Rishika Bhagwatkar

We present a diverse benchmark for robust fine-tuning, investigating the effect of several design and training configurations. Our analysis both confirms and challenges prior assumptions, highlighting promising directions for further research.

We introduce a first-of-its-kind benchmark for real-world commonsense anomalies supporting three open-ended tasks. Through the high-quality examples and cognitively grounded annotations, we comprehensively evaluate the anomaly detection and understanding capabilities of popular vision-language models. 

We investigate the impact of model design choices on adversarial robustness in VLMs. More interestingly, we propose several inexpensive but highly effective prompt engineering techniques that provide substantial robustness improvements.

We present Lag-Llama a foundation model for univariate time series forecasting using a transformer architecture with lags as covariates. Pretrained on diverse data, it excels in zero-shot generalization and achieves state-of-the-art performance when fine-tuned, outperforming existing models.

In this work we hypothesize that real-world images and their corresponding synthetic images are different views of the same abstract representation. To enhance the quality of domain-invariant features, we increase the mutual information between the two inputs.

In this work, we present various limitations and drawbacks faced by current autonomous pipelines along with solutions to mitigate the same.

We posit that languages are linguistic transforms that map abstract meaning to sentences. We attempt to extract and investigate this abstract space by optimizing the Barlow Twins objective between latent representations of parallel sentences.

Just like sentences are series of words, videos are series of images. Inspired by the success of large language models in predicting language, we attempt to generate videos using a GPT and a novel Attention-based Discretized Autoencoder. 

In this work we study and discuss several approaches for generating videos, either using Generative Adversarial Networks (GANs) to sequential models like LSTMs. Further, we compare the strengths and weakness of each approach with the underlying motivation to provide a broad and rigorous review on the subject.