Project

Optimizing Generative Models through Joint Training of VAE and DDPM

• Period: Nov 2024 – Dec 2024

• Project Description:

  This project aimed to optimize the performance of generative models by integrating a Variational Autoencoder (VAE) with a Denoising Diffusion Probabilistic Model (DDPM). The project was conducted in three main stages:

  1. Independent Training: VAE and DDPM were trained separately to evaluate their individual performance.
  2. Joint End-to-End Training: The two models were integrated and trained jointly in an end-to-end manner, leveraging VAE for latent space representation and DDPM for improved sampling quality.
  3. Adaptive Conditioning with Flow-Based Prior: A flow-based prior was introduced to enable adaptive conditioning, leading to more refined and high-quality generation results.

[Step1 result]

[Step2 result]

[Step3 result]

Project

Enhancing DCN-V2 Performance with Transformer-Based Architecture

• Period: Oct 2024 – Dec 2024

• Project Description:

  This project proposed a novel architecture to enhance the performance of the DCN-V2 model.

  A hybrid structure was designed by integrating Transformer modules into the traditional DCN-V2 architecture, combining CrossNet and Transformer layers.

  In the proposed design, CrossNet captures explicit feature interactions, while the Transformer learns complex and implicit relationships.

  The output from the Transformer was used to generate stronger feature representations, which were then combined with the original DCN-V2 pipeline for final prediction.

  The model was trained on the Criteo dataset and outperformed various baseline models, demonstrating the effectiveness of the proposed approach.

• Architecture: