This repository contains my complete implementation for CMU 10-714: Deep Learning Systems.
The course provides a full-stack understanding of modern deep learning systems — from high-level framework design and automatic differentiation, down to low-level hardware acceleration and production deployment.

Throughout the course, I built Needle, a deep learning framework developed entirely from scratch.
Needle supports autograd, GPU acceleration, loss functions, data loaders, and optimizers, and can train a variety of modern neural network architectures including CNNs, RNNs, LSTMs, and Transformers.

10-714: Deep Learning Systems (CMU)
This course covers the fundamental building blocks of modern deep learning frameworks.
Students design and implement all components step-by-step, bridging the gap between theory and production-level systems.

Topics include:

• Computational graphs and reverse-mode automatic differentiation
• Tensor operations and broadcasting
• GPU acceleration using CUDA kernels
• Neural network layers and modules (Linear, Conv2D, BatchNorm, etc.)
• Loss functions and optimization algorithms
• Sequence models (RNN, LSTM)
• Transformer architectures
• Dataset loaders, training loops, and optimization pipelines

Needle is a minimalist deep learning framework that demonstrates:

• Automatic differentiation via a dynamic computation graph
• CPU and GPU backends for tensor operations
• Training utilities (optimizers, data loaders, model serialization)
• Neural network modules implemented on top of the autograd engine
• End-to-end models, including:
  • Convolutional Neural Networks (CNNs)
  • Recurrent Neural Networks (RNNs)
  • Long Short-Term Memory (LSTM)
  • Transformers for sequence modeling