Applied AI Systems
Our recent projects have a focus on deep learning explainability [1,6], model adaptation [4], graph neural networks [6,7] and knowledge graph enabled data sharing techniques. We apply our techniques on applications in a variety of domains ranging from fast radio burst search (FRB) radio astronomy [2,3] to crop yield prediction [4,5] and bio-origin classification [8] in agriculture.
Sub projects
Crop Yield Prediction with Knowledge Transferring
Crop yield prediction is important for precision agriculture, food security and grain marketing. Many data analytics methods have been developed for crop yield prediction. However, they often face data collection problems, either because farmers may not be willing to share their data due to commercial consideration or privacy concerns, or because data collected are unevenly distributed among regions and crop types. This results in issues such as how representative a data collection is for developing a machine learning model that is useful for a variety of fields.
We investigate how to transfer a model trained using one dataset to be applicable to another target dataset to address this problem. This will greatly reduce the data collection efforts from farmers while improving the prediction accuracy. The resulting transfer method enables a federated learning system to run on heterogeneous datasets where data owners can collaborate to gain deeper insight of crop growth with an optimized data collection procedure and make robust prediction of the yield.
Cost-effective Cattle Origin Classification
The project is aimed to distinguish the origins of beef products from farms in different regions of Australia based on single-nucleotide polymorphisms (SNP) data in the cattle genome. We develop a deep neural network model to learn the difference between breeds. We then applied model explanation techniques to understand the important features leading to the model decision. The important features are used as signature SNPs. This approach can significantly reduce the genome sample collection cost.
Mission Alignment
Our research aligns with the vision of multiple CSIRO Missions such as Trusted AgriFood Exports Mission and Infectious Disease Resilience Mission.
Relevent Publications
[1] Wu, H., Wang, C., Nock, R., Wang, W., Yin, J., Lu, K. and Zhu, L., 2020. SMINT: Toward interpretable and robust model sharing for deep neural networks. ACM Transactions on the Web (TWEB), 14(3), pp.1-28.
[2] Zhang, C., Wang, C., Hobbs, G., Russell, C.J., Li, D., Zhang, S.B., Dai, S., Wu, J.W., Pan, Z.C., Zhu, W.W. and Toomey, L., 2020. Applying saliency-map analysis in searches for pulsars and fast radio bursts. Astronomy & Astrophysics, 642, p.A26.
[3] Yang, X., Zhang, S.B., Wang, J.S., Hobbs, G., Sun, T.R., Manchester, R.N., Geng, J.J., Russell, C.J., Luo, R., Tang, Z.F. and Wang, C., 2021. 81 New Candidate Fast Radio Bursts in Parkes Archive. Monthly Notices of the Royal Astronomical Society.
[4] Wang C., Klinkmueller C., Lu Q., Herrmann C. and Chen S. 2020. Federated Learning for Crop Yield Prediction, CSIRO technical report.
[5] Al-Shammari, D., Whelan, B.M., Wang, C., Bramley, R.G., Fajardo, M. and Bishop, T.F., 2021. Impact of spatial resolution on the quality of crop yield predictions for site-specific crop management. Agricultural and Forest Meteorology, 310, p.108622.
[6] Wu, H., Wang, C., Tyshetskiy, Y., Docherty, A., Lu, K. and Zhu, L., 2019. Adversarial Examples for Graph Data: Deep Insights into Attack and Defense. IJCAI.
[7] Sun, K., Koniusz, P. and Wang, Z., 2020, August. Fisher-bures adversary graph convolutional networks. In Uncertainty in Artificial Intelligence (pp. 465-475). PMLR.
[8] Xu X., Wang C., Lu Q. and Zhu L., 2020. Data and model uncertainty management for biological origin classification applications, in Work Package 3: Biological Origin in the Supply Chain Integrity Phase 2 Initiative, CSIRO technical report.