About

Sushil Thapa

Sushil Thapa

Senior Machine Learning Engineer ยท Apple

Generative AI, large language models, multimodal machine learning, computer vision, and NLP.

Experience

2025–Present
Senior Machine Learning Engineer, Apple.
2022–2025
Software Engineer, Machine Learning, Intel.
2021
Researcher, Stanford AI Lab, Stanford University.
2019–2020
Machine Learning / Deep Learning Instructor, AI Microdegree™, fuse.ai.
2017–2020
Senior Machine Learning Engineer, Fusemachines.
2016–2017
Research and Development Engineer, Spark Tech Pvt. Ltd.
2014–2016
Robotics Engineer, Robotics & Automation Center. Represented Nepal at international robotics competitions: Techfest (IIT Bombay) and Kshitij (IIT Kharagpur).

Education

2021–2022
M.S. in Computer Science, New Mexico Tech. Research at Los Alamos National Laboratory.
2017–2018
Artificial Intelligence MicroMasters™, Columbia University (ColumbiaX).
2012–2016
Bachelor of Engineering, Institute of Engineering, Thapathali Campus, Tribhuvan University.

Publications and Preprints

Contrastive self-supervised learning paradigms
Survey on Self-Supervised Multimodal Representation Learning and Foundation Models
S. Thapa
TL;DR: A survey of self-supervised methods for multimodal representation learning and the rise of multimodal foundation models.
TL+KD architecture
On Effects of Knowledge Distillation on Transfer Learning
S. Thapa
TL;DR: Combining knowledge distillation with transfer learning (TL+KD), and how distillation during fine-tuning affects generalization and robustness.
Abstention logit separating in- and out-of-distribution
An Effective Baseline for Robustness to Distributional Shift
S. Thulasidasan, S. Thapa, S. Dhaubhadel, G. Chennupati, T. Bhattacharya, J. Bilmes
TL;DR: An abstention-class baseline for out-of-distribution detection that often beats prior state of the art on vision and NLP benchmarks.