NETO’S

RESEARCH

2021

Adversarial Multi-task Learning Enhanced Physics-informed Neural Networks for Solving Partial Differential Equations

Will be published in: 2021 International Joint Conference on Neural Networks (IJCNN)

Recently, researchers have utilized neural networks to accurately solve partial differential equations (PDEs), enabling the mesh-free method for scientific computation. Unfortunately, the network performance drops when encountering a high nonlinearity domain. To improve the generalizability, we introduce the novel approach of employing multi-task learning techniques, the uncertainty-weighting loss and the gradients surgery, in the context of learning PDE solutions. The multi-task scheme exploits the benefits of learning shared representations, controlled by cross-stitch modules, between multiple related PDEs, which are obtainable by varying the PDE parameterization coefficients, to generalize better on the original PDE. Encouraging the network pay closer attention to the high nonlinearity domain regions that are more challenging to learn, we also propose adversarial training for generating supplementary high-loss samples, similarly distributed to the training distribution. In the experiment, our proposed methods are found to be effective and reduce the error on the unseen data points as compared to the previous approaches in various PDE examples, including high-dimensional stochastic PDEs.

A novel approach to solve partial differential equations (PDEs) at high accuracy using multi-task neural networks, which is trained with adversarial examples. Our models have surpassed the previous approaches, such as Physics-informed neural networks (PINNs), in terms of prediction accuracy.

2020

Encoder-decoder  based  convolutional  neural  networks  with  multi-scale-aware modules  for  crowd  counting

Conference: International Conference on Pattern Recognition, ICPR, Italy, January 10-15, 2021.

In this paper, we propose two modified neural networks based on dual path multi-scale fusion networks (SFANet) and SegNet for accurate and efficient crowd counting. Inspired by SFANet, the first model, which is named M-SFANet, is attached with atrous spatial pyramid pooling (ASPP) and context-aware module (CAN). The encoder of M-SFANet is enhanced with ASPP containing parallel atrous convolutional layers with different sampling rates and hence able to extract multi-scale features of the target object and incorporate larger context. To further deal with scale variation throughout an input image, we leverage the CAN module which adaptively encodes the scales of the contextual information. The combination yields an effective model for counting in both dense and sparse crowd scenes. Based on the SFANet decoder structure, M-SFANet’s decoder has dual paths, for density map and attention map generation. The second model is called M-SegNet, which is produced by replacing the bilinear upsampling in SFANet with max unpooling that is used in SegNet. This change provides a faster model while providing competitive counting performance. Designed for high-speed surveillance applications, M-SegNet has no additional multi-scale-aware module in order to not increase the complexity. Both models are encoder-decoder based architectures and are end-to-end trainable. We conduct extensive experiments on five crowd counting datasets and one vehicle counting dataset to show that these modifications yield algorithms that could improve state-of-the-art crowd counting methods.

I invented neural crowd counter which achieved the state-of-the-art performance in 2020 on various datasets. See this URL for the historical leaderboard and this URL for the code.

2019

Using label noise filtering and ensemble method for sentiment analysis on Thai social data

Published in: 2019 14th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP)

Sentiment analysis is an essential task for social listening, especially in service and product analysis. Prior works on sentiment analysis, especially in Thai language, mostly focus on the improvement of model architecture without considering error propagation from word tokenizers or noisy text from social media. In this paper, three contributions are proposed for implementing social analysis model. First, text pre-processing is used to mitigate noise from input texts. Second, robustness towards word segmentation is enhanced by using an ensemble process with two tokenizers. Lastly, the training process inspired by Co-training method is proposed in order to filter label noise within the data. In the experiments, the model achieves 2.56% improvement on the average macro f-l score when compared with the baseline models in social media data.

I was one of the early researchers for building Thai sentiment analysis software @KBTG. This research tackled the problem of wrong word segmentation and label inconsistency using an ensemble method with label noise filtering technique. This was the best technical paper @iSAI-NLP 2019 (https://isai-nlp2019.aiat.or.th/).

A Light-weighted and Accurate Crowd Level Estimation Model

Conference: 11th ECTI-CARD 2019, Ubon Ratchathani, Thailand

This study aims to develop a light-weighted and accurate crowd-levels estimation model using the transfer learning method from the multi-column convolutional neural network (MCNN) trained for crowd counting task. We make our own training dataset by using pseudo-label predicted from an existing state-of-the-art model, CSRNet, and classify the output into various sets of crowd-levels by k-means clustering algorithm, from low to high. We also propose a more generalized crowd-counting model that yields similar performance to CSRNet on the ShanghaiTech B test set but achieves better performance in our own dataset without any further training. With this method, we are able to train our network from the dataset that we collected locally without human effort on labeling the data. The network architecture is only a branch of MCNN to reduce model complexity. Finally, we evaluate our model with the test set and report its accuracy.

I implemented state-of-the-art neural networks for Crowd counting task, which are able to run on edges devices such as Raspberry Pi. This was my undergraduate thesis @Chulalongkorn university and also won the second-best project @Thailand National software competition (NSC) 2019 (https://www.nectec.or.th/nsc/).