Interactive Visualization of Machine Learning and Computer Vision with Dash by Xing Han
With the rise of advanced Neural Networks, Computer Vision and Machine Learning have become more and more useful, with novel and diverse applications. However, to achieve that high performance, state-of-the-art models have become increasingly more complex, and their results have become harder to interpret. To tackle this problem, I decided to create a suite of Dash apps that are able to improve the interpretability of the training and output of ML models.
The Object Detection App helps the user better understand the result of Object Detection algorithms beyond simple bounding boxes. When you play a video inside the app, not only can you see the bounding boxes drawn around the objects present in the video, but you can also view heatmaps and analytical pie charts generated by the output of the ML algorithm.
For all Deep Learning models, keeping track of accuracy and loss is an essential part of the training process, since they indicate how good your models are. The Live Model Training Viewer is a real-time visualization app that monitors core metrics of your Tensorflow model during the training so that you can quickly detect anomalies when you are training unstable algorithms or early prototypes. Its simplicity makes it extremely easy to setup, and provide much faster feedback than traditional viewers (e.g. Tensorboard).
I’m Xing Han, and I study CS at McGill with a focus on Deep Learning. I’m currently an intern at Plotly, where I work on incorporating Computer Vision and Machine Learning into Dash apps. Reach me on LinkedIn at / xing-han-lu