اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدMatch What Matters: Generative Implicit Feature Replay for Continual Learning
79
0
0.0
(
0
)
تأليف
Kevin Thandiackal
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Neural networks are prone to catastrophic forgetting when trained incrementally on different tasks. In order to prevent forgetting, most existing methods retain a small subset of previously seen samples, which in turn can be used for joint training with new tasks. While this is indeed effective, it may not always be possible to store such samples, e.g., due to data protection regulations. In these cases, one can instead employ generative models to create artificial samples or features representing memories from previous tasks. Following a similar direction, we propose GenIFeR (Generative Implicit Feature Replay) for class-incremental learning. The main idea is to train a generative adversarial network (GAN) to generate images that contain realistic features. While the generator creates images at full resolution, the discriminator only sees the corresponding features extracted by the continually trained classifier. Since the classifier compresses raw images into features that are actually relevant for classification, the GAN can match this target distribution more accurately. On the other hand, allowing the generator to create full resolution images has several benefits: In contrast to previous approaches, the feature extractor of the classifier does not have to be frozen. In addition, we can employ augmentations on generated images, which not only boosts classification performance, but also mitigates discriminator overfitting during GAN training. We empirically show that GenIFeR is superior to both conventional generative image and feature replay. In particular, we significantly outperform the state-of-the-art in generative replay for various settings on the CIFAR-100 and CUB-200 datasets.
قيم البحث
اقرأ أيضاً
Continual learning consists in incrementally training a model on a sequence of datasets and testing on the union of all datasets. In this paper, we examine continual learning for the problem of sound classification, in which we wish to refine already
trained models to learn new sound classes. In practice one does not want to maintain all past training data and retrain from scratch, but naively updating a model with new data(sets) results in a degradation of already learned tasks, which is referred to as catastrophic forgetting. We develop a generative replay procedure for generating training audio spectrogram data, in place of keeping older training datasets. We show that by incrementally refining a classifier with generative replay a generator that is 4% of the size of all previous training data matches the performance of refining the classifier keeping 20% of all previous training data. We thus conclude that we can extend a trained sound classifier to learn new classes without having to keep previously used datasets.
Continual learning is the problem of learning new tasks or knowledge while protecting old knowledge and ideally generalizing from old experience to learn new tasks faster. Neural networks trained by stochastic gradient descent often degrade on old ta
sks when trained successively on new tasks with different data distributions. This phenomenon, referred to as catastrophic forgetting, is considered a major hurdle to learning with non-stationary data or sequences of new tasks, and prevents networks from continually accumulating knowledge and skills. We examine this issue in the context of reinforcement learning, in a setting where an agent is exposed to tasks in a sequence. Unlike most other work, we do not provide an explicit indication to the model of task boundaries, which is the most general circumstance for a learning agent exposed to continuous experience. While various methods to counteract catastrophic forgetting have recently been proposed, we explore a straightforward, general, and seemingly overlooked solution - that of using experience replay buffers for all past events - with a mixture of on- and off-policy learning, leveraging behavioral cloning. We show that this strategy can still learn new tasks quickly yet can substantially reduce catastrophic forgetting in both Atari and DMLab domains, even matching the performance of methods that require task identities. When buffer storage is constrained, we confirm that a simple mechanism for randomly discarding data allows a limited size buffer to perform almost as well as an unbounded one.
Learning a sequence of tasks without access to i.i.d. observations is a widely studied form of continual learning (CL) that remains challenging. In principle, Bayesian learning directly applies to this setting, since recursive and one-off Bayesian up
dates yield the same result. In practice, however, recursive updating often leads to poor trade-off solutions across tasks because approximate inference is necessary for most models of interest. Here, we describe an alternative Bayesian approach where task-conditioned parameter distributions are continually inferred from data. We offer a practical deep learning implementation of our framework based on probabilistic task-conditioned hypernetworks, an approach we term posterior meta-replay. Experiments on standard benchmarks show that our probabilistic hypernetworks compress sequences of posterior parameter distributions with virtually no forgetting. We obtain considerable performance gains compared to existing Bayesian CL methods, and identify task inference as our major limiting factor. This limitation has several causes that are independent of the considered sequential setting, opening up new avenues for progress in CL.
Regularization-based methods are beneficial to alleviate the catastrophic forgetting problem in class-incremental learning. With the absence of old task images, they often assume that old knowledge is well preserved if the classifier produces similar
output on new images. In this paper, we find that their effectiveness largely depends on the nature of old classes: they work well on classes that are easily distinguishable between each other but may fail on more fine-grained ones, e.g., boy and girl. In spirit, such methods project new data onto the feature space spanned by the weight vectors in the fully connected layer, corresponding to old classes. The resulting projections would be similar on fine-grained old classes, and as a consequence the new classifier will gradually lose the discriminative ability on these classes. To address this issue, we propose a memory-free generative replay strategy to preserve the fine-grained old classes characteristics by generating representative old images directly from the old classifier and combined with new data for new classifier training. To solve the homogenization problem of the generated samples, we also propose a diversity loss that maximizes Kullback Leibler (KL) divergence between generated samples. Our method is best complemented by prior regularization-based methods proved to be effective for easily distinguishable old classes. We validate the above design and insights on CUB-200-2011, Caltech-101, CIFAR-100 and Tiny ImageNet and show that our strategy outperforms existing memory-free methods with a clear margin. Code is available at https://github.com/xmengxin/MFGR
Deep neural networks have shown promise in several domains, and the learned data (task) specific information is implicitly stored in the network parameters. Extraction and utilization of encoded knowledge representations are vital when data is no lon
ger available in the future, especially in a continual learning scenario. In this work, we introduce {em flashcards}, which are visual representations that {em capture} the encoded knowledge of a network as a recursive function of predefined random image patterns. In a continual learning scenario, flashcards help to prevent catastrophic forgetting and consolidating knowledge of all the previous tasks. Flashcards need to be constructed only before learning the subsequent task, and hence, independent of the number of tasks trained before. We demonstrate the efficacy of flashcards in capturing learned knowledge representation (as an alternative to the original dataset) and empirically validate on a variety of continual learning tasks: reconstruction, denoising, task-incremental learning, and new-instance learning classification, using several heterogeneous benchmark datasets. Experimental evidence indicates that: (i) flashcards as a replay strategy is { em task agnostic}, (ii) performs better than generative replay, and (iii) is on par with episodic replay without additional memory overhead.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
