Lecture
MIT 6.S093: Introduction to Human-Centered Artificial Intelligence (AI)
- Learning-based approaches like deep and machine learning are expected to dominate real-world applications over the next century, though they will not be provably safe, fair, or fully explainable in any step of the pipeline without extreme constraints or constant human supervision.
- Future AI systems will require a symbiotic relationship with humans for operation and annotation, as perfect autonomous systems are predicted to remain unachievable for the next 100 years, with human oversight triggered when systems specify uncertainty exceeding a specific threshold.
- The critical research direction for creating intelligent systems is machine teaching and data selection optimization, which aims to reduce annotation requirements by several orders of magnitude through active learning processes where machines query humans rather than relying on brute-force datasets.
- OpenAI, DeepMind, and MIT are currently pursuing the continuous monitoring and re-engineering of reward functions during training, with near-term safety strategies focusing on disagreement signals between multiple AI systems to trigger human intervention.
- Perceptual research is advancing toward detecting physical, mental, and social states via face, emotion, and speech recognition, yet current technology remains far from high-accuracy emotion detection, with a grand challenge defined as achieving 95% accuracy in classifying a person's desire for solitude after 30 days of data collection.
- Major grand challenges include building an AI to represent US Congress beliefs, passing the Turing test for social bots, and creating autonomous vehicles capable of handling hundreds of billions of miles involving diverse demographics such as teenagers aged 16 to 18 and older adults unfamiliar with AI.
- Success in artificial intelligence is predicted to rely on learning through natural human interaction rather than costly offline annotation, with MIT specifically predicting high-accuracy detection of driver disengagement from Autopilot systems using disagreement-based frameworks.