Understanding Moravec's Paradox
7 days ago
- #Machine Learning
- #Robotics
- #Artificial Intelligence
- Moravec's paradox highlights that reasoning requires less computation than sensorimotor and perception tasks, contrary to common misinterpretations.
- The difficulty of tasks for machines can be understood through two components: search space size and reward sparsity.
- Chess is easier for machines due to a relatively small search space and frequent rewards, unlike robotics which involves large action spaces and sparse rewards.
- Humans evolved complex sensorimotor skills through billions of years of evolutionary search under natural selection.
- Simulating future states is crucial for tasks like chess but challenging in robotics due to environmental complexity.
- Neural networks, like LLMs, succeed by reducing search space via pre-training and handling sparse rewards through fine-tuning methods like RLHF.
- Reinforcement learning struggles with large search spaces and sparse rewards unless aided by pre-training or simulators.
- Moravec's paradox suggests that task difficulty for AI depends on search space and reward sparsity, predicting ease or challenge in solving future problems.