Karush Suri

Karush Suri

Research Engineer at Google X

Karush Suri

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

About me

About me

About me

Posts

Discrete Stochastic Optimization

5 minute read

Published:

This post will cover stochastic optimization with discrete latent random variables. Unlike continuous random variables, discrete random variables encode data in a few bits. This allows us to capture relevant information effectively. However, differentiating through discrete variables is challenging. We will look at the challenges posed in discrete stochastic optimization and reparameterization methods which overcome these challenges.

Coarsening Graphs with Neural Networks

11 minute read

Published:

With the rise of large-scale graphs for relational learning, graph coarsening emerges as a computationally viable alternative. We revisit the principles that aim to improve data-driven graph coarsening with adjustable coarsened structures.

Variational Generalization Bounds

8 minute read

Published:

Recent advancements in generalization bounds have led to the development of tight information theoretic and data-dependent measures. Although generalization bounds reduce bias in estimates, they often suffer from tractability during empirical evaluation. The lack of a uniform criterion for estimation of Mutual Information (MI) and selection of divergence measures in conventional bounds hinders utility to sparse distributions. To that end, we revisit generalization through the lens of variational bounds. We identify hindrances based on bias, variance and learning dynamics which prevent accurate approximations of data distributions. Our empirical evaluation carried out on large-scale unsupervised visual recognition tasks highlights the necessity for variational bounds as generalization objectives for learning complex data distributions. Approximated estimates demonstrate low variance and improved convergence in comparison to conventional generalization bounds. Lastly, based on observed hindrances, we propose a theoretical alternative which aims to improve learning and tightness of variational generalization bounds. The proposed approach is motivated by contraction theory and yields a lower bound on MI.

Evolution-based Soft Actor Critic

6 minute read

Published:

Concepts and applications of Reinforcement Learning (RL) have seen a tremendous growth over the past decade. These consist of applications in arcade games, board games and lately, robotic control tasks. Primary reason for this growth is the usage of computationally efficient function approximators such as neural networks. Modern-day RL algorithms make use of parallelization to reduce training times and boost agent's performance through effective exploration giving rise to scalable methods, commonly referred to as Scalable Reinforcement Learning (SRL). However, a number of open problems such as approximation bias, lack of scalability in the case of long time horizons and lack of diverse exploration restrict the application of SRL to complex control and robotic tasks.

Stacked Capsule Autoencoders

4 minute read

Published:

Capsule Networks(CapsNet) have been growing in application and development ever since the radical breakthrough of Vector Capsules1. The fundamental idea behind colecting more information about the presence of an object in an image such as its pose, angle and depth remain a non-trivial open problem. CapsNet is a step forward in the direction and addresses the issue by spatially abstracting and taking into account rotational equivariance. This post is a review of the recent direction proposed the new Stacked Capsule Autoencoder paper2.

Benchmarking Policy Search using CyclicMDP

5 minute read

Published:

Policy search is a crucial aspect in Reinforcement Learning as it directly relates to the optimization of the algorithm in weight space. Various environments are used for benchmarking policy search including the famous ATARI 2600 games and MuJoCo control suite. However, various environments have longer horizons which force the agent to perform better at continuous timesteps. This is often a non-trivial problem when dealing with policy-based approaches. This post introduces a new environment called the CyclicMDP which is a long horizon discrete problem presented to the agent. The environment is competitive to modern-day RL benchmarks is the sense that it presents the agent with a simple problem but at a very long (ideally infinite) horizon.

DQN with Atari in 6 Minutes

6 minute read

Published:

This post will introduce Q-Learning, one of the most famous algorithms in Reinforcement Learning which utilizes Neural Networks as function approximators. All tutorials in the Reinforcement Learning sections require prior knowledge of Deep Neural Networks and it is recommended that you have a look at the notebooks in the Deep Learning section.

Capsule Networks for Digit Recognition in PyTorch

8 minute read

Published:

We are going to implement the Capsule Network from the recent paper on Dynamic Routing Between Capsules. The network consists of the novel Primary and Digit Caps Layers which perform nested convolutions. Dynamic routing, or more specifically Routing by Agreement, takes place in the Digit Caps Capsule Layer which we will discuss in more detail. So let's get started and begin with or standard imports from the PyTorch module.

portfolio

publications

talks

teaching