i am a second year ms student at stanford in the computer science department. before that, i was a naive undergrad at iit bombay. i (pretend to) work on neural nonsense.
you can find me in one of my two natural eigenstates (images on the left).
i am interested in the mathematical foundations of machine learning, which translates to learning and infering over high-dimensions efficiently. i am interested in reinforcement learning. in the past, i have worked on sparse recovery, tractable inference, and sequential decision-making problems. i have done some internships where i did funky deep learning.
i am a sota neural net capable of (a) few-shot meta adaptation, (b) robust to adversarial attacks, and (c) in-context learning abilities. i can be trained very easily by showing me few slices of pizza instead to expensive gpus. i like working on fun real-world applications of machine learning. in my free time, i like to cook.
publications
- pac mode estimation using ppr martingale confidence sequences
- shubham jain, rohan shah, sanit gupta*, denil mehta*, inderjeet nair*, jian vora*, sushil khyalia, sourav das, vinay riberio, shivaram kalyankrishnan
- asymptotically optimal mode estimation of a discrete distribution by construcing confidence sequences (1v1, 1vr); applications to election polls and contract verification in blockchains
- recovery of joint probability distribution from
one-way marginals: low rank tensors and random projections
- jian vora, karthik gurumoorthy, ajit rajwade
- model a joint pmf as a low-rank tensor, recover the mode factors from 1D marginals estimated from random projections of data
- compressive signal recovery under sensing matrix errors combined
with unknown measurement gains
- jian vora, ajit rajwade
- compressive recovery when the sensing matrix is misspecified and there are unknown sensor gains
preprints
- scoring black-box models for adversarial robustness
- jian vora, pranay reddy samala
- adversarially robust models have sharper explanations and sparser lime weights, use this as a good subsitute for robust accuracy where trying to find attacks to the model can be hard
- plug&play multimodal generative model
allowing tractable inference
- jian vora, isabel valera, guy van den broeck, antonio vergari
- learn a joint distribution over multiple modalities while allowing for efficient marginalization, conditioning, likelihood evaluation using probabilisitc circuits on the fused latent space
- joint
speech-text pre-training with application to text adaptation for asr
- jian vora, cekic metehan, dhanush bekal, karel mundnich, srikanth ronanki, katrin kirchoff
- a hubert-based shared encoder for speech and text modalities for better coherence in embeddings using a small amount of paired multimodal data, application to text adaptation for asr
other selected research + course projects
- efficient learning of log-concave mixtures
- jian vora, vivek borkar
- random projections of data drawn from a mixture of log-concave densities are provably distributed as a gaussian mixture in the subspace
- tractable cooperative multi-agent reinforcement learning
- learn a joint policy over actions of all agents allowing for efficient inference by modeling q-function to be a factor graph
- improving inference in latent variable models
- improved inference in VAEs by reducing two gaps -- approximation gap by using hierarchical VAEs and amortization gap by performing unamortized inference
- a study on continuous optimization schemes for machine learning
- literature survey of three papers regarding continuous optimization in deep learning and reinforcement learning
- spatio-temporal action detection and classification
- participated in the trecvid'19 challenge which involved performing action detection and classification in videos. proposed a stage-wise architecture of object detection followed by tracking and activity classification
- continual learning for keyword spotting and speaker identification
- proposed a joint model to perform simultaneous kws and sid based on an interspeech 2021 challenge
- conditional style-gan for audio generative modeling
- Modified stylegan to allow for conditioning and trained on audio spectrograms
teaching
fall 2022: course assistant
for cs224v: conversational virtual assistants with deep learning at stanford
winter 2022: course assistant
for cs236g: generative adversarial networks at stanford
spring 2021: teaching assistant
for ma111: vector calculus at iitb
service
reviewer: tmlr, ieee tsp, aaai
leadership: manager, electronics and robotics club