Efficient deep learning

This course addresses the challenges posed by modern neural networks, which require significant memory and computing power, making it difficult to deploy them on mobile and peripheral devices. In addition, the increasing scale and complexity of neural networks makes training very resource-intensive, often creating bottlenecks that slow the progress of AI applications. The course is divided into two main parts: improving inference efficiency and optimizing the training process.
In the first part, students will focus on improving inference efficiency by evaluating the effectiveness of neural networks and applying various compression techniques, such as pruning, tensor factorization and quantization, to create smaller, faster models without loss of accuracy. The second part of the course is dedicated to optimizing the training process, addressing the challenges of scaling and training complexity in modern AI models. Students will explore techniques such as memory-saving methods, including re-materialization (activation checkpointing) and offloading, as well as different types of parallelism-data, tensor, model, and pipeline parallelism-that are essential for efficient training.
Profiling neural networks to identify bottlenecks is emphasized throughout the course, helping students understand and solve performance problems for both inference and training. By the end of the course, students will have acquired the skills needed to optimize neural network performance and successfully deploy and train advanced AI models in real-world scenarios.

Deep learning tools