research

research topics explored as part of coursework.

Yu C., Hillier D. Modha N., Liu X. Y., PACE: Prototypical Adaptive Contrastive Embeddings

Deep Learning and Vision (CS701) with Prof. Sun Qianru

  • Built a novel, 2-staged approach for a constrained (<600MB model size and only IN-1K PTMs) 10-stage incremental fine-grained image classification task.
  • Model Performance topped the leaderboard, with 79.8% LA and 85.6% AA, using contrastive fine-tuning in stage 1 and logistic regression head trained on k-means prototypes of training samples in stage 2.
  • Link to Project GitHub and Project Report.