Introduction

I began my open-source journey largely by contributing to Keras.io and the adjoining libraries within the Keras Ecosystem. Listed below are some of my contributions till date. If you’re looking to make a contribution of your own or would like to collaborate, please feel free to reach out via email to discuss the same.

Blog Posts

• When Recurrence meets Transformers: This example, in joint work with Aritra Roy Gosthipaty, was done to implement a novel research paper known as Temporal Latent Bottleneck: Synthesis of Fast and Slow Processing Mechanisms in Sequence Learning. We implemented this with no references except for the paper, since no code was released by the authors. This was also subsequently the first TensorFlow/Keras implementation of the paper at the time.

• Using the Forward-Forward Algorithm for Image Classification: This example was one very close to me, as I wrote it as a pastime experiment while playing around with a new paper released by Prof. Geoffrey Hinton that detailed a technique for performing optimization without any backpropagation, but by instead performing 2 specialized forward-passes. This was a challenge to implement at first due to the inherent design of almost all Deep Learning frameworks to support backpropagation natively, but I created a workaround to make an acceptable implementation at the time. This was (arguably) the first TensorFlow/Keras implementation of this algorithm.

• Image Segmentation using Composable Fully-Convolutional Networks: This example details how to make use of Fully-Convolutional Networks for Semantic Segmentation. It details the implementations of the FCN-32S, FCN-16S, and FCN-8S models and their performance on a certain dataset.

• Electroencephalogram Signal Classification for action identification: This was my first example ever, that I wrote to get started with contributing to Keras.io as a side-project. It details the use of Conv1D layers to perform action identification from simple EEG scans of subjects. There are better solutions that exist now for doing this, but this example stays on as a baseline for what can be done and how an initial end-to-end pipeline would look like.