Dheeraj Mekala

Dheeraj Mekala

Ph.D. Candidate in Computer Science

University of California, San Diego

Biography

I am a Ph.D. Candidate in the Computer Science department at the University of California, San Diego working with Prof. Jingbo Shang. I completed my Bachelor Of Technology in Computer Science And Engineering from the Indian Institute of Technology, Kanpur in 2017. For summer 2023, I am interning at FAIR London, Meta AI with Dr. Jane Yu and Dr. Jason Weston, working on improving the tool use capability of large language models. Previously, I interned at Microsoft Semantic Machines in 2022, and Amazon Science in 2021.

Current Research

My research is centered on understanding data and the development of data-driven approaches to enhance NLP pipelines, with a particular emphasis on reducing annotation and training costs. I’m actively exploring the following inquiries:

  • Data Quality for Performance I examine the data landscape through the lens of difficulty, diversity, and noise, striving to understand whether we could find the optimal volume of data necessary for achieving specific performance targets.

  • Leveraging Noisy Data Given that noisy data is often more readily available and cost-effective to obtain compared to clean data, I’m exploring the extent to which we can advance NLP tasks with noisy data and weak supervision [1], [2], [3].

  • Enhancing Model Awareness How can we empower language models to be self-aware of their capabilities and train them to provide well-calibrated predictions, making their scores more reliable? I’m also interested in enabling them to use external tools & resources when they are not confident.

I’m also extremely enthusiastic about designing goal-driven language assistants, and this area presents an abundance of intriguing questions to explore. In my vision statement, I delve deep into some of the questions including:

  • Empowering Language Assistants One fundamental feature I envision for language assistants is their capacity to seamlessly integrate with a diverse range of tools. How can we train these assistants to adapt and effectively utilize tools that are new or previously unseen?

  • Enhancing Collaboration Building an assistant that can collaborate with humans to simplify their tasks is a pivotal challenge. This entails improving the assistant’s ability to ask clarifying questions, thereby fostering a productive partnership.

  • Proactivity in Assistants Are existing language assistants proactive, and how do we define and quantify their proactiveness? Exploring ways to train an assistant to be proactive is another intriguing aspect of this research.

Apart from Academics, I enjoy spending time playing Ukulele, playing Football(soccer) and I rarely write too. Checkout my blog!

Interests

  • Machine Learning
  • Natural Language Processing

Education

  • PhD in Computer Science, 2025 (expected)

    University of California, San Diego

  • MS in Computer Science, 2021

    University of California, San Diego

  • B.Tech. in Computer Science, 2017

    Indian Institute of Technology, Kanpur

Publications

Quickly discover relevant content by filtering publications.
(2024). TOOLVERIFIER: Generalization to New Tools via Self-Verification. Arxiv.

PDF

(2024). MORL-Prompt: An Empirical Analysis of Multi-Objective Reinforcement Learning for Discrete Prompt Optimization. Arxiv.

PDF

(2023). SELFOOD: Self-Supervised Out-Of-Distribution Detection via Learning to Rank. EMNLP Findings 2023.

PDF Code

(2022). ZEROTOP: Zero-Shot Task-Oriented Semantic Parsing using Large Language Models. EMNLP 2023.

PDF

(2022). LOPS: Learning Order Inspired Pseudo-Label Selection for Weakly Supervised Text Classification. EMNLP Findings 2022.

PDF Code

(2022). Leveraging QA Datasets to Improve Generative Data Augmentation. EMNLP 2022.

PDF Code

(2022). Progressive Sentiment Analysis for Code-Switched Text Data. EMNLP Findings 2022.

PDF Code

(2021). Coarse2Fine: Fine-grained Text Classification on Coarsely-grained Annotated Data. EMNLP 2021.

PDF Code

(2021). BFClass: A Backdoor-free Text Classification Framework. EMNLP Findings 2021.

PDF Code

Experience

 
 
 
 
 

Data Scientist

Sprinklr

Apr 2018 – Jul 2019 Gurgaon, India

Part of Machine Learning team:

  • Architected and built most of Sprinklr AI’s visual insights module that is now used by over 1200 Sprinklr clients.
  • Developed in-house computer vision models for visual sentiment, gender, age, inappropriate content detection in images and videos.
  • Implementation was done using asynchronous programming and as a result, throughput was increased by 65% and total resources cost reduced by 50%.
  • Built in-house computer vision model that identifies the font and suggests similar fonts from an image.
  • Developed a dockerized auto-scaling python-based framework which is deployed in kubernetes for image classification. It works over a stream of data published to Kafka and thus is auto-scaled based on lag in Kafka queue.
  • Developed a scalable system capable of running classification models over 500 million messages per day using the latest technologies like Caffe, Tensorflow, Kafka and Elasticsearch.
  • Deployed a centralized monitoring environment(Grafana, InfluxDB) which gather system metrics as well as docker run-time metrics.
 
 
 
 
 

Product Engineer

Sprinklr

Jul 2017 – Apr 2018 Gurgaon, India

Part of Paid Advertising team:

  • Implemented an end to end pipeline that incorporates DoubleClick tracking in ads for integrated reporting.
  • Expanded the reach of the product by integrating Ads APIs of various social media channels like LinkedIn, Twitter, Google DCM.
  • Researched, Designed and Implemented core functionalities in backend code to improve the feature of importing and exporting ads which is the primary way, the users undergo to create ads.
 
 
 
 
 

Machine Learning Intern

Microsoft India

May 2016 – Jul 2016 Bangalore, India
Developed tree-based models for predicting the ideal assignment candidate for a case in Microsoft Dynamics CRM.
 
 
 
 
 

Software Development Intern

ASnTech & Engineering Services

Dec 2015 – Jan 2015 Hyderabad, India
Designed and implemented an algorithm to speed up search queries related to the location of a vehicle, from 120 seconds to 5 seconds.

Recent Posts

March 2021 NLP Reading:

The blog post summarizing a few papers from EMNLP 2020 and some recent papers that I have enjoyed reading in March 2021.

Popular Topics

Contact