Core Focus

• Foundational Concepts
  • Basic Deep Learning principles
  • Convolutional Neural Networks (CNN)
  • Domain generalization techniques
• Advanced Learning Paradigms
  • Few-shot learning
  • Zero-shot (open-set) learning
• Neural Network Architectures
  • Graph Convolutional Neural Networks (GCNN)
  • Attention mechanisms
  • Transformers for NLP and Computer Vision
• Modern Applications
  • Transformer architectures in practice
  • Supervised and self-supervised learning with foundation models
  • Multimodal learning with foundation models

Introduction to Learning Principles

Supervised Learning

graph LR
    A[Unlabeled Images] --> B[Learning DNN]
    B --> C[Classification]
    D[Labeled Images] --> C

Transfer Learning

Transfer learning begins with large-scale supervised pretraining, where a Deep Neural Network (DNN) is trained on extensive labeled datasets like ImageNet-1k. This pretrained model can then be effectively adapted for downstream tasks, even with limited labeled data.

Some Other Examples

LT-ViT

Multi-modal analysis

<aside> 🧠

Supervised Learning Limitations

A bottleneck for building generalized models due to:

• Human labels being under-complete
• Human bias in labeling leading to less rich pretrained models for downstream tasks </aside>

Self-Supervised Learning

This leads to Self-supervised Learning, which aims to learn meaningful representations from unlabeled data by creating self-supervised tasks. This approach allows models to leverage vast amounts of unlabeled data, making them more robust and generalizable.

Think of it like educating a child—providing guidance rather than constant supervision.