Abstract:Class-Incremental Semantic Segmentation(CISS) aims to learn new classes without forgetting the old ones, using only the labels of the new classes. To achieve this, two popular strategies are employed: 1) pseudo-labeling and knowledge distillation to preserve prior knowledge; and 2) background weight transfer, which leverages the broad coverage of background in learning new classes by transferring background weight to the new class classifier. However, the first strategy heavily relies on the old model in detecting old classes while undetected pixels are regarded as the background, thereby leading to the background shift towards the old classes(i.e., misclassification of old class as background). Additionally, in the case of the second approach, initializing the new class classifier with background knowledge triggers a similar background shift issue, but towards the new classes. To address these issues, we propose a background-class separation framework for CISS. To begin with, selective pseudo-labeling and adaptive feature distillation are to distill only trustworthy past knowledge. On the other hand, we encourage the separation between the background and new classes with a novel orthogonal objective along with label-guided output distillation. Our state-of-the-art results validate the effectiveness of these proposed methods.
Abstract:While interests in tabular deep learning has significantly grown, conventional tree-based models still outperform deep learning methods. To narrow this performance gap, we explore the innovative retrieval mechanism, a methodology that allows neural networks to refer to other data points while making predictions. Our experiments reveal that retrieval-based training, especially when fine-tuning the pretrained TabPFN model, notably surpasses existing methods. Moreover, the extensive pretraining plays a crucial role to enhance the performance of the model. These insights imply that blending the retrieval mechanism with pretraining and transfer learning schemes offers considerable potential for advancing the field of tabular deep learning.
Abstract:Recently, as the spread of smart devices increases, the amount of data collected through sensors is increasing. A lifelog is a kind of big data to analyze behavior patterns in the daily life of individuals collected from various smart de-vices. However, sensor data is a low-level signal that makes it difficult for hu-mans to recognize the situation directly and cannot express relations clearly. It is also difficult to identify the daily behavior pattern because it records heterogene-ous data by various sensors. In this paper, we propose a method to define a graph structure with node and edge and to extract the daily behavior pattern from the generated lifelog graph. We use the graph convolution method to embeds the lifelog graph and maps it to low dimension. The graph convolution layer im-proves the expressive power of the daily behavior pattern by implanting the life-log graph in the non-Euclidean space and learns the patterns of graphs. Experi-mental results show that the proposed method automatically extracts meaningful user patterns from UbiqLog dataset. In addition, we confirm the usefulness by comparing our method with existing methods to evaluate performance.