Ashwini Pokle

I am a sixth year PhD student in the Machine Learning Department, part of the School of Computer Science at Carnegie Mellon University. I am fortunate to be advised by Prof. Zico Kolter.

Previously, I completed my Master's in Computer Science from Stanford University. During my Master's, I also worked as a Research Assistant at Stanford Vision and Learning Lab (SVL) under supervision of Prof. Silvio Savarese. Prior to joining Stanford, I worked as a Software Development Engineer at Amazon where I was a part of the Prime Video team and Fulfillment Center Technologies team.

I received my Bachelor’s in Computer Science from Birla Institute of Technology and Science, Pilani, India. I have also spent a summer at USC Melady lab where I worked with Prof. Yan Liu.

I have previously interned at FAIR (Meta AI) and Bosch AI.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

Research

I am broadly interested in generative models (diffusion models, flows etc.) and deep equilibrium models (DEQs).

Training-free Linear Image Inverses via Flows

We propose a training-free method to solve linear inverse problems using pretrained flow models. Our experiments indicate that images restored via a conditional Optimal transport path, a specific parametrization of flows, are perceptually superior compared to those restored via diffusion paths.

Consistency Models Made Easy

We propose an alternative scheme for training Consistency Models (CM), vastly improving the efficiency of building such models. We express CM trajectories via a particular differential equation and argue that diffusion models can be viewed as a special case of CMs with a specific discretization. We can thus fine-tune a consistency model starting from a pre-trained diffusion model and progressively approximate the full consistency condition to stronger degrees over the training process. Our resulting method, which we term Easy Consistency Tuning (ECT), achieves vastly improved training times while indeed improving upon the quality of previous methods: for example, ECT achieves a 2-step FID of 2.73 on CIFAR10 within 1 hour on a single A100 GPU, matching Consistency Distillation trained of hundreds of GPU hours.

Deep Equilibrium Based Neural Operators for Steady-State PDEs

We demonstrate the benefits of weight-tying as an effective architectural choice for neural operators when applied to steady-state PDEs. We propose FNO-DEQ, a deep equilibrium model based architecture for solving steady-state PDEs that outperforms non-weight tied baselines with 4-6x parameters.

One-Step Diffusion Distillation via Deep Equilibrium Models

We propose Generative Equilibrium Transformer (GET), a deep equilibrium model-based variant of Vision Transformer that can efficiently distill sampling chain of diffusion models (EDM) into a single-step generative model.

Deep Equilibrium Approaches to Diffusion Models

We write the sampling process of DDIM as a deep equilibrium model. This allows us to do fast parallel sampling of DDIM by batching the workload, and enables use of memory efficient O(1) implicit gradients to backprop through the diffusion chain.

Path independent equilibrium models can better exploit test-time computation

We show that equilibrium models display strong upwards generalization on hard algorithmic tasks when these models converge to the same steady-state behaviour regardless of initialization, a phenomenon we term as “path independence”.

Contrasting the Landscape of contrastive and non-contrastive learning

We show through theoretical results and controlled experiments on simple data models that non-contrastive losses have a preponderance of non-collapsed bad minima. Moreover, we show that the training process does not avoid these minima.

Visually-Grounded Library of Behaviors for Generalizing Manipulation Across Objects, Configurations and Views

We propose a method for manipulating diverse objects across a wide range of initial and goal configurations and camera placements. We disentangle the standard image-to-action mapping into two separate modules: (1) a behavior selector which selects the behaviors that can successfully perform the desired tasks on the object in hand, and (2) a library of behaviors each of which conditions on extrinsic and abstract object properties to predict actions to execute over time.

Deep Local Trajectory Planning and Control for Robot Navigation

We proposed a navigation system that combines ideas from hierarchical planning and machine learning. The system uses a traditional global planner to compute optimal paths towards a goal, and a deep local trajectory planner and velocity controller to compute motion commands.

Translating Navigation Instructions in Natural Language to a High-Level Plan for Behavioral Robot Navigation

We propose an end-to-end deep learning model for translating free-form natural language instructions to a high-level plan for behavioral robot navigation. The proposed model uses attention mechanisms to connect information from user instructions with a topological representation of the environment.

Design and source code from Leonid Keselman's fork of Jon Barron's website