Investigated the impact of spurious correlations and dataset artifacts on pre-trained models in NLI

Used contrast sets, checklist sets, and adversarial challenge sets to assess model behavior and identify weaknesses in performance on challenging examples.

Implemented methods like adversarial data augmentation to enhance model generalization and reduce bias from spurious correlations.

Conducted in-depth analysis to identify specific error types and implemented focused training strategies to address performance issues in targeted subsets of data.

Evaluated model performance with accuracy metrics and visualizations to understand the effectiveness of fixes and identify improvements in generalization to real-world scenarios.

Analyzed and verified outputs from large language models (LLMs) like ChatGPT by conducting fact-checking against Wikipedia through two methods.

Method 1: Implemented a word overlap method for fact-checking, using metrics like cosine similarity and Jaccard similarity to predict supported vs. not supported facts.

Method 2: Applied a pre-trained DeBERTa-v3 model for textual entailment to classify facts as supported or not supported based on premise-hypothesis pairs.

Optimized the entailment model with pruning based on word overlap to improve performance and meet accuracy requirements.

Conducted error analysis on the entailment model's predictions, categorizing false positives and false negatives into fine-grained error types for further evaluation.

Implemented a Transformer language model to predict character occurrences and next characters in sequences using PyTorch.

Built a simplified Transformer encoder with self-attention and positional encodings to predict character occurrences (before or before-and-after) in sequences.

Extended the model to predict the next character in a sequence, achieving a perplexity ≤ 7 and training within 10 minutes.

Evaluated performance on perplexity and log-likelihood, requiring the model to output valid probability distributions for predictions.

Implemented optimization with SGD for a quadratic function and explore step size tuning.

Built a Deep Averaging Network (DAN) using pre-trained word embeddings and fine-tune model parameters.

Implemented batching in the neural network and handle padding for sequences of varying lengths.

Enhanced model generalization by addressing misspelled words with spelling correction or prefix embeddings.

Implemented perceptron and logistic regression classifiers for binary sentiment classification using unigram features.

Explored feature extraction techniques like unigrams, bigrams, and custom features to improve model performance.

Experimented with different learning schedules and feature modifications to optimize accuracy.

Designed machine learning models, including MLP, Transformers, and CNNs, to predict vehicle trajectories for autonomous driving tasks.

Utilized lane boundary and waypoint data to train models for trajectory prediction, with metrics focusing on longitudinal and lateral error.

Implemented a Transformer-based planner leveraging cross-attention and learned query embeddings for improved trajectory predictions.

Built a CNN planner to predict lane boundaries directly from images, bypassing segmentation and depth estimation challenges.

Developed convolutional networks for image classification, segmentation, and detection tasks.

Implemented models to classify racing track images and predict road segmentation masks with depth estimation.

Built a training pipeline from scratch, including data augmentation, optimization, and model evaluation using metrics like mIoU and MAE.

Experimented with advanced techniques like residual connections, skip connections, and multi-task loss functions.

Trained neural networks to classify SuperTuxKart images using linear, MLP, and residual models.

Built data loaders, loss functions, and models with PyTorch.

Used TensorBoard for tracking training and validation metrics.

Accelerated training with GPU resources for efficiency.

Ensured robust evaluation with train, validation, and test splits.