paperlist.ai

ACML 2022 Papers

Asian Conference on Machine Learning

20252024202320222021202020192018201720162015

Decision breakdown

83papers

Counted from the official paper list; may differ slightly from the number the conference announced.

  • Poster
    83100.0%

Top 10 topics (click to filter)

No topic data yet.

Trends for all 207 research topics

Browse all papers

All83Poster83
A Novel Graph Aggregation Method Based on Feature Distribution Around Each Ego-node for Heterophily

Shuichiro Haruta, Tatsuya Konishi, Mori Kurokawa

Expand abstract

In this paper, we propose a novel graph aggregation method based on feature distribution around each ego-node (a node to which features are aggregated) for heterophily. In heterophily graphs, labels of neighboring nodes can be uniformly distributed. In such case, aggregated features by existing GNNs will be always similar regardless of the label of ego-node and fail to capture useful information for a node classification task. Since the existing methods basically ignore label distribution around the ego-node, we attempt to handle heterophily graphs through dynamic aggregations so that nodes with similar vicinity characteristics exhibit similar behavior. In particular, we adjust the amount of aggregation based on the features generated by higher-order neighbors, since they reflect the label distribution around each ego-node. By doing this, we can take the influence of distant nodes into account while adapting local structures of each node. Extensive experiments demonstrate that the proposed method achieves higher performance in heterophily graphs by up to 14.68% compared with existing methods.

Poster
A Novel Differentiable Mixed-Precision Quantization Search Framework for Alleviating the Matthew Effect and Improving Robustness

Hengyi Zhou, Hongyi He, Wanchen Liu, Yuhai Li, Haonan Zhang, Longjun Liu

Expand abstract

Network quantization is an effective and widely-used model compression technique. Recently, several works apply differentiable neural architectural search (NAS) methods to mixed-precision quantization (MPQ) and achieve encouraging results. However, the nature of differentiable architecture search can lead to the Matthew Effect in the mixed-precision. The candidates with higher bit-widths would be trained maturely earlier while the candidates with lower bit-widths may never have the chance to express the desired function. To address this issue, we propose a novel mixed-precision quantization framework. The mixed-precision search is resolved as a distribution learning problem, which alleviates the Matthew effect and improves the generalization ability. Meanwhile, different from generic differentiable NAS methods, search space will grow rapidly as the depth of the network increases in the mixed-precision quantization search. This makes the supernet harder to train and the search process unstable. To this end, we add a skip connection with a gradually decreasing architecture weight between convolutional layers in the supernet to improve robustness. The skip connection will help the optimization of the search process and will not participate in the bit width competition. Extensive experiments on CIFAR-10 and ImageNet demonstrate the effectiveness of the proposed methods. For example, when quantizing ResNet-50 on ImageNet, we achieve a state-of-the-art 156.10x Bitops compression rate while maintaining a 75.87$%$ accuracy.

Poster
3D Manifold Topology Based Medical Image Data Augmentation

Jisui Huang, Na Lei

Expand abstract

Data augmentation is an effective and universal technique for improving the generalization performance of deep neural networks. Current data augmentation implementations usually involve geometric and photometric transformations. However, none of them considers the topological information in images, which is an important global invariant of the three-dimensional manifold. In our implementation, we design a novel method that finds the generator of the first homology group, i.e. closed loops cannot shrink to a point, of 3D image and erases the bounding box of a random loop. To the best of our knowledge, it is the first time that data augmentation based on the first homology group of the three-dimensional image is applied in medical image augmentation. Our numerical experiments demonstrate that the proposed approach outperforms the state-of-the-art method.

Poster