DC25_PAPER_Track14_FoundationalModelApproachForSIPI_Kashyap_V2.pdf

1、 Information Classification:General Foundational Model Approach for SI/PI Analysis Using Large Language Model Techniques Priyank Kashyap,Hewlett Packard Enterprise()Yeujiang Wen,Hewlett Packard Enterprise()Yongjin Choi,Hewlett Packard Enterprise()Chris Cheng,Hewlett Packard Enterprise()Information C

2、lassification:General Abstract AI/ML techniques have been widely developed for SI/PI applications;however,a few challenges limit their wide adoption for SI/PI analysis.Training a good model requires collecting a large dataset and training hardware resources.However,an increased feature space or a mo

3、dified performance target requires new data collection and a complete retraining of the deep learning(DL)model.Further,modeling approaches utilizing neural networks do not offer insight into the physical aspects of the underlying model.This paper uses a transformer model to predict a channels impuls

4、e response given its input and output waveform.It then shows that the model contains a feature-rich latent space within the transformer stack.We explore this space and demonstrate that we can extract physical channel parameters to help us identify channel characteristics.The paper then establishes a

5、 framework for fine-tuning by extending the underlying datas range,specifically including new via topologies and increasing the operational bitrate.Then,the paper explores extending the base model to a different domain,modeling the impact of equalization with the channel,and demonstrating the abilit

6、y of transfer learning with the base model.This approach demonstrates the ability to future-proof our existing models and training efforts through computationally efficient fine-tuning and transfer learning techniques.Information Classification:General Author(s)Biography Priyank Kashyap is an AI/ML