April : Week 2 Seminar

InnovAI April Week 2 Seminar

On April 10, the InnovAI April Week 2 Seminar was held from 7:00 PM to 9:00 PM in D407. The session featured a presentation by Wonbin Oh on SimCLR and CLIP.

SimCLR is a simple framework for contrastive learning of visual representations that uses data augmentations to create positive pairs from the same image and contrast them against other images in the batch. By training a neural network to maximize agreement between augmented views using a contrastive loss in the latent space, SimCLR learns useful features without requiring labeled data. This method achieves competitive performance on downstream tasks after fine-tuning, demonstrating the power of self-supervised learning with minimal architectural changes.

CLIP learns visual representations by leveraging natural language supervision from large-scale internet data. It trains an image encoder and a text encoder to align their outputs in a shared embedding space using contrastive learning, so that matching image-text pairs have higher similarity. This multimodal approach enables zero-shot learning, allowing the model to perform a wide range of vision tasks by simply providing textual descriptions, without task-specific fine-tuning.

Following the presentation, a Q&A session was held to discuss related topics.