails to address the domain shift between base and novel classes in few-shot learning. In this work, we propose a novel robust meta-learning algorithm, NestedMAML, which learns to assign weights to training tasks or instances. We consider weights as hyper-parameters and iteratively optimize them using a small set of validation tasks set in a nested bi-level optimization approach (in contrast to the standard bi-level optimization in MAML). We then apply NestedMAML in the meta-training stage, which involves (1) several tasks sampled from a distribution different from the meta-test task distribution, or (2) some data samples with noisy labels. Extensive experiments on synthetic and real-world datasets demonstrate that NestedMAML efficiently mitigates the effects of "unwanted" tasks or instances, leading to significant improvement over the state-of-the-art robust meta-learning methods.
Hebbian meta-learning has recently shown promise to solve hard reinforcement learning problems, allowing agents to adapt to some degree to changes in the environment. However, because each synapse in these approaches can learn a very specific learning rule, the ability to generalize to very different situations is likely reduced. We hypothesize that limiting the number of Hebbian learning rules through a "genomic bottleneck" can act as a regularizer leading to better generalization across changes to the environment. We test this hypothesis by decoupling the number of Hebbian learning rules from the number of synapses and systematically varying the number of Hebbian learning rules. The results in this paper suggest that simultaneously learning the Hebbian learning rules and their assignment to synapses is a difficult optimization problem, leading to poor performance in the environments tested. However, parallel research to ours finds that it is indeed possible to reduce the number of learning rules by clustering similar rules together. How to best implement a "genomic bottleneck" algorithm is thus an important research direction that warrants further investigation.
This work is motivated by the recent advances in Deep Neural Networks (DNNs) and their widespread applications in human-machine interfaces. DNNs have been recently used for detecting the intended hand gesture through processing of surface electromyogram (sEMG) signals. The ultimate goal of these approaches is to realize high-performance controllers for prosthetic. However, although DNNs have shown superior accuracy than conventional methods when large amounts of data are available for training, their performance substantially decreases when data are limited. Collecting large datasets for training may be feasible in research laboratories, but it is not a practical approach for real-life applications. Therefore, there is an unmet need for the design of a modern gesture detection technique that relies on minimal training data while providing high accuracy. Here we propose an innovative and novel "Few-Shot Learning" framework based on the formulation of meta-learning, referred to as the FS-HGR, to address this need. Few-shot learning is a variant of domain adaptation with the goal of inferring the required output based on just one or a few training examples. More specifically, the proposed FS-HGR quickly generalizes after seeing very few examples from each class. The proposed approach led to 85.94% classification accuracy on new repetitions with few-shot observation (5-way 5-shot), 81.29% accuracy on new subjects with few-shot observation (5-way 5-shot), and 73.36% accuracy on new gestures with few-shot observation (5-way 5-shot).
The support/query (S/Q) episodic training strategy has been widely used in modern meta-learning algorithms and is believed to improve their generalization ability to test environments. This paper conducts a theoretical investigation of this training strategy on generalization. From a stability perspective, we analyze the generalization error bound of generic meta-learning algorithms trained with such strategy. We show that the S/Q episodic training strategy naturally leads to a counterintuitive generalization bound of $O(1/\sqrt{n})$, which only depends on the task number $n$ but independent of the inner-task sample size $m$. Under the common assumption $m<<n$ for few-shot learning, the bound of $O(1/\sqrt{n})$ implies strong generalization guarantees for modern meta-learning algorithms in the few-shot regime. To further explore the influence of training strategies on generalization, we propose a leave-one-out (LOO) training strategy for meta-learning and compare it with S/Q training. Experiments on standard few-shot regression and classification tasks with popular meta-learning algorithms validate our analysis.
Neural Memory Networks (NMNs) have received increased attention in recent years compared to deep architectures that use a constrained memory. Despite their new appeal, the success of NMNs hinges on the ability of the gradient-based optimiser to perform incremental training of the NMN controllers, determining how to leverage their high capacity for knowledge retrieval. This means that while excellent performance can be achieved when the training data is consistent and well distributed, rare data samples are hard to learn from as the controllers fail to incorporate them effectively during model training. Drawing inspiration from the human cognition process, in particular the utilisation of neuromodulators in the human brain, we propose to decouple the learning process of the NMN controllers to allow them to achieve flexible, rapid adaptation in the presence of new information. This trait is highly beneficial for meta-learning tasks where the memory controllers must quickly grasp abstract concepts in the target domain, and adapt stored knowledge. This allows the NMN controllers to quickly determine which memories are to be retained and which are to be erased, and swiftly adapt their strategy to the new task at hand. Through both quantitative and qualitative evaluations on multiple public benchmarks, including classification and regression tasks, we demonstrate the utility of the proposed approach. Our evaluations not only highlight the ability of the proposed NMN architecture to outperform the current state-of-the-art methods, but also provide insights on how the proposed augmentations help achieve such superior results. In addition, we demonstrate the practical implications of the proposed learning strategy, where the feedback path can be shared among multiple neural memory networks as a mechanism for knowledge sharing.
Most environmental data come from a minority of well-monitored sites. An ongoing challenge in the environmental sciences is transferring knowledge from monitored sites to unmonitored sites. Here, we demonstrate a novel transfer learning framework that accurately predicts depth-specific temperature in unmonitored lakes (targets) by borrowing models from well-monitored lakes (sources). This method, Meta Transfer Learning (MTL), builds a meta-learning model to predict transfer performance from candidate source models to targets using lake attributes and candidates' past performance. We constructed source models at 145 well-monitored lakes using calibrated process-based modeling (PB) and a recently developed approach called process-guided deep learning (PGDL). We applied MTL to either PB or PGDL source models (PB-MTL or PGDL-MTL, respectively) to predict temperatures in 305 target lakes treated as unmonitored in the Upper Midwestern United States. We show significantly improved performance relative to the uncalibrated process-based General Lake Model, where the median RMSE for the target lakes is $2.52^{\circ}C$. PB-MTL yielded a median RMSE of $2.43^{\circ}C$; PGDL-MTL yielded $2.16^{\circ}C$; and a PGDL-MTL ensemble of nine sources per target yielded $1.88^{\circ}C$. For sparsely monitored target lakes, PGDL-MTL often outperformed PGDL models trained on the target lakes themselves. Differences in maximum depth between the source and target were consistently the most important predictors. Our approach readily scales to thousands of lakes in the Midwestern United States, demonstrating that MTL with meaningful predictor variables and high-quality source models is a promising approach for many kinds of unmonitored systems and environmental variables.
Deep pre-trained Transformer models have achieved state-of-the-art results over a variety of natural language processing (NLP) tasks. By learning rich language knowledge with millions of parameters, these models are usually overparameterized and significantly increase the computational overhead in applications. It is intuitive to address this issue by model compression. In this work, we propose a method, called Single-Shot Meta-Pruning, to compress deep pre-trained Transformers before fine-tuning. Specifically, we focus on pruning unnecessary attention heads adaptively for different downstream tasks. To measure the informativeness of attention heads, we train our Single-Shot Meta-Pruner (SMP) with a meta-learning paradigm aiming to maintain the distribution of text representations after pruning. Compared with existing compression methods for pre-trained models, our method can reduce the overhead of both fine-tuning and inference. Experimental results show that our pruner can selectively prune 50% of attention heads with little impact on the performance on downstream tasks and even provide better text representations. The source code will be released in the future.
Target-driven visual navigation is a challenging problem that requires a robot to find the goal using only visual inputs. Many researchers have demonstrated promising results using deep reinforcement learning (deep RL) on various robotic platforms, but typical end-to-end learning is known for its poor extrapolation capability to new scenarios. Therefore, learning a navigation policy for a new robot with a new sensor configuration or a new target still remains a challenging problem. In this paper, we introduce a learning algorithm that enables rapid adaptation to new sensor configurations or target objects with a few shots. We design a policy architecture with latent features between perception and inference networks and quickly adapt the perception network via meta-learning while freezing the inference network. Our experiments show that our algorithm adapts the learned navigation policy with only three shots for unseen situations with different sensor configurations or different target colors. We also analyze the proposed algorithm by investigating various hyperparameters.
Most conventional supervised super-resolution (SR) algorithms assume that low-resolution (LR) data is obtained by downscaling high-resolution (HR) data with a fixed known kernel, but such an assumption often does not hold in real scenarios. Some recent blind SR algorithms have been proposed to estimate different downscaling kernels for each input LR image. However, they suffer from heavy computational overhead, making them infeasible for direct application to videos. In this work, we present DynaVSR, a novel meta-learning-based framework for real-world video SR that enables efficient downscaling model estimation and adaptation to the current input. Specifically, we train a multi-frame downscaling module with various types of synthetic blur kernels, which is seamlessly combined with a video SR network for input-aware adaptation. Experimental results show that DynaVSR consistently improves the performance of the state-of-the-art video SR models by a large margin, with an order of magnitude faster inference time compared to the existing blind SR approaches.
To prevent the leakage of private information while enabling automated machine intelligence, there is an emerging trend to integrate federated learning and Neural Architecture Search (NAS). Although promising as it may seem, the coupling of difficulties from both two tenets makes the algorithm development quite challenging. In particular, how to efficiently search the optimal neural architecture directly from massive non-iid data of clients in a federated manner remains to be a hard nut to crack. To tackle this challenge, in this paper, by leveraging the advances in proxy-less NAS, we propose a Federated Direct Neural Architecture Search (FDNAS) framework that allows hardware-aware NAS from decentralized non-iid data of clients. To further adapt for various data distributions of clients, inspired by meta-learning, a cluster Federated Direct Neural Architecture Search (CFDNAS) framework is proposed to achieve client-aware NAS, in the sense that each client can learn a tailored deep learning model for its particular data distribution. Extensive experiments on real-world non-iid datasets show state-of-the-art accuracy-efficiency trade-offs for various hardware and data distributions of clients. Our codes will be released publicly upon paper acceptance.
Few-shot image classification is challenging due to the lack of ample samples in each class. Such a challenge becomes even tougher when the number of classes is very large, i.e., the large-class few-shot scenario. In this novel scenario, existing approaches do not perform well because they ignore confusable classes, namely similar classes that are difficult to distinguish from each other. These classes carry more information. In this paper, we propose a biased learning paradigm called Confusable Learning, which focuses more on confusable classes. Our method can be applied to mainstream meta-learning algorithms. Specifically, our method maintains a dynamically updating confusion matrix, which analyzes confusable classes in the dataset. Such a confusion matrix helps meta learners to emphasize on confusable classes. Comprehensive experiments on Omniglot, Fungi, and ImageNet demonstrate the efficacy of our method over state-of-the-art baselines.
Meta-learning automatically infers an inductive bias by observing data from a number of related tasks. The inductive bias is encoded by hyperparameters that determine aspects of the model class or training algorithm, such as initialization or learning rate. Meta-learning assumes that the learning tasks belong to a task environment, and that tasks are drawn from the same task environment both during meta-training and meta-testing. This, however, may not hold true in practice. In this paper, we introduce the problem of transfer meta-learning, in which tasks are drawn from a target task environment during meta-testing that may differ from the source task environment observed during meta-training. Novel information-theoretic upper bounds are obtained on the transfer meta-generalization gap, which measures the difference between the meta-training loss, available at the meta-learner, and the average loss on meta-test data from a new, randomly selected, task in the target task environment. The first bound, on the average transfer meta-generalization gap, captures the meta-environment shift between source and target task environments via the KL divergence between source and target data distributions. The second, PAC-Bayesian bound, and the third, single-draw bound, account for this shift via the log-likelihood ratio between source and target task distributions. Furthermore, two transfer meta-learning solutions are introduced. For the first, termed Empirical Meta-Risk Minimization (EMRM), we derive bounds on the average optimality gap. The second, referred to as Information Meta-Risk Minimization (IMRM), is obtained by minimizing the PAC-Bayesian bound. IMRM is shown via experiments to potentially outperform EMRM.
Neural network has been recognized with its accomplishments on tackling various natural language understanding (NLU) tasks. Methods have been developed to train a robust model to handle multiple tasks to gain a general representation of text. In this paper, we implement the model-agnostic meta-learning (MAML) and Online aware Meta-learning (OML) meta-objective under the continual framework for NLU tasks. We validate our methods on selected SuperGLUE and GLUE benchmark.
Recent works found that fine-tuning and joint training---two popular approaches for transfer learning---do not always improve accuracy on downstream tasks. First, we aim to understand more about when and why fine-tuning and joint training can be suboptimal or even harmful for transfer learning. We design semi-synthetic datasets where the source task can be solved by either source-specific features or transferable features. We observe that (1) pre-training may not have incentive to learn transferable features and (2) joint training may simultaneously learn source-specific features and overfit to the target. Second, to improve over fine-tuning and joint training, we propose Meta Representation Learning (MeRLin) to learn transferable features. MeRLin meta-learns representations by ensuring that a head fit on top of the representations with target training data also performs well on target validation data. We also prove that MeRLin recovers the target ground-truth model with a quadratic neural net parameterization and a source distribution that contains both transferable and source-specific features. On the same distribution, pre-training and joint training provably fail to learn transferable features. MeRLin empirically outperforms previous state-of-the-art transfer learning algorithms on various real-world vision and NLP transfer learning benchmarks.
Joining multiple decision-makers together is a powerful way to obtain more sophisticated decision-making systems, but requires to address the questions of division of labor and specialization. We investigate in how far information constraints in hierarchies of experts not only provide a principled method for regularization but also to enforce specialization. In particular, we devise an information-theoretically motivated on-line learning rule that allows partitioning of the problem space into multiple sub-problems that can be solved by the individual experts. We demonstrate two different ways to apply our method: (i) partitioning problems based on individual data samples and (ii) based on sets of data samples representing tasks. Approach (i) equips the system with the ability to solve complex decision-making problems by finding an optimal combination of local expert decision-makers. Approach (ii) leads to decision-makers specialized in solving families of tasks, which equips the system with the ability to solve meta-learning problems. We show the broad applicability of our approach on a range of problems including classification, regression, density estimation, and reinforcement learning problems, both in the standard machine learning setup and in a meta-learning setting.
Continual learning (CL) studies the problem of learning a sequence of tasks, one at a time, such that the learning of each new task does not lead to the deterioration in performance on the previously seen ones while exploiting previously learned features. This paper presents Bilevel Continual Learning (BiCL), a general framework for continual learning that fuses bilevel optimization and recent advances in meta-learning for deep neural networks. BiCL is able to train both deep discriminative and generative models under the conservative setting of the online continual learning. Experimental results show that BiCL provides competitive performance in terms of accuracy for the current task while reducing the effect of catastrophic forgetting. This is a concurrent work with [1]. We submitted it to AAAI 2020 and IJCAI 2020. Now we put it on the arxiv for record. Different from [1], we also consider continual generative model as well. At the same time, the authors are aware of a recent proposal on bilevel optimization based coreset construction for continual learning [2]. [1] Q. Pham, D. Sahoo, C. Liu, and S. C. Hoi. Bilevel continual learning. arXiv preprint arXiv:2007.15553, 2020. [2] Z. Borsos, M. Mutny, and A. Krause. Coresets via bilevel optimization for continual learning and streaming. arXiv preprint arXiv:2006.03875, 2020
A key problem in the theory of meta-learning is to understand how the task distributions influence transfer risk, the expected error of a meta-learner on a new task drawn from the unknown task distribution. In this paper, focusing on fixed design linear regression with Gaussian noise and a Gaussian task (or parameter) distribution, we give distribution-dependent lower bounds on the transfer risk of any algorithm, while we also show that a novel, weighted version of the so-called biased regularized regression method is able to match these lower bounds up to a fixed constant factor. Notably, the weighting is derived from the covariance of the Gaussian task distribution. Altogether, our results provide a precise characterization of the difficulty of meta-learning in this Gaussian setting. While this problem setting may appear simple, we show that it is rich enough to unify the "parameter sharing" and "representation learning" streams of meta-learning; in particular, representation learning is obtained as the special case when the covariance matrix of the task distribution is unknown. For this case we propose to adopt the EM method, which is shown to enjoy efficient updates in our case. The paper is completed by an empirical study of EM. In particular, our experimental results show that the EM algorithm can attain the lower bound as the number of tasks grows, while the algorithm is also successful in competing with its alternatives when used in a representation learning context.
The generalization capability of machine learning models, which refers to generalizing the knowledge for an "unseen" domain via learning from one or multiple seen domain(s), is of great importance to develop and deploy machine learning applications in the real-world conditions. Domain Generalization (DG) techniques aim to enhance such generalization capability of machine learning models, where the learnt feature representation and the classifier are two crucial factors to improve generalization and make decisions. In this paper, we propose Discriminative Adversarial Domain Generalization (DADG) with meta-learning-based cross-domain validation. Our proposed framework contains two main components that work synergistically to build a domain-generalized DNN model: (i) discriminative adversarial learning, which proactively learns a generalized feature representation on multiple "seen" domains, and (ii) meta-learning based cross-domain validation, which simulates train/test domain shift via applying meta-learning techniques in the training process. In the experimental evaluation, a comprehensive comparison has been made among our proposed approach and other existing approaches on three benchmark datasets. The results shown that DADG consistently outperforms a strong baseline DeepAll, and outperforms the other existing DG algorithms in most of the evaluation cases.