A hybrid approach combines learned deep-learning modules with discrete algorithmic or symbolic modules. Pattern recognition and intuition come from learned models; reasoning, search, memory, planning, and causal understanding come from programs. This division matters because either an end-to-end neural network or standalone software can be insufficient on its own. section 19.4.5 section 20.3
AlphaGo and autonomous driving
AlphaGo combines human-designed Monte Carlo Tree Search with learned value and policy networks. A Waymo self-driving system combines a hand-coded 3D world model with deep-learning perception modules that update it from the vehicle’s surroundings. In both cases, learned continuous models supply perception or evaluation while discrete programs organize decisions and action. section 19.4.5
Use the distinction between the two sides to reason about a hybrid design:
- Deep-learning modules handle geometric primitives such as pattern recognition, perception, and intuition. section 19.4.5
- Algorithmic modules handle program-centric primitives such as reasoning, search, memory, and symbolic manipulation. section 19.4.5 section 20.3
- A future system may learn to use control flow, variables, recursion, and complex data structures dynamically instead of relying on painstakingly hardcoded logic. section 20.3
Figure 19.15 A learned program relying on both geometric primitives (pattern recognition, intuition) and algorithmic primitives (reasoning, search, memory)
The figure separates geometric primitives for pattern recognition and intuition from algorithmic primitives for reasoning, search, and memory. section 19.4.5
