Nishi Mehta Email and Phone Number

• Formulated a novel method to explain the text generated by GPT like causal la by attributing the most relevant training instances from the corpus• Computed Integrated Gradients across the neurons of feed-forward layer to identify the knowledge neurons present in the transformer• Utilized Graph Neural Network Autoencoder to encode the attention patterns contained in the Attention Heads• Extracted the most influential Training Datapoints for any specific text generation by combining the information stored in knowledge neurons and Attention Heads• Developed a method that outperforms prior art methods on multiple datasets by a significant margin while being Resource Efficient