[Last updated: July 2021]
I have recently started working on personal projects in machine / deep learning. This is an overview of the (currently very short list of) projects that I have done so far.
To finally act upon my interest in machine learning, I am currently trying to gain and keep up an overall understanding of the fast advancing field, as well as become a competent practitioner and (hopefully soon) an active contributor in some specific areas.
Here are the links to my repos:
- Time Series Anomaly Detection via Prediction and Reconstruction
- Feature Visualization: Mini Lucid TF2
This project is to implement examples of two approaches to time series anomaly detection, one using prediction methods and one using reconstruction methods.
For each approach, we have selected a particular deep learning model -- DeepAR for prediction, TadGAN for reconstruction -- and demonstrated the end-to-end process with a dataset from the Numenta Anomaly Benchmark repository. The demonstrations entail, respectively, the use of SageMaker SDK for DeepAR training and inference, and a re-implementation of TadGAN in TensorFlow 2.
This project is to re-implement part of Lucid in TensorFlow 2. Lucid is a package for research in neural network interpretability, and is built on TensorFlow 1.
Specifically, we have so far covered roughly the part for feature visualization. The key idea is to visualize features learned in a convolutional neural network by maximizing the activations of its different components from an image.