Van Tai (David) Nguyen

Led a project aiming to develop an AI imaging doctor based on deep learning to classify the skin lesion types. (GitHub to be disclosed)

· Coaching ML life cycle to junior staff.

· Decided the collection (approx. 3GB of images) image data of 7 classes of skin lesion.

· Analysed and chose soft attention-based Inception Resnet V2 model. Monitored the training of the DL model.

· Improved the classification results (98% accuracy) by applying the soft-attention-based Resnet V2 model. Analysed the results based on the classification report (achieved), ROC and AUC.

· Deployed model to GCP cloud.

End-to-End DL workflow to develop accurate and fast-computed model of 3D magnetic field distribution (the DL model consumes 4.6s whereas the FEM took more than 39000s). (GitHub to be disclosed)

· Reduced the computational time to generate 160 GB of random ground truth samples by 12 times due to deriving the semi-analytical expressions for the magnetic field and applying parallel algorithms. (Git Repo)

· Accelerated loading data during training deep learning model by converting the raw data into the TFRecord format.

· Improved the accuracy of the prediction to 0.99/1 by tuning hyperparameters with the assessment due to two sample z-test.

· Saved time on re-training deep learning model by proposing one sample z-test for the out-of-training intervals in the production.

· Distributed new knowledge by sharing the codes in a GitHub repository and publishing a high-quality journal article.

Developed and deployed a data-driven fast-computed model of magnetic force interaction between permanent magnets (the ML model required 10-4 s to execute which is in multiple orders of magnitude faster than the FEM). (Git Repo)

· Designed efficient AI architecture based on the MLP which reduce the number of input features from 7 to 5 without decreasing accuracy of the prediction.

· Decreased the difference between the predicted and ground truth to less than 4.2 % by tuning the hyperparameters.

· Analysed the importance of each input features towards their predictive power using a permutation strategy.

· Improved the user’s experience by developing a friendly software interface with the back end written in Flask and front end written on HTML and CSS. Deployed the model on EC2 instance (AWS) and GCP.

Developed model of the bead width and height from Wire and Arc additive manufacturing. (Code on Colab)

· Led a group of three people as a ML expert through effective encouragement and communication.

· Demonstrated the significance of the process parameters to the width and height using ANOVA.

· Quantified the uncertainty in the predicted results by utilizing the probabilistic, Bayesian and probabilistic Bayesian models.

· Assisted in writing high quality code for the project and analyzed the global sensitivity based on the first order Sobol index.

- Tweet sentiment analysis using LSTM.

- Predicted lung disease using transfer learning (VGG, Inception and MobileNet).

- Built Power BI dashboards for sale data.

- Analysed COVID data using Microsoft SQL (T-SQL).

- Developed the upper arm exoskeleton to assist people with disability or power augmentation.

- Derived fast-computed semi-analytical expressions of the magnetic field distribution.