Ya-nan Han

Ya-nan Han, Jian-wei Liu

The goal of continual learning is to provide intelligent agents that are capable of learning continually a sequence of tasks using the knowledge obtained from previous tasks while performing well on prior tasks. However, a key challenge in this continual learning paradigm is catastrophic forgetting, namely adapting a model to new tasks often leads to severe performance degradation on prior tasks. Current memory-based approaches show their success in alleviating the catastrophic forgetting problem by replaying examples from past tasks when new tasks are learned. However, these methods are infeasible to transfer the structural knowledge from previous tasks i.e., similarities or dissimilarities between different instances. Furthermore, the learning bias between the current and prior tasks is also an urgent problem that should be solved. In this work, we propose a new method, named Online Continual Learning via the Knowledge Invariant and Spread-out Properties (OCLKISP), in which we constrain the evolution of the embedding features via Knowledge Invariant and Spread-out Properties (KISP). Thus, we can further transfer the inter-instance structural knowledge of previous tasks while alleviating the forgetting due to the learning bias. We empirically evaluate our proposed method on four popular benchmarks for continual learning: Split CIFAR 100, Split SVHN, Split CUB200 and Split Tiny-Image-Net. The experimental results show the efficacy of our proposed method compared to the state-of-the-art continual learning algorithms.

Ya-nan Han, Jian-wei Liu, Bing-biao Xiao et al.

Recent years have witnessed enormous progress of online learning. However, a major challenge on the road to artificial agents is concept drift, that is, the data probability distribution would change where the data instance arrives sequentially in a stream fashion, which would lead to catastrophic forgetting and degrade the performance of the model. In this paper, we proposed a new Bilevel Online Deep Learning (BODL) framework, which combine bilevel optimization strategy and online ensemble classifier. In BODL algorithm, we use an ensemble classifier, which use the output of different hidden layers in deep neural network to build multiple base classifiers, the important weights of the base classifiers are updated according to exponential gradient descent method in an online manner. Besides, we apply the similar constraint to overcome the convergence problem of online ensemble framework. Then an effective concept drift detection mechanism utilizing the error rate of classifier is designed to monitor the change of the data probability distribution. When the concept drift is detected, our BODL algorithm can adaptively update the model parameters via bilevel optimization and then circumvent the large drift and encourage positive transfer. Finally, the extensive experiments and ablation studies are conducted on various datasets and the competitive numerical results illustrate that our BODL algorithm is a promising approach.

Zhi-yan Song, Jian-wei Liu, Lu-ning Zhang et al.

Abstract. Most of the real world data we encounter are asynchronous event sequence, so the last decades have been characterized by the implementation of various point process into the field of social networks,electronic medical records and financial transactions. At the beginning, Hawkes process and its variants which can simulate simultaneously the self-triggering and mutual triggering patterns between different events in complex sequences in a clear and quantitative way are more popular.Later on, with the advances of neural network, neural Hawkes process has been proposed one after another, and gradually become a research hotspot. The proposal of the transformer Hawkes process (THP) has gained a huge performance improvement, so a new upsurge of the neural Hawkes process based on transformer is set off. However, THP does not make full use of the information of occurrence time and type of event in the asynchronous event sequence. It simply adds the encoding of event type conversion and the location encoding of time conversion to the source encoding. At the same time, the learner built from a single transformer will result in an inescapable learning bias. In order to mitigate these problems, we propose a tri-transformer Hawkes process (Tri-THP) model, in which the event and time information are added to the dot-product attention as auxiliary information to form a new multihead attention. The effectiveness of the Tri-THP is proved by a series of well-designed experiments on both real world and synthetic data.