Aditya Rastogi

I am a Machine Learning Engineer at benshi.ai in Barcelona, Spain. At Benshi, I implement deep learning algorithms (mostly contextual bandits) to develop interventions in digital health. I also work with Go and PostgreSQL to develop and maintain our data pipeline.

I graduated with a dual degree (bachelors and masters) from the Department of Computer Science and Engineering at IIT Kharagpur in Jul'21. I am fortunate to have been advised by Prof. Partha Pratim Chakrabarti and Prof. Aritra Hazra for my bachelor's and master's theses, in which I worked on self-supervised learning in computer vision.

I am interested in developing agents that can operate in the real world. I want to develop models that can do reasoning, in addition to the current tasks that they are good at. I have also worked as a digital consultant in Tokyo, Japan, because of which I can speak a bit of Japanese as well.

Meanwhile, I am also working on developing my blog further, mostly looking at creating backends for large applications these days. I also have a strong passion for Math (especially Advanced Linear Algebra) because of its fundamental essence.

I have done research internships at the University of British Columbia, Vancouver and the University of Sydney. I worked on large scale pattern matching and facial landmarks detection, respectively, in these internships. I have also interned at Goldman Sachs in Summer 2020. I like to simplify fascinating complex concepts (which often requires going back to first principles) through my blog posts, and visualizations.

Email  /  CV  /  Transcript  /  Twitter  /  Github  /  LinkedIn

An open-source synthetic data generator of user behavior for testing reinforcement learning algorithms in the context of mobile health interventions using open-source SDK for tracking, organizing, and labeling data ready for machine learning models.

Reflecting upon the potential of transformers and scaling them with an efficient version: Switch Transformers.

Neural Network Architecture, Self-learning, Board Symmetries, Playout Cap Randomization and other pieces to make it all work.

Understanding Monte Carlo Tree Search - the search algorithm that sits at the core of AlphaZero.

Coding the Chain Reaction game, and understanding amplification and distillation of knowledge in AlphaZero.

In this post, I try to explain the biological vision system, primarily in humans. I start with the techniques used to understand how vision works, and then look at how the vision system is wired in our brains.

Here’s how I used Selenium and Twilio’s Voice API in Python to make a phone call when vaccine slots got available near my location in India.

The SimCLR framework requires large batch-sizes to form a good representation space, because the negative pairs are generated in the same batch. Momentum Contrast V2 combines SimCLR's design improvements with the original MoCo framework to do self-supervised learning with less computational costs.

SimCLR is a simple framework for contrastive learning of visual representations. It showed that composition of data augmentations like color jittering and random crop play a critical role in learning good visual representations.

The GNU Toolchain is a set of programming tools in Linux systems that programmers can use to make and compile their code to produce a program or a library. This post explains this toolchain that contains GNU m4, GNU Make, GNU Bison, GCC, GNU Binutils, GNU Debugger and the GNU build system.

This post discusses a simple gradient approach ( by Simonyan et. al) to obtain saliency maps for a trained neural network in PyTorch.

This post solves the racetrack problem in reinforcement learning in a detailed step-by-step manner. It starts with constructing the racetrack environment in Pygame and then proceeds with solving this problem with the off-policy Monte Carlo control algorithm.

This post explains the policy iteration algorithm in Reinforcement Learning and uses it to solve Jack’s car rental problem given in the Sutton & Barto book.

AI

This is a simulation in which the goal of the rockets is to reach the yellow circle while avoiding the obstacles and the borders along their path. This page shows how a genetic algorithm solves this path planning problem.

This page is an implementation of the Expectation-Maximization (EM) algorithm from scratch. It fits mixture gaussian density to the points provided on the screen.

Games

Just like flappy bird, the goal is to swim between columns of pipes without hitting them. The spacebar can be used to control the fish.

You need to save your block from the other falling blocks in this game. Increase your score and climb up the levels. The falling blocks follow Poisson distribution to ensure that the occurrence of one falling event does not affect the probability that a second falling event will occur. The expectation of this distribution increases as the game progresses.

This is the classic 15 puzzle problem. I became quite interested in solving it in as less number of steps as I could. The goal is to place the tiles in order, by making moves that use the empty space. I find it quite interesting how we humans can find good heuristic functions based on both logic and intuition to navigate in the state space to arrive at the goal state and that too with constrained memory.

Other

An implementation of the Gift Wrapping algorithm for convex hull.

Perlin noise is a type of gradient noise used by visual effects artists to increase the appearance of realism in computer graphics. This page uses a 2D perlin noise distribution to create a force field in which a thousand particles are dropped on the screen and their locus is displayed to create beautiful designs.

Design courtesy of Jon Barron.