We argue that one of the main obstacles for developing effective Continual Reinforcement Learning (CRL) algorithms is the negative transfer issue occurring when the new task to learn arrives. Through comprehensive experimental validation, we demonstrate that such issue frequently exists in CRL and cannot be effectively addressed by several recent work on mitigating plasticity loss of RL agents. To that end, we develop Reset & Distill (R&D), a simple yet highly effective method, to overcome the negative transfer problem in CRL. R&D combines a strategy of resetting the agent's online actor and critic networks to learn a new task and an offline learning step for distilling the knowledge from the online actor and previous expert's action probabilities. We carried out extensive experiments on long sequence of Meta-World tasks and show that our method consistently outperforms recent baselines, achieving significantly higher success rates across a range of tasks. Our findings highlight the importance of considering negative transfer in CRL and emphasize the need for robust strategies like R&D to mitigate its detrimental effects.
We address the problem of network calibration adjusting miscalibrated confidences of deep neural networks. Many approaches to network calibration adopt a regularization-based method that exploits a regularization term to smooth the miscalibrated confidences. Although these approaches have shown the effectiveness on calibrating the networks, there is still a lack of understanding on the underlying principles of regularization in terms of network calibration. We present in this paper an in-depth analysis of existing regularization-based methods, providing a better understanding on how they affect to network calibration. Specifically, we have observed that 1) the regularization-based methods can be interpreted as variants of label smoothing, and 2) they do not always behave desirably. Based on the analysis, we introduce a novel loss function, dubbed ACLS, that unifies the merits of existing regularization methods, while avoiding the limitations. We show extensive experimental results for image classification and semantic segmentation on standard benchmarks, including CIFAR10, Tiny-ImageNet, ImageNet, and PASCAL VOC, demonstrating the effectiveness of our loss function.
We address the problem of incremental semantic segmentation (ISS) recognizing novel object/stuff categories continually without forgetting previous ones that have been learned. The catastrophic forgetting problem is particularly severe in ISS, since pixel-level ground-truth labels are available only for the novel categories at training time. To address the problem, regularization-based methods exploit probability calibration techniques to learn semantic information from unlabeled pixels. While such techniques are effective, there is still a lack of theoretical understanding of them. Replay-based methods propose to memorize a small set of images for previous categories. They achieve state-of-the-art performance at the cost of large memory footprint. We propose in this paper a novel ISS method, dubbed ALIFE, that provides a better compromise between accuracy and efficiency. To this end, we first show an in-depth analysis on the calibration techniques to better understand the effects on ISS. Based on this, we then introduce an adaptive logit regularizer (ALI) that enables our model to better learn new categories, while retaining knowledge for previous ones. We also present a feature replay scheme that memorizes features, instead of images directly, in order to reduce memory requirements significantly. Since a feature extractor is changed continually, memorized features should also be updated at every incremental stage. To handle this, we introduce category-specific rotation matrices updating the features for each category separately. We demonstrate the effectiveness of our approach with extensive experiments on standard ISS benchmarks, and show that our method achieves a better trade-off in terms of accuracy and efficiency.
Class-incremental semantic segmentation (CISS) labels each pixel of an image with a corresponding object/stuff class continually. To this end, it is crucial to learn novel classes incrementally without forgetting previously learned knowledge. Current CISS methods typically use a knowledge distillation (KD) technique for preserving classifier logits, or freeze a feature extractor, to avoid the forgetting problem. The strong constraints, however, prevent learning discriminative features for novel classes. We introduce a CISS framework that alleviates the forgetting problem and facilitates learning novel classes effectively. We have found that a logit can be decomposed into two terms. They quantify how likely an input belongs to a particular class or not, providing a clue for a reasoning process of a model. The KD technique, in this context, preserves the sum of two terms (i.e., a class logit), suggesting that each could be changed and thus the KD does not imitate the reasoning process. To impose constraints on each term explicitly, we propose a new decomposed knowledge distillation (DKD) technique, improving the rigidity of a model and addressing the forgetting problem more effectively. We also introduce a novel initialization method to train new classifiers for novel classes. In CISS, the number of negative training samples for novel classes is not sufficient to discriminate old classes. To mitigate this, we propose to transfer knowledge of negatives to the classifiers successively using an auxiliary classifier, boosting the performance significantly. Experimental results on standard CISS benchmarks demonstrate the effectiveness of our framework.
We address the task of person search, that is, localizing and re-identifying query persons from a set of raw scene images. Recent approaches are typically built upon OIMNet, a pioneer work on person search, that learns joint person representations for performing both detection and person re-identification (reID) tasks. To obtain the representations, they extract features from pedestrian proposals, and then project them on a unit hypersphere with L2 normalization. These methods also incorporate all positive proposals, that sufficiently overlap with the ground truth, equally to learn person representations for reID. We have found that 1) the L2 normalization without considering feature distributions degenerates the discriminative power of person representations, and 2) positive proposals often also depict background clutter and person overlaps, which could encode noisy features to person representations. In this paper, we introduce OIMNet++ that addresses the aforementioned limitations. To this end, we introduce a novel normalization layer, dubbed ProtoNorm, that calibrates features from pedestrian proposals, while considering a long-tail distribution of person IDs, enabling L2 normalized person representations to be discriminative. We also propose a localization-aware feature learning scheme that encourages better-aligned proposals to contribute more in learning discriminative representations. Experimental results and analysis on standard person search benchmarks demonstrate the effectiveness of OIMNet++.
We address the problem of generalized zero-shot semantic segmentation (GZS3) predicting pixel-wise semantic labels for seen and unseen classes. Most GZS3 methods adopt a generative approach that synthesizes visual features of unseen classes from corresponding semantic ones (e.g., word2vec) to train novel classifiers for both seen and unseen classes. Although generative methods show decent performance, they have two limitations: (1) the visual features are biased towards seen classes; (2) the classifier should be retrained whenever novel unseen classes appear. We propose a discriminative approach to address these limitations in a unified framework. To this end, we leverage visual and semantic encoders to learn a joint embedding space, where the semantic encoder transforms semantic features to semantic prototypes that act as centers for visual features of corresponding classes. Specifically, we introduce boundary-aware regression (BAR) and semantic consistency (SC) losses to learn discriminative features. Our approach to exploiting the joint embedding space, together with BAR and SC terms, alleviates the seen bias problem. At test time, we avoid the retraining process by exploiting semantic prototypes as a nearest-neighbor (NN) classifier. To further alleviate the bias problem, we also propose an inference technique, dubbed Apollonius calibration (AC), that modulates the decision boundary of the NN classifier to the Apollonius circle adaptively. Experimental results demonstrate the effectiveness of our framework, achieving a new state of the art on standard benchmarks.
We address the problem of weakly-supervised semantic segmentation (WSSS) using bounding box annotations. Although object bounding boxes are good indicators to segment corresponding objects, they do not specify object boundaries, making it hard to train convolutional neural networks (CNNs) for semantic segmentation. We find that background regions are perceptually consistent in part within an image, and this can be leveraged to discriminate foreground and background regions inside object bounding boxes. To implement this idea, we propose a novel pooling method, dubbed background-aware pooling (BAP), that focuses more on aggregating foreground features inside the bounding boxes using attention maps. This allows to extract high-quality pseudo segmentation labels to train CNNs for semantic segmentation, but the labels still contain noise especially at object boundaries. To address this problem, we also introduce a noise-aware loss (NAL) that makes the networks less susceptible to incorrect labels. Experimental results demonstrate that learning with our pseudo labels already outperforms state-of-the-art weakly- and semi-supervised methods on the PASCAL VOC 2012 dataset, and the NAL further boosts the performance.
In recent years, $Q$-learning has become indispensable for model-free reinforcement learning (MFRL). However, it suffers from well-known problems such as under- and overestimation bias of the value, which may adversely affect the policy learning. To resolve this issue, we propose a MFRL framework that is augmented with the components of model-based RL. Specifically, we propose to estimate not only the $Q$-values but also both the transition and the reward with a shared network. We further utilize the estimated reward from the model estimators for $Q$-learning, which promotes interaction between the estimators. We show that the proposed scheme, called Model-augmented $Q$-learning (MQL), obtains a policy-invariant solution which is identical to the solution obtained by learning with true reward. Finally, we also provide a trick to prioritize past experiences in the replay buffer by utilizing model-estimation errors. We experimentally validate MQL built upon state-of-the-art off-policy MFRL methods, and show that MQL largely improves their performance and convergence. The proposed scheme is simple to implement and does not require additional training cost.
Experience replay, which enables the agents to remember and reuse experience from the past, plays a significant role in the success of off-policy reinforcement learning (RL). To utilize the experience replay efficiently, experience transitions should be sampled with consideration of their significance, such that the known prioritized experience replay (PER) further allows to sample more important experience. Yet, the conventional PER may result in generating highly biased samples due to considering a single metric such as TD-error and computing the sampling rate independently for each experience. To tackle this issue, we propose a Neural Experience Replay Sampler (NERS), which adaptively evaluates the relative importance of a sampled transition by obtaining context from not only its (local) values that characterize itself such as TD-error or the raw features but also other (global) transitions. We validate our framework on multiple benchmark tasks for both continuous and discrete controls and show that the proposed framework significantly improves the performance of various off-policy RL methods. Further analysis confirms that the improvements indeed come from the use of diverse features and the consideration of the relative importance of experiences.
Energy and resource efficient training of DNNs will greatly extend the applications of deep learning. However, there are three major obstacles which mandate accurate calculation in high precision. In this paper, we tackle two of them related to the loss of gradients during parameter update and backpropagation through a softmax nonlinearity layer in low precision training. We implemented SGD with Kahan summation by employing an additional parameter to virtually extend the bit-width of the parameters for a reliable parameter update. We also proposed a simple guideline to help select the appropriate bit-width for the last FC layer followed by a softmax nonlinearity layer. It determines the lower bound of the required bit-width based on the class size of the dataset. Extensive experiments on various network architectures and benchmarks verifies the effectiveness of the proposed technique for low precision training.