8.2 Deep learning architectures and applications

There are three main subfields in machine learning: speech and natural language processing (NLP), computer vision, and reinforcement learning. NLP has been successfully applied in business intelligence, voice assistant, machine translation, autocompletion and autocorrection, automated customer services, etc.

Computer vision is the core technology in self-driving cars, security (surveillance cameras, facial recognition), photo/video generation (which are terrifyingly good), and other entertainment services such as photo editing, photo filters, face swap.

Reinforcement learning is harder to deploy as the real-world environment is so much more complex than simulation, but we’ve seen the use of RL in ads bidding optimization, unmanned aerial vehicles (such as drones), and various robotic applications such as warehouse and production robots.

A company or a team might focus on a subfield. For example, the Siri team at Apple might focus on speech and natural language understanding, and the Autopilot team at Tesla might be more interested in computer vision. However, techniques in one subfield can be used for another, and there are tasks that have components from different subfields. There’s undoubtedly value in being a world-class expert in your niche subfield, but to get there, you might need to have knowledge of other subfields.