Deep Learning

Exploiting meta-learning mechanism to enhance model generalization with dynamically changing workload of database, with the aim of decreasing model retraining cost and improving overall knobs evaluation performance.

Designed two phases (detect & classify) to identify birds species from sound segment; Applied Audio Spectrogram Transformer (AST), Vision Transformer (ViT), AlexNet, CNN-RNN network as backbone to both detect whether bird sound exists in segment and which kinds they are.

Propose a two components (prediction & decision) solution to help Alibaba Cloud (cloud host provider) maintain best-profits supply chain; Utilized LSTM, ARIMA, DeepAR time-sequential models to predict future demands; Designed dynamic safety stock policy to support decision.

Reproduced classical machine learning algorithm: Logistic Regression, Decision Tree(ID3), KNN, GMM, K-means, MLP, MRF, SVM and validated efficacy on public benchmarks.