Deep learning approaches are nowadays ubiquitously used to tackle computer vision tasks such as semantic segmentation, requiring large datasets and substantial computational power. Continual learning for semantic segmentation (CSS) is an emerging trend that consists in updating an old model by sequentially adding new classes. However, continual learning methods are usually prone to catastrophic forgetting. This issue is further aggravated in CSS where, at each step, old classes from previous iterations are collapsed into the background. In this paper, we propose Local POD, a multi-scale pooling distillation scheme that preserves long- and short-range spatial relationships at feature level. Furthermore, we design an entropy-based pseudo-labelling of the background w.r.t. classes predicted by the old model to deal with background shift and avoid catastrophic forgetting of the old classes. Finally, we introduce a novel rehearsal method that is particularly suited for segmentation. Our approach, called PLOP, significantly outperforms state-of-the-art methods in existing CSS scenarios, as well as in newly proposed challenging benchmarks.
In class-incremental learning, an agent with limited resources needs to learn a sequence of classification tasks, forming an ever growing classification problem, with the constraint of not being able to access data from previous tasks. The main difference with task-incremental learning, where a task-ID is available at inference time, is that the learner also needs to perform cross-task discrimination, i.e. distinguish between classes that have not been seen together. Approaches to tackle this problem are numerous and mostly make use of an external memory (buffer) of non-negligible size. In this paper, we ablate the learning of cross-task features and study its influence on the performance of basic replay strategies used for class-IL. We also define a new forgetting measure for class-incremental learning, and see that forgetting is not the principal cause of low performance. Our experimental results show that future algorithms for class-incremental learning should not only prevent forgetting, but also aim to improve the quality of the cross-task features, and the knowledge transfer between tasks. This is especially important when tasks contain limited amount of data.
This paper introduces a solid state-of-the-art baseline for a class-incremental semantic segmentation (CISS) problem. While the recent CISS algorithms utilize variants of the knowledge distillation (KD) technique to tackle the problem, they failed to fully address the critical challenges in CISS causing the catastrophic forgetting; the semantic drift of the background class and the multi-label prediction issue. To better address these challenges, we propose a new method, dubbed SSUL-M (Semantic Segmentation with Unknown Label with Memory), by carefully combining techniques tailored for semantic segmentation. Specifically, we claim three main contributions. (1) defining unknown classes within the background class to help to learn future classes (help plasticity), (2) freezing backbone network and past classifiers with binary cross-entropy loss and pseudo-labeling to overcome catastrophic forgetting (help stability), and (3) utilizing tiny exemplar memory for the first time in CISS to improve both plasticity and stability. The extensively conducted experiments show the effectiveness of our method, achieving significantly better performance than the recent state-of-the-art baselines on the standard benchmark datasets. Furthermore, we justify our contributions with thorough ablation analyses and discuss different natures of the CISS problem compared to the traditional class-incremental learning targeting classification. The official code is available at https://github.com/clovaai/SSUL.
This paper proposes two novel knowledge transfer techniques for class-incremental learning (CIL). First, we propose data-free generative replay (DF-GR) to mitigate catastrophic forgetting in CIL by using synthetic samples from a generative model. In the conventional generative replay, the generative model is pre-trained for old data and shared in extra memory for later incremental learning. In our proposed DF-GR, we train a generative model from scratch without using any training data, based on the pre-trained classification model from the past, so we curtail the cost of sharing pre-trained generative models. Second, we introduce dual-teacher information distillation (DT-ID) for knowledge distillation from two teachers to one student. In CIL, we use DT-ID to learn new classes incrementally based on the pre-trained model for old classes and another model (pre-)trained on the new data for new classes. We implemented the proposed schemes on top of one of the state-of-the-art CIL methods and showed the performance improvement on CIFAR-100 and ImageNet datasets.
Modern computer vision applications suffer from catastrophic forgetting when incrementally learning new concepts over time. The most successful approaches to alleviate this forgetting require extensive replay of previously seen data, which is problematic when memory constraints or data legality concerns exist. In this work, we consider the high-impact problem of Data-Free Class-Incremental Learning (DFCIL), where an incremental learning agent must learn new concepts over time without storing generators or training data from past tasks. One approach for DFCIL is to replay synthetic images produced by inverting a frozen copy of the learner's classification model, but we show this approach fails for common class-incremental benchmarks when using standard distillation strategies. We diagnose the cause of this failure and propose a novel incremental distillation strategy for DFCIL, contributing a modified cross-entropy training and importance-weighted feature distillation, and show that our method results in up to a 25.1% increase in final task accuracy (absolute difference) compared to SOTA DFCIL methods for common class-incremental benchmarks. Our method even outperforms several standard replay based methods which store a coreset of images.
We propose a novel approach for class incremental online learning in a limited data setting. This problem setting is challenging because of the following constraints: (1) Classes are given incrementally, which necessitates a class incremental learning approach; (2) Data for each class is given in an online fashion, i.e., each training example is seen only once during training; (3) Each class has very few training examples; and (4) We do not use or assume access to any replay/memory to store data from previous classes. Therefore, in this setting, we have to handle twofold problems of catastrophic forgetting and overfitting. In our approach, we learn robust representations that are generalizable across tasks without suffering from the problems of catastrophic forgetting and overfitting to accommodate future classes with limited samples. Our proposed method leverages the meta-learning framework with knowledge consolidation. The meta-learning framework helps the model for rapid learning when samples appear in an online fashion. Simultaneously, knowledge consolidation helps to learn a robust representation against forgetting under online updates to facilitate future learning. Our approach significantly outperforms other methods on several benchmarks.
Neural networks are prone to catastrophic forgetting when trained incrementally on different tasks. Popular incremental learning methods mitigate such forgetting by retaining a subset of previously seen samples and replaying them during the training on subsequent tasks. However, this is not always possible, e.g., due to data protection regulations. In such restricted scenarios, one can employ generative models to replay either artificial images or hidden features to a classifier. In this work, we propose Genifer (GENeratIve FEature-driven image Replay), where a generative model is trained to replay images that must induce the same hidden features as real samples when they are passed through the classifier. Our technique therefore incorporates the benefits of both image and feature replay, i.e.: (1) unlike conventional image replay, our generative model explicitly learns the distribution of features that are relevant for classification; (2) in contrast to feature replay, our entire classifier remains trainable; and (3) we can leverage image-space augmentations, which increase distillation performance while also mitigating overfitting during the training of the generative model. We show that Genifer substantially outperforms the previous state of the art for various settings on the CIFAR-100 and CUB-200 datasets.
Training on class-imbalanced data usually results in biased models that tend to predict samples into the majority classes, which is a common and notorious problem. From the perspective of energy-based model, we demonstrate that the free energies of categories are aligned with the label distribution theoretically, thus the energies of different classes are expected to be close to each other when aiming for ``balanced'' performance. However, we discover a severe energy-bias phenomenon in the models trained on class-imbalanced dataset. To eliminate the bias, we propose a simple and effective method named Energy Aligning by merely adding the calculated shift scalars onto the output logits during inference, which does not require to (i) modify the network architectures, (ii) intervene the standard learning paradigm, (iii) perform two-stage training. The proposed algorithm is evaluated on two class imbalance-related tasks under various settings: class incremental learning and long-tailed recognition. Experimental results show that energy aligning can effectively alleviate class imbalance issue and outperform state-of-the-art methods on several benchmarks.
Deep Learning (DL) has been utilized pervasively in the Internet of Things (IoT). One typical application of DL in IoT is device identification from wireless signals, namely Non-cryptographic Device Identification (NDI). However, learning components in NDI systems have to evolve to adapt to operational variations, such a paradigm is termed as Incremental Learning (IL). Various IL algorithms have been proposed and many of them require dedicated space to store the increasing amount of historical data, and therefore, they are not suitable for IoT or mobile applications. However, conventional IL schemes can not provide satisfying performance when historical data are not available. In this paper, we address the IL problem in NDI from a new perspective, firstly, we provide a new metric to measure the degree of topological maturity of DNN models from the degree of conflict of class-specific fingerprints. We discover that an important cause for performance degradation in IL enabled NDI is owing to the conflict of devices' fingerprints. Second, we also show that the conventional IL schemes can lead to low topological maturity of DNN models in NDI systems. Thirdly, we propose a new Channel Separation Enabled Incremental Learning (CSIL) scheme without using historical data, in which our strategy can automatically separate devices' fingerprints in different learning stages and avoid potential conflict. Finally, We evaluated the effectiveness of the proposed framework using real data from ADS-B (Automatic Dependent Surveillance-Broadcast), an application of IoT in aviation. The proposed framework has the potential to be applied to accurate identification of IoT devices in a variety of IoT applications and services. Data and code available at IEEE Dataport (DOI: 10.21227/1bxc-ke87) and \url{https://github.com/pcwhy/CSIL}}
Incrementally training deep neural networks to recognize new classes is a challenging problem. Most existing class-incremental learning methods store data or use generative replay, both of which have drawbacks, while 'rehearsal-free' alternatives such as parameter regularization or bias-correction methods do not consistently achieve high performance. Here, we put forward a new strategy for class-incremental learning: generative classification. Rather than directly learning the conditional distribution p(y|x), our proposal is to learn the joint distribution p(x,y), factorized as p(x|y)p(y), and to perform classification using Bayes' rule. As a proof-of-principle, here we implement this strategy by training a variational autoencoder for each class to be learned and by using importance sampling to estimate the likelihoods p(x|y). This simple approach performs very well on a diverse set of continual learning benchmarks, outperforming generative replay and other existing baselines that do not store data.
We present an approach for continual learning (CL) that is based on fully probabilistic (or generative) models of machine learning. In contrast to, e.g., GANs that are "generative" in the sense that they can generate samples, fully probabilistic models aim at modeling the data distribution directly. Consequently, they provide functionalities that are highly relevant for continual learning, such as density estimation (outlier detection) and sample generation. As a concrete realization of generative continual learning, we propose Gaussian Mixture Replay (GMR). GMR is a pseudo-rehearsal approach using a Gaussian Mixture Model (GMM) instance for both generator and classifier functionalities. Relying on the MNIST, FashionMNIST and Devanagari benchmarks, we first demonstrate unsupervised task boundary detection by GMM density estimation, which we also use to reject untypical generated samples. In addition, we show that GMR is capable of class-conditional sampling in the way of a cGAN. Lastly, we verify that GMR, despite its simple structure, achieves state-of-the-art performance on common class-incremental learning problems at very competitive time and memory complexity.
While many works on Continual Learning have shown promising results for mitigating catastrophic forgetting, they have relied on supervised training. To successfully learn in a label-agnostic incremental setting, a model must distinguish between learned and novel classes to properly include samples for training. We introduce a novelty detection method that leverages network confusion caused by training incoming data as a new class. We found that incorporating a class-imbalance during this detection method substantially enhances performance. The effectiveness of our approach is demonstrated across a set of image classification benchmarks: MNIST, SVHN, CIFAR-10, CIFAR-100, and CRIB.
Few-shot class-incremental learning (FSCIL) aims to design machine learning algorithms that can continually learn new concepts from a few data points, without forgetting knowledge of old classes. The difficulty lies in that limited data from new classes not only lead to significant overfitting issues but also exacerbate the notorious catastrophic forgetting problems. Moreover, as training data come in sequence in FSCIL, the learned classifier can only provide discriminative information in individual sessions, while FSCIL requires all classes to be involved for evaluation. In this paper, we address the FSCIL problem from two aspects. First, we adopt a simple but effective decoupled learning strategy of representations and classifiers that only the classifiers are updated in each incremental session, which avoids knowledge forgetting in the representations. By doing so, we demonstrate that a pre-trained backbone plus a non-parametric class mean classifier can beat state-of-the-art methods. Second, to make the classifiers learned on individual sessions applicable to all classes, we propose a Continually Evolved Classifier (CEC) that employs a graph model to propagate context information between classifiers for adaptation. To enable the learning of CEC, we design a pseudo incremental learning paradigm that episodically constructs a pseudo incremental learning task to optimize the graph parameters by sampling data from the base dataset. Experiments on three popular benchmark datasets, including CIFAR100, miniImageNet, and Caltech-USCD Birds-200-2011 (CUB200), show that our method significantly outperforms the baselines and sets new state-of-the-art results with remarkable advantages.
Few-shot class-incremental learning (FSCIL), which targets at continuously expanding model's representation capacity under few supervisions, is an important yet challenging problem. On the one hand, when fitting new tasks (novel classes), features trained on old tasks (old classes) could significantly drift, causing catastrophic forgetting. On the other hand, training the large amount of model parameters with few-shot novel-class examples leads to model over-fitting. In this paper, we propose a learnable expansion-and-compression network (LEC-Net), with the aim to simultaneously solve catastrophic forgetting and model over-fitting problems in a unified framework. By tentatively expanding network nodes, LEC-Net enlarges the representation capacity of features, alleviating feature drift of old network from the perspective of model regularization. By compressing the expanded network nodes, LEC-Net purses minimal increase of model parameters, alleviating over-fitting of the expanded network from a perspective of compact representation. Experiments on the CUB/CIFAR-100 datasets show that LEC-Net improves the baseline by 5~7% while outperforms the state-of-the-art by 5~6%. LEC-Net also demonstrates the potential to be a general incremental learning approach with dynamic model expansion capability.
In class-incremental learning, the objective is to learn a number of classes sequentially without having access to the whole training data. However, due to a problem known as catastrophic forgetting, neural networks suffer substantial performance drop in such settings. The problem is often approached by experience replay, a method which stores a limited number of samples to be replayed in future steps to reduce forgetting of the learned classes. When using a pretrained network as a feature extractor, we show that instead of training a single classifier incrementally, it is better to train a number of specialized classifiers which do not interfere with each other yet can cooperatively predict a single class. Our experiments on CIFAR100 dataset show that the proposed method improves the performance over SOTA by a large margin.
Learning continually from non-stationary data streams is a long-standing goal and a challenging problem in machine learning. Recently, we have witnessed a renewed and fast-growing interest in continual learning, especially within the deep learning community. However, algorithmic solutions are often difficult to re-implement, evaluate and port across different settings, where even results on standard benchmarks are hard to reproduce. In this work, we propose Avalanche, an open-source end-to-end library for continual learning research based on PyTorch. Avalanche is designed to provide a shared and collaborative codebase for fast prototyping, training, and reproducible evaluation of continual learning algorithms.
We address the problem of class incremental learning, which is a core step towards achieving adaptive vision intelligence. In particular, we consider the task setting of incremental learning with limited memory and aim to achieve better stability-plasticity trade-off. To this end, we propose a novel two-stage learning approach that utilizes a dynamically expandable representation for more effective incremental concept modeling. Specifically, at each incremental step, we freeze the previously learned representation and augment it with additional feature dimensions from a new learnable feature extractor. This enables us to integrate new visual concepts with retaining learned knowledge. We dynamically expand the representation according to the complexity of novel concepts by introducing a channel-level mask-based pruning strategy. Moreover, we introduce an auxiliary loss to encourage the model to learn diverse and discriminate features for novel concepts. We conduct extensive experiments on the three class incremental learning benchmarks and our method consistently outperforms other methods with a large margin.
This paper presents a practical and simple yet efficient method to effectively deal with the catastrophic forgetting for Class Incremental Learning (CIL) tasks. CIL tends to learn new concepts perfectly, but not at the expense of performance and accuracy for old data. Learning new knowledge in the absence of data instances from previous classes or even imbalance samples of both old and new classes makes CIL an ongoing challenging problem. These issues can be tackled by storing exemplars belonging to the previous tasks or by utilizing the rehearsal strategy. Inspired by the rehearsal strategy with the approach of using generative models, we propose ClaRe, an efficient solution for CIL by remembering the representations of learned classes in each increment. Taking this approach leads to generating instances with the same distribution of the learned classes. Hence, our model is somehow retrained from the scratch using a new training set including both new and the generated samples. Subsequently, the imbalance data problem is also solved. ClaRe has a better generalization than prior methods thanks to producing diverse instances from the distribution of previously learned classes. We comprehensively evaluate ClaRe on the MNIST benchmark. Results show a very low degradation on accuracy against facing new knowledge over time. Furthermore, contrary to the most proposed solutions, the memory limitation is not problematic any longer which is considered as a consequential issue in this research area.
Online class-incremental continual learning (CL) studies the problem of learning new classes continually from an online non-stationary data stream, intending to adapt to new data while mitigating catastrophic forgetting. While memory replay has shown promising results, the recency bias in online learning caused by the commonly used Softmax classifier remains an unsolved challenge. Although the Nearest-Class-Mean (NCM) classifier is significantly undervalued in the CL community, we demonstrate that it is a simple yet effective substitute for the Softmax classifier. It addresses the recency bias and avoids structural changes in the fully-connected layer for new classes. Moreover, we observe considerable and consistent performance gains when replacing the Softmax classifier with the NCM classifier for several state-of-the-art replay methods. To leverage the NCM classifier more effectively, data embeddings belonging to the same class should be clustered and well-separated from those with a different class label. To this end, we contribute Supervised Contrastive Replay (SCR), which explicitly encourages samples from the same class to cluster tightly in embedding space while pushing those of different classes further apart during replay-based training. Overall, we observe that our proposed SCR substantially reduces catastrophic forgetting and outperforms state-of-the-art CL methods by a significant margin on a variety of datasets.
Human beings not only have the ability to recognize novel unseen classes, but also can incrementally incorporate the new classes to existing knowledge preserved. However, zero-shot learning models assume that all seen classes should be known beforehand, while incremental learning models cannot recognize unseen classes. This paper introduces a novel and challenging task of Incrementally Zero-Shot Detection (IZSD), a practical strategy for both zero-shot learning and class-incremental learning in real-world object detection. An innovative end-to-end model -- IZSD-EVer was proposed to tackle this task that requires incrementally detecting new classes and detecting the classes that have never been seen. Specifically, we propose a novel extreme value analyzer to detect objects from old seen, new seen, and unseen classes, simultaneously. Additionally and technically, we propose two innovative losses, i.e., background-foreground mean squared error loss alleviating the extreme imbalance of the background and foreground of images, and projection distance loss aligning the visual space and semantic spaces of old seen classes. Experiments demonstrate the efficacy of our model in detecting objects from both the seen and unseen classes, outperforming the alternative models on Pascal VOC and MSCOCO datasets.