Projects

GAN-based Anomaly detection

Explored Optical-Flow and RNN-based autoencoder networks for understanding object dynamics in video datasets. Designed Generative Adversarial Network (GAN) to model the Spatio-temporal features of normal/abnormal events. Introduced the Channel-Attention (CA) module to conditionally enhance focus on the foreground objects. Achieved 4X faster computation from LSTM by employing the Temporal Shift Module (TSM) in the 2D CNN layer. Improved the AUC/EER performance metrics by 3-5% from the state-of-the-art models on benchmark datasets. Documented the procedure in a research paper that is presently being reviewed for publication

Read more

K-means for Evolving Datastream

Implemented the paper K-means for Evolving Data Streams. Utilized K-means++ initialization technique for improved initial centroid selection. Incorporated a restart mechanism to adapt to concept drift, ensuring accurate clustering results. Employed surrogate error functions to approximate true error, facilitating continuous adaptation without explicit drift detection.

Read more