中国科学院自动化研究所机构知识库

在非平稳环境中，人工学习系统需要面临不断学习新任务的挑战。如果以经典深度学习模式在非平稳环境中学习新任务，一方面在面对数据量极少的任务时模型容易发生过拟合问题；另一方面，在学习新任务后会导致对模型解决旧任务的能力发生灾难遗忘问题。相比之下，人类可以在非平稳环境中连续不断地学习新任务而不忘记旧任务能力，并且从学习过的任务中总结可迁移的先验知识帮助面对新任务只需利用少量样本就可以快速学习。因此，为了克服深度学习模型这两方面的局限性，构建更加智能的人工学习系统，本文从如何使深度学习模型学习可迁移的先验知识来帮助学习新任务，以及如何能够持续学习新任务两个方面进行了研究，具体贡献如下：

研究了一种面向小样本任务的元学习方法。该研究的目的是希望能够在模型中学习出可迁移先验知识，来快速学习新的小样本任务。该方法由两个关键模块组成，即目标网络和元网络。目标网络是一个针对小样本任务设计的分类网络，这里采用了匹配网络的计算结构；元网络则可以根据小样本任务的训练样本直接生成目标网络的功能参数。在元网络中设计了任务上下文编码器和条件权重生成器，任务上下文编码器使用任务训练集样本的特征统计信息来编码任务上下文特征，权重生成器则根据任务上下文特征编码生成出目标网络不同层的功能参数。我们通过任务情景式的训练方式将可迁移的先验知识嵌入在元网络中，并且提出了能够利用不同任务间的共享信息的任务间标准化策略来帮助模型训练。在训练之后，对面新的小样本任务，元网络可以直接根据任务的训练样本生成出目标网络解决该任务的功能参数，而不需要进一步的微调，从而达到快速学习和适应的目的。在合成数据集上的实验展示了方法的直观思路，通过可视化目标网络在不同任务上决策边界。在两个小样本学习的基准数据集上进行实验，实验结果证明了方法的有效性，并且在one-shot任务上达到了最前沿的效果。

研究了一种人脸图像属性转换知识迁移的方法。该研究的目的是希望能在模型中学习出人脸属性转换的可迁移先验知识，在面对新的人脸图像时，可以只根据语义信息对人脸图像进行转换，同时不改变人脸图像的身份特征。在该方法中，我们设计了一种条件再循环生成式对抗网络，它由一组条件生成器和具有辅助分类器的判别器组成，我们将人脸图像输入到生成器中并指定它所要变换的目标属性，进行属性转换之后，将生成的人脸图像重新输入到条件生成器中，并且指定它原来的人脸属性，将其变换回原来的人脸。为了能够保证转换之后的人脸具有相同的身份特征，保持再循环一致性，我们提出了再循环重建损失；为了引导条件生成器可以生成具有指定属性的人脸图像，我们在判别器中加入了多标签分类损失作为辅助损失引导训练。在训练过程中，我们只使用了人脸图像的属性语义标签，将每次人脸属性变换够看做是一个任务，然后从大量的人脸属性变换任务中学习人脸属性转换的可迁移先验知识。在一个代表性人脸数据集上进行实验，通过可视化的方式定性地评估了人脸图像转换效果，可以对多种属性进行相互转换，证实了方法的有效性。

研究了一种基于伪样本预演的类别增量学习方法。该研究受到互补学习系统理论的启发，利用了人工神经网络中不同的记忆类型来缓解持续学习中的灾难遗忘问题。一方面，我们提出了一种增量概念学习网络，它是一种带有外存记忆模块的分类网络，可以通过在概念记忆矩阵中增加概念向量的方式使网络能够学习新的类别；为了缓解灾难遗忘问题，我们提出了一种概念对比损失函数，在持续学习过程中约束不同类别的特征具有更小的类别方差，减小持续学习过程中网络权重的变化；另一方面，提出使用条件生成式对抗网络作为概念回忆网络，对旧类别概念进行记忆巩固，在学习新类别的时候对旧类别进行记忆回放，以及一种在线平衡回忆的训练策略来缓解灾难遗忘问题。为了展示方法的直观思路，我们可视化了类别增量学习过程中学习不同概念类别时特征分布的变化。在几个代表性的类别增量学习数据集上进行实验，与基于基于伪样本预演的持续方法相比达到的当时最前沿的效果，证实了方法的有效性。
    
研究了一种基于特征零空间梯度转换的持续学习方法。基于随机梯度下降算法优化的神经网络模型，在持续学习过程中容易发生灾难遗忘。主要原因是由于在学习新任务数据后，神经网络对于旧任务数据的映射关系发生了严重的改变导致的。因此，如果可以在优化新任务过程中调整梯度更新的方向，使神经网络在学习新任务时尽可能少地改变旧任务数据的映射关系，那么就可以缓解灾难遗忘问题。于是，我们提出了特征零空间梯度转换的方法，在线性层中将来自新目标函数的梯度转换到旧数据特征向量张成空间的零空间上，从而可以在学习新任务数据时保持每个线性层对旧数据特征映射关系不改变，进而保持整个网络对于旧数据样本的特征映射关系尽可能少地改变，从而缓解灾难遗忘问题。为了提高特征零空间梯度转换函数的计算效率，提出了一种近似零空间投影矩阵的在线累积计算方法，以及三种不同的批数据特征矩阵构造的策略。在合成数据集上，通过可视化类别增量学习过程中神经网络分类器决策边界的变化，展示了方法的直观思路。值得注意的是，该方法在持续学习过程中，不需要使用任何旧任务的历史数据，在几个代表性数据集构造的类别增量学习任务上，明显超过其他基于正则化的持续学习方法。通过进一步的实验探索，发现该方法可以通过改变超参数来控制神经网络在持续学习过程中的稳定性-可塑性平衡，意味着生物角度的解释性。

In non-stationary environments, artificial learning systems need to face the challenge of continuously learning new tasks. However, there exist limitations while using traditional deep learning paradigm to learn new tasks in non-stationary environments. On the one hand, the deep learning models are prone to overfitting when the training data of a new task is scarce. On the other hand, after learning a new task, the deep learning models will catastrophically forget how to solve previous tasks. In contrast, humans are capable of continually learning new tasks without forgetting old ones in non-stationary environments. Furthermore, we can also obtain transferable prior knowledge from learning different tasks which can help to learn new tasks rapidly using only a few examples. Therefore, to overcome these limitations in deep learning models and build more intelligent artificial learning systems, our research concentrates on two aspects. The first is how to make deep learning models obtain transferable prior knowledge to help to learn new tasks, and the second is how to make deep learning models capable of continually learning new tasks. Our contributions are as follows:

A novel meta learning method for few-shot learning tasks. The main purpose of this research is to learn the transferable prior knowledge for rapidly learning and adapting to new few-shot tasks. This approach contains two key components, e.g. TargetNet module and MetaNet module. The TargetNet is a classification network that is designed for few-shot learning, and we use directly use the architecture of matching networks. The MetaNet can directly produce functional weights for TargetNet using the training samples of a few-shot task. In the MetaNet, we have designed a task context encoder and a conditional weight generator. The task context encoder aims to encode all training samples of a task and generate a task context feature. The conditional weight generator can generate the functional weights of TargetNet using the task context feature. Through episodic training procedure, MetaNet can learn the transferable prior knowledge. We also propose an intertask normalization strategy to utilize the common information shared among tasks during training. After training, for an unseen few-shot task, MetaNet can directly use the training samples to produce the functional weights of TargetNet for solving this task. No further fine-tuning is required, hence the goal of fast learning and adaptation is achieved. We have conducted several experiments on synthetic datasets to show the intuition of our approach, through visualizing the decision boundaries of TargetNet for different few-shot tasks. The experiment results on two few-shot learning benchmarks show the validity of our approach, and we have achieved the state-of-the-art performance on the one-shot learning task.

An approach of knowledge transfer for facial attribute transformation. The main purpose of this research is to learn transferable prior knowledge of facial attribute transformation. It can transform different attributes of an unseen facial image without changing the personal identity. In this approach, we design a conditional recycle generative adversarial network, which consists of a conditional generator and a discriminator with auxiliary classifier. During training, we first input the original facial image and target facial attributes to the generator and produce a transformed facial image. Then, we recycle the transformed facial image and original facial attributes to the generator and produce a facial image that should be completely the same as the original one. In order to maintain the facial personal identity and keep recycle consistency, we propose a recycle-consistent loss. To guide the generator to produce the facial image with specified attributes, we add a multi-label classification loss in the discriminator. During training, we only use the facial attribute labels. We can regard each facial image attribute transformation as a task. Through learning a lot of facial image attribute transformation tasks, the generator could learn the transferable prior knowledge of facial attribute transformation which can be applied to new facial images. In several experiments on a facial image dataset, we qualitatively compare and evaluate the results which prove the validity of our approach.

A novel rehearsal-based class incremental learning approach. Inspired by complementary learning systems theory, we propose to utilize different types of memory in artificial neural networks to alleviate the catastrophic forgetting problem during continual learning. On the one hand, we propose an incremental concept learning network which is a remolded classification network with a dynamic external memory module. During learning new classes, it can dynamically increase the corresponding number of concept vectors in the concept memory matrix. To alleviate the catastrophic forgetting problem, we propose a concept contrastive loss that can reduce the intra-class feature variance during continual learning. On the other hand, we propose the concept memory recall network which is a conditional generative adversarial network, to consolidate learned concepts and recall old concepts during learning new categories. We also propose an online balanced recall strategy to mitigate the catastrophic forgetting problem. To show the intuition of our approach, we visualize the feature variation of different learned classes during continual learning. We conduct class incremental learning experiments on several datasets to show the validity of our approach. Comparing with rehearsal-based continual learning methods, we have achieved state-of-the-art performance.

A continual learning approach based on feature null space gradient conversion. The neural network model, which is optimized by stochastic gradient descent algorithms, is prone to happen catastrophic forgetting during continual learning. The main reason lies in that the feature mapping relation of old task data in the neural network is heavily changed after learning new task data. Therefore, if we can adjust the new loss gradients into the directs that lead to the change of feature mapping relation as small as possible while learning new task data, the catastrophic forgetting problem can be mitigated. Hence，we propose the feature null space gradient conversion method. In each linear layer, we should project the gradient from new task loss into the null space of the space spanned by old task data feature vectors. In this way, the old feature mapping relation in the linear layer will not be changed during learning new task data. Furthermore, the old task data feature mapping relation in the whole network will be minimally changed and the catastrophic forgetting will be alleviated. In order to improve the computational efficiency of feature null space gradient conversion function, we propose an online accumulative way to approximate the null space projector as well as three different batch feature matrix construction strategies。To exhibit the intuition, we conduct the experiments on a synthetic dataset and visualize the variation of decision boundaries of neural network classifier during incrementally learn new classes. It is worth noting that our approach does not require any historical data during continual learning. Comparing with other regularization based approaches in the same condition, our results obviously outperform. In several exploration experiments, we find that our approach could control the stability-plasticity balance of the neural network during continual learning through adjust a hyperparameter, which means some biological explanations.