Yash's Pic

About Me

A Georgia Tech master's student working with Prof. Zsolt Kira and PhD student James Smith on continual learning of large pretrained models. In summer 2022, I worked with the Visual Search and Augmented Reality team at Amazon as an Applied Scientist intern. After graduating in 2020 from College of Engineering, Pune, I worked for a year at D. E. Shaw as a Software Developer. In the pandemic summer, I contributed to Kubeflow as a Google Summer of Code student. Previously, I have interned at the Vision and AI Lab of the Indian Institute of Science, Bangalore, working with Prof. R. Venkatesh Babu and PhD student Jogendra Nath Kundu on seq2seq modeling of human motion sequences.

Contact Details

  Contact Number:
+1-(404)-820-5409
  Email:
yashjakhotiya@gatech.edu
mailsforyashj@gmail.com

Resume

Education Details

  Georgia Institute of Technology
  GPA: 4.0/4.0
  Master of Science in Computer Science
  Dual Specialization in Machine Learning and Robotics
  College of Engineering, Pune
  CGPA: 9.2/10
  Bachelor of Technology, Computer Engineering
  Minor in Financial Engineering

Blog

moore-penrose-inversion

pinv >>> inv

A flow on matrix ranks, invertibility, singular value decomposition, Moore-Penrose inverse and the least squares method • Nov 7, 2021

linked-tori

Can you, like, REALLY explain me this?

A personal list of resources to look back to • Oct 11, 2021

kubelet_to_kernel

Containers, Container Runtimes, and What Kubernetes 'Docker' Deprecation Really Means

Kubelet to kernel space, and everything in between • Dec 20, 2020

kubeflow_components

GSoC '20: Kubeflow Customer User Journey Notebooks with Tensorflow 2.x Keras

An important milestone in my journey of open source • Aug 23, 2020

Research Publications

assistive-robotics

Improving Assistive Robotics with Deep Reinforcement Learning

NeurIPS '22, Deep Reinforcement Learning Workshop

•Yash Jakhotiya, Iman Haque
Assistive Robotics is a class of robotics concerned with aiding humans in daily care tasks that they may be inhibited from doing due to disabilities or age. While research has demonstrated that classical control methods can be used to design policies to complete these tasks, these methods can be difficult to generalize to a variety of instantiations of a task. Reinforcement learning can provide a solution to this issue, wherein robots are trained in simulation and their policies are transferred to real-world machines. In this work, we replicate a published baseline for training robots on three tasks in the Assistive Gym environment, and we explore the usage of a Recurrent Neural Network policy and Phasic Policy Gradient learning to augment the original work. Our baseline implementation meets or exceeds the baseline of the original work, however, we found that our explorations into the new methods was not as effective as we anticipated. We discuss the results of our baseline and analyze why our new methods were not as successful.

transformer-malware-detector

Adversarial Attacks On Transformers-Based Malware Detectors

NeurIPS '22, Machine Learning Safety Workshop

•Yash Jakhotiya, Heramb Patil, Jugal Rawlani
Signature-based malware detectors have proven to be insufficient as even a small change in malignant executable code can bypass these signature-based detectors. Many machine learning-based models have been proposed to efficiently detect a wide variety of malware. Many of these models are found to be susceptible to adversarial attacks - attacks that work by generating intentionally designed inputs that can force these models to misclassify. Our work aims to explore vulnerabilities in the current state of the art malware detectors to adversarial attacks. We train a Transformers-based malware detector, carry out adversarial attacks resulting in a misclassification rate of 23.9% and propose defenses that reduce this misclassification rate to half. An implementation of our work can be found at this https url.

relation_network

It Takes One to Know One? Idiomaticity Detection using Zero and One Shot Learning

NAACL '22, The 16th International Workshop on Semantic Evaluation

•Yash Jakhotiya*, Vaibhav Kumar*, Ashwin Pathak*, Raj Shah*
Large Language Models have been successful in a wide variety of Natural Language Processing tasks by capturing the compositionality of the text representations. In spite of their great success, these vector representations fail to capture meaning of idiomatic multi-word expressions (MWEs). In this paper, we focus on the detection of idiomatic expressions by using binary classification. We use a dataset consisting of the literal and idiomatic usage of MWEs in English and Portuguese. Thereafter, we perform the classification in two different settings: zero shot and one shot, to determine if a given sentence contains an idiom or not. N shot classification for this task is defined by N number of common idioms between the training and testing sets. In this paper, we train multiple Large Language Models in both the settings and achieve an F1 score (macro) of 0.73 for the zero shot setting and an F1 score (macro) of 0.85 for the one shot setting. An implementation of our work can be found at this https URL.

Projects

quora-logo

Can we avoid repeated questions by identifying similar intent?. • Oct 2021 - Dec 2021
In an online QnA platform like Edstem, Piazza, Stack Overflow or Quora, multiple questions with the same intent leads to redundancy, and waste of time, space and resources. We train a two stage model to identify if an incoming question is similar to an already asked one - a BERT Siamese Network, and PCA followed by a clustering algorithm for efficient inference. Our tuned classification accuracy is 86%. PCA reduces number of principal components to 15 with 99% variance retained. Finally, an elbow-determined 100 clusters help reduce our time overhead 100 times.

disentanglement-cat

Exploring Disentanglement in Discriminative Modeling. • Oct 2021 - Dec 2021
Can enforcing disentanglement in vision representation space lead to more generalizable and transferable discriminative models? We enforce disentanglement by introducing a novel KSG mutual information based regularizer. To measure disentanglement, we also contribute the first open-source implementation of Unsigned Correlation Coefficient (Wang et al. 2020). Experiments on CIFAR-10 showed improvement in classification accuracy by ~4% after disentanglement, which we then validated with GradCam and Activation Grid outputs.

ucsd_anomaly_example

Anomaly Detection in a Supervised Way. • Sep 2019 - Oct 2019
The UCSD Anomaly detection dataset is used in a supervised fashion here. A CNN Autoencoder is trained on the constituent frames to learn latent space representations. These latent space representations are then passed to a 2-layered LSTM network to detect if a sequence of frames contains an anomaly.

equity-price-movement-prediction

Time Series Modeling using Deep Learning with Credit Suisse India. • Feb 2019 - May 2019
This project's aim was time series modeling using deep learning. It was part of an elective course offered by Credit Suisse India. The code is structured into modules which can be reused. The core model used was a single-layered LSTM followed by a series of Fully Connected layers. On normalized data, an MSE of the order of 10^-3 was achieved.

bounding-box-prediction

Bounding Box Prediction around a Primary Object in an Image using Deep Learning. • Jan 2019 - Mar 2019
This project was a part of 'Flipkart GRiD – Te[a]ch The Machines | 2019' competition. An IoU (intersection over union) of 0.8 was achieved here on the unseen test dataset. Transfer learning not being allowed, the core network design along with its weights were arrived upon in a novel manner. Keras with a Tensorflow backend for training and matplotlib for plotting are used here.

human-detection

Human Detection in a Live Video Stream with HP Enterprises India. • Sep 2018 - Dec 2018
The objective here was to detect humans in a live video stream and to draw bounding boxes around them. Faster R-CNN with Inception-v2 pre-trained on the COCO dataset is used here for the task. Apart from Tensorflow, OpenCV and matplotlib are also used for various sub-tasks.

ctags

Support for Ctags in a Small Open Source Editor. • Sep 2017 - Nov 2017
This project's motive has been addition of tags-based code navigation functionality to bric, a small open source editor. The implementation uses UNIX's Exuberant Ctags generated tags file for navigation. Final PR to the editor was merged and the project code now resides in the editor source code. Many bugs were removed from and few small functionalities were added to the original source code while working on the project.

imu

Orientation Determination Using a 9-axis IMU. • Mar 2017 - Apr 2017
The end goal here was accurate determination of roll, pitch and yaw using IMU LSM9DS1. The microcontroller used was Atmel's ATSAME70. All the 9 axes of the IC were used to prevent drifts in absolute directions arising from building up of small errors in each of the 6 accelerometer-gyroscope axes. Complementary filters, which are not as accurate as Kalman filters but were sufficient for the project, were used to merge the sensor signals.