The focus of recent meta-learning research has been on the development of learning algorithms that can quickly adapt to test time tasks with limited data and low computational cost. Few-shot learning is widely used as one of the standard benchmarks in meta-learning. In this work, we show that a simple baseline: learning a supervised or self-supervised representation on the meta-training set, followed by training a linear classifier on top of this representation, outperforms state-of-the-art few-shot learning methods. An additional boost can be achieved through the use of self-distillation. This demonstrates that using a good learned embedding model can be more effective than sophisticated meta-learning algorithms. We believe that our findings motivate a rethinking of few-shot image classification benchmarks and the associated role of meta-learning algorithms. Code is available at: http://github.com/WangYueFt/rfs/.
We present Meta Pseudo Labels, a semi-supervised learning method that achieves a new state-of-the-art top-1 accuracy of 90.2% on ImageNet, which is 1.6% better than the existing state-of-the-art. Like Pseudo Labels, Meta Pseudo Labels has a teacher network to generate pseudo labels on unlabeled data to teach a student network. However, unlike Pseudo Labels where the teacher is fixed, the teacher in Meta Pseudo Labels is constantly adapted by the feedback of the student's performance on the labeled dataset. As a result, the teacher generates better pseudo labels to teach the student. Our code will be available at https://github.com/google-research/google-research/tree/master/meta_pseudo_labels.
Object frequency in the real world often follows a power law, leading to a mismatch between datasets with long-tailed class distributions seen by a machine learning model and our expectation of the model to perform well on all classes. We analyze this mismatch from a domain adaptation point of view. First of all, we connect existing class-balanced methods for long-tailed classification to target shift, a well-studied scenario in domain adaptation. The connection reveals that these methods implicitly assume that the training data and test data share the same class-conditioned distribution, which does not hold in general and especially for the tail classes. While a head class could contain abundant and diverse training examples that well represent the expected data at inference time, the tail classes are often short of representative training data. To this end, we propose to augment the classic class-balanced learning by explicitly estimating the differences between the class-conditioned distributions with a meta-learning approach. We validate our approach with six benchmark datasets and three loss functions.
Meta-learning leverages related source tasks to learn an initialization that can be quickly fine-tuned to a target task with limited labeled examples. However, many popular meta-learning algorithms, such as model-agnostic meta-learning (MAML), only assume access to the target samples for fine-tuning. In this work, we provide a general framework for meta-learning based on weighting the loss of different source tasks, where the weights are allowed to depend on the target samples. In this general setting, we provide upper bounds on the distance of the weighted empirical risk of the source tasks and expected target risk in terms of an integral probability metric (IPM) and Rademacher complexity, which apply to a number of meta-learning settings including MAML and a weighted MAML variant. We then develop a learning algorithm based on minimizing the error bound with respect to an empirical IPM, including a weighted MAML algorithm, $\alpha$-MAML. Finally, we demonstrate empirically on several regression problems that our weighted meta-learning algorithm is able to find better initializations than uniformly-weighted meta-learning algorithms, such as MAML.
In a real-world setting, object instances from new classes can be continuously encountered by object detectors. When existing object detectors are applied to such scenarios, their performance on old classes deteriorates significantly. A few efforts have been reported to address this limitation, all of which apply variants of knowledge distillation to avoid catastrophic forgetting. We note that although distillation helps to retain previous learning, it obstructs fast adaptability to new tasks, which is a critical requirement for incremental learning. In this pursuit, we propose a meta-learning approach that learns to reshape model gradients, such that information across incremental tasks is optimally shared. This ensures a seamless information transfer via a meta-learned gradient preconditioning that minimizes forgetting and maximizes knowledge transfer. In comparison to existing meta-learning methods, our approach is task-agnostic, allows incremental addition of new-classes and scales to high-capacity models for object detection. We evaluate our approach on a variety of incremental learning settings defined on PASCAL-VOC and MS COCO datasets, where our approach performs favourably well against state-of-the-art methods.
Visuomotor control (VMC) is an effective means of achieving basic manipulation tasks such as pushing or pick-and-place from raw images. Conditioning VMC on desired goal states is a promising way of achieving versatile skill primitives. However, common conditioning schemes either rely on task-specific fine tuning - e.g. using one-shot imitation learning (IL) - or on sampling approaches using a forward model of scene dynamics i.e. model-predictive control (MPC), leaving deployability and planning horizon severely limited. In this paper we propose a conditioning scheme which avoids these pitfalls by learning the controller and its conditioning in an end-to-end manner. Our model predicts complex action sequences based directly on a dynamic image representation of the robot motion and the distance to a given target observation. In contrast to related works, this enables our approach to efficiently perform complex manipulation tasks from raw image observations without predefined control primitives or test time demonstrations. We report significant improvements in task success over representative MPC and IL baselines. We also demonstrate our model's generalisation capabilities in challenging, unseen tasks featuring visual noise, cluttered scenes and unseen object geometries.
Graph classification aims to extract accurate information from graph-structured data for classification and is becoming more and more important in graph learning community. Although Graph Neural Networks (GNNs) have been successfully applied to graph classification tasks, most of them overlook the scarcity of labeled graph data in many applications. For example, in bioinformatics, obtaining protein graph labels usually needs laborious experiments. Recently, few-shot learning has been explored to alleviate this problem with only given a few labeled graph samples of test classes. The shared sub-structures between training classes and test classes are essential in few-shot graph classification. Exiting methods assume that the test classes belong to the same set of super-classes clustered from training classes. However, according to our observations, the label spaces of training classes and test classes usually do not overlap in real-world scenario. As a result, the existing methods don't well capture the local structures of unseen test classes. To overcome the limitation, in this paper, we propose a direct method to capture the sub-structures with well initialized meta-learner within a few adaptation steps. More specifically, (1) we propose a novel framework consisting of a graph meta-learner, which uses GNNs based modules for fast adaptation on graph data, and a step controller for the robustness and generalization of meta-learner; (2) we provide quantitative analysis for the framework and give a graph-dependent upper bound of the generalization error based on our framework; (3) the extensive experiments on real-world datasets demonstrate that our framework gets state-of-the-art results on several few-shot graph classification tasks compared to baselines.
We tackle a novel few-shot learning challenge, which we call few-shot semantic edge detection, aiming to localize crisp boundaries of novel categories using only a few labeled samples. We also present a Class-Agnostic Few-shot Edge detection Network (CAFENet) based on meta-learning strategy. CAFENet employs a semantic segmentation module in small-scale to compensate for lack of semantic information in edge labels. The predicted segmentation mask is used to generate an attention map to highlight the target object region, and make the decoder module concentrate on that region. We also propose a new regularization method based on multi-split matching. In meta-training, the metric-learning problem with high-dimensional vectors are divided into small subproblems with low-dimensional sub-vectors. Since there is no existing dataset for few-shot semantic edge detection, we construct two new datasets, FSE-1000 and SBD-$5^i$, and evaluate the performance of the proposed CAFENet on them. Extensive simulation results confirm the performance merits of the techniques adopted in CAFENet.
Few-shot learning aims to handle previously unseen tasks using only a small amount of new training data. In preparing (or meta-training) a few-shot learner, however, massive labeled data are necessary. In the real world, unfortunately, labeled data are expensive and/or scarce. In this work, we propose a few-shot learner that can work well under the semi-supervised setting where a large portion of training data is unlabeled. Our method employs explicit task-conditioning in which unlabeled sample clustering for the current task takes place in a new projection space different from the embedding feature space. The conditioned clustering space is linearly constructed so as to quickly close the gap between the class centroids for the current task and the independent per-class reference vectors meta-trained across tasks. In a more general setting, our method introduces a concept of controlling the degree of task-conditioning for meta-learning: the amount of task-conditioning varies with the number of repetitive updates for the clustering space. Extensive simulation results based on the miniImageNet and tieredImageNet datasets show state-of-the-art semi-supervised few-shot classification performance of the proposed method. Simulation results also indicate that the proposed task-adaptive clustering shows graceful degradation with a growing number of distractor samples, i.e., unlabeled sample images coming from outside the candidate classes.
Face recognition systems are usually faced with unseen domains in real-world applications and show unsatisfactory performance due to their poor generalization. For example, a well-trained model on webface data cannot deal with the ID vs. Spot task in surveillance scenario. In this paper, we aim to learn a generalized model that can directly handle new unseen domains without any model updating. To this end, we propose a novel face recognition method via meta-learning named Meta Face Recognition (MFR). MFR synthesizes the source/target domain shift with a meta-optimization objective, which requires the model to learn effective representations not only on synthesized source domains but also on synthesized target domains. Specifically, we build domain-shift batches through a domain-level sampling strategy and get back-propagated gradients/meta-gradients on synthesized source/target domains by optimizing multi-domain distributions. The gradients and meta-gradients are further combined to update the model to improve generalization. Besides, we propose two benchmarks for generalized face recognition evaluation. Experiments on our benchmarks validate the generalization of our method compared to several baselines and other state-of-the-arts. The proposed benchmarks will be available at https://github.com/cleardusk/MFR.
Plant diseases serve as one of main threats to food security and crop production. It is thus valuable to exploit recent advances of artificial intelligence to assist plant disease diagnosis. One popular approach is to transform this problem as a leaf image classification task, which can be then addressed by the powerful convolutional neural networks (CNNs). However, the performance of CNN-based classification approach depends on a large amount of high-quality manually labeled training data, which are inevitably introduced noise on labels in practice, leading to model overfitting and performance degradation. To overcome this problem, we propose a novel framework that incorporates rectified meta-learning module into common CNN paradigm to train a noise-robust deep network without using extra supervision information. The proposed method enjoys the following merits: i) A rectified meta-learning is designed to pay more attention to unbiased samples, leading to accelerated convergence and improved classification accuracy. ii) Our method is free on assumption of label noise distribution, which works well on various kinds of noise. iii) Our method serves as a plug-and-play module, which can be embedded into any deep models optimized by gradient descent based method. Extensive experiments are conducted to demonstrate the superior performance of our algorithm over the state-of-the-arts.
Detecting rare objects from a few examples is an emerging problem. Prior works show meta-learning is a promising approach. But, fine-tuning techniques have drawn scant attention. We find that fine-tuning only the last layer of existing detectors on rare classes is crucial to the few-shot object detection task. Such a simple approach outperforms the meta-learning methods by roughly 2~20 points on current benchmarks and sometimes even doubles the accuracy of the prior methods. However, the high variance in the few samples often leads to the unreliability of existing benchmarks. We revise the evaluation protocols by sampling multiple groups of training examples to obtain stable comparisons and build new benchmarks based on three datasets: PASCAL VOC, COCO and LVIS. Again, our fine-tuning approach establishes a new state of the art on the revised benchmarks. The code as well as the pretrained models are available at https://github.com/ucbdrive/few-shot-object-detection.
Bayesian statistical inference loses predictive optimality when generative models are misspecified. Working within an existing coherent loss-based generalisation of Bayesian inference, we show existing Modular/Cut-model inference is coherent, and write down a new family of Semi-Modular Inference (SMI) schemes, indexed by an influence parameter, with Bayesian inference and Cut-models as special cases. We give a meta-learning criterion and estimation procedure to choose the inference scheme. This returns Bayesian inference when there is no misspecification. The framework applies naturally to Multi-modular models. Cut-model inference allows directed information flow from well-specified modules to misspecified modules, but not vice versa. An existing alternative power posterior method gives tunable but undirected control of information flow, improving prediction in some settings. In contrast, SMI allows tunable and directed information flow between modules. We illustrate our methods on two standard test cases from the literature and a motivating archaeological data set.
Deep reinforcement learning can learn effective policies for a wide range of tasks, but is notoriously difficult to use due to instability and sensitivity to hyperparameters. The reasons for this remain unclear. When using standard supervised methods (e.g., for bandits), on-policy data collection provides "hard negatives" that correct the model in precisely those states and actions that the policy is likely to visit. We call this phenomenon "corrective feedback." We show that bootstrapping-based Q-learning algorithms do not necessarily benefit from this corrective feedback, and training on the experience collected by the algorithm is not sufficient to correct errors in the Q-function. In fact, Q-learning and related methods can exhibit pathological interactions between the distribution of experience collected by the agent and the policy induced by training on that experience, leading to potential instability, sub-optimal convergence, and poor results when learning from noisy, sparse or delayed rewards. We demonstrate the existence of this problem, both theoretically and empirically. We then show that a specific correction to the data distribution can mitigate this issue. Based on these observations, we propose a new algorithm, DisCor, which computes an approximation to this optimal distribution and uses it to re-weight the transitions used for training, resulting in substantial improvements in a range of challenging RL settings, such as multi-task learning and learning from noisy reward signals. Blog post presenting a summary of this work is available at: https://bair.berkeley.edu/blog/2020/03/16/discor/.
The pattern of Electroencephalogram (EEG) signal differs significantly across different subjects, and poses challenge for EEG classifiers in terms of 1) effectively adapting a learned classifier onto a new subject, 2) retaining knowledge of known subjects after the adaptation. We propose an efficient transfer learning method, named Meta UPdate Strategy (MUPS-EEG), for continuous EEG classification across different subjects. The model learns effective representations with meta update which accelerates adaptation on new subject and mitigate forgetting of knowledge on previous subjects at the same time. The proposed mechanism originates from meta learning and works to 1) find feature representation that is broadly suitable for different subjects, 2) maximizes sensitivity of loss function for fast adaptation on new subject. The method can be applied to all deep learning oriented models. Extensive experiments on two public datasets demonstrate the effectiveness of the proposed model, outperforming current state of the arts by a large margin in terms of both adapting on new subject and retain knowledge of learned subjects.
Building in silico models to predict chemical properties and activities is a crucial step in drug discovery. However, drug discovery projects are often characterized by limited labeled data, hindering the applications of deep learning in this setting. Meanwhile advances in meta-learning have enabled state-of-the-art performances in few-shot learning benchmarks, naturally prompting the question: Can meta-learning improve deep learning performance in low-resource drug discovery projects? In this work, we assess the efficiency of the Model-Agnostic Meta-Learning (MAML) algorithm - along with its variants FO-MAML and ANIL - at learning to predict chemical properties and activities. Using the ChEMBL20 dataset to emulate low-resource settings, our benchmark shows that meta-initializations perform comparably to or outperform multi-task pre-training baselines on 16 out of 20 in-distribution tasks and on all out-of-distribution tasks, providing an average improvement in AUPRC of 7.2% and 14.9% respectively. Finally, we observe that meta-initializations consistently result in the best performing models across fine-tuning sets with $k \in \{16, 32, 64, 128, 256\}$ instances.
Many aspects of human reasoning, including language, require learning rules from very little data. Humans can do this, often learning systematic rules from very few examples, and combining these rules to form compositional rule-based systems. Current neural architectures, on the other hand, often fail to generalize in a compositional manner, especially when evaluated in ways that vary systematically from training. In this work, we present a neuro-symbolic model which learns entire rule systems from a small set of examples. Instead of directly predicting outputs from inputs, we train our model to induce the explicit system of rules governing a set of previously seen examples, drawing upon techniques from the neural program synthesis literature. Our rule-synthesis approach outperforms neural meta-learning techniques in three domains: an artificial instruction-learning domain used to evaluate human learning, the SCAN challenge datasets, and learning rule-based translations of number words into integers for a wide range of human languages.
Continual learning studies agents that learn from streams of tasks without forgetting previous ones while adapting to new ones. Two recent continual-learning scenarios have opened new avenues of research. In meta-continual learning, the model is pre-trained to minimize catastrophic forgetting of previous tasks. In continual-meta learning, the aim is to train agents for faster remembering of previous tasks through adaptation. In their original formulations, both methods have limitations. We stand on their shoulders to propose a more general scenario, OSAKA, where an agent must quickly solve new (out-of-distribution) tasks, while also requiring fast remembering. We show that current continual learning, meta-learning, meta-continual learning, and continual-meta learning techniques fail in this new scenario. We propose Continual-MAML, an online extension of the popular MAML algorithm as a strong baseline for this scenario. We empirically show that Continual-MAML is better suited to the new scenario than the aforementioned methodologies, as well as standard continual learning and meta-learning approaches.
Being able to quickly adapt to changes in dynamics is paramount in model-based control for object manipulation tasks. In order to influence fast adaptation of the inverse dynamics model's parameters, data efficiency is crucial. Given observed data, a key element to how an optimizer updates model parameters is the loss function. In this work, we propose to apply meta-learning to learn structured, state-dependent loss functions during a meta-training phase. We then replace standard losses with our learned losses during online adaptation tasks. We evaluate our proposed approach on inverse dynamics learning tasks, both in simulation and on real hardware data. In both settings, the structured and state-dependent learned losses improve online adaptation speed, when compared to standard, state-independent loss functions.
Off-Policy Actor-Critic (Off-PAC) methods have proven successful in a variety of continuous control tasks. Normally, the critic's action-value function is updated using temporal-difference, and the critic in turn provides a loss for the actor that trains it to take actions with higher expected return. In this paper, we introduce a novel and flexible meta-critic that observes the learning process and meta-learns an additional loss for the actor that accelerates and improves actor-critic learning. Compared to the vanilla critic, the meta-critic network is explicitly trained to accelerate the learning process; and compared to existing meta-learning algorithms, meta-critic is rapidly learned online for a single task, rather than slowly over a family of tasks. Crucially, our meta-critic framework is designed for off-policy based learners, which currently provide state-of-the-art reinforcement learning sample efficiency. We demonstrate that online meta-critic learning leads to improvements in avariety of continuous control environments when combined with contemporary Off-PAC methods DDPG, TD3 and the state-of-the-art SAC.