MIT 6.S093: Introduction to Human-Centered Artificial Intelligence (AI)

Core Thesis and Predictions

• Pure learning-based approaches (e.g., deep learning) will continue to dominate real-world applications over fine-tuned optimization-based models in the 21st century.
• Learning-based systems cannot be provably safe, fair, or explainable without deep integration of human supervision during both training and operation.
• The primary pathway to smarter AI is not solely better algorithms, but "machine teaching," which optimizes the selection of data for human annotation rather than just model parameters.
• Neural networks generalize reality based on sparse samples, inherently creating uncertainty that requires human intervention to manage.

Human-Centered Annotation (Machine Teaching)

• Active Learning: Algorithms should query humans for the most informative data subsets, reducing the need for brute-force annotation of massive datasets.
• Objective vs. Subjective Annotation:
  • Objective: Tasks like ImageNet labeling where single humans can accurately classify clear visual data.
  • Subjective: Ambiguous, ethical, or low-emotion tasks where crowd intelligence is required to converge on a consensus.
• Research Goals:
  • Minimize data requirements by orders of magnitude (e.g., achieving state-of-the-art on COCO object detection using only Wikipedia text/images).
  • Achieve state-of-the-art accuracy on MNIST training on a single example per digit.
• Reward Engineering: Humans must continuously adjust loss functions to encode societal values, preventing AI from exploiting reward loopholes (e.g., maximizing "green turbos" instead of finishing a race).

Real-World Operation and Interaction

• Uncertainty Signaling: AI systems must provide quantifiable uncertainty measures to trigger human supervision when operating in high-stakes domains (medical diagnosis, autonomous driving).
• Human Sensing: Systems must perceive human physical, mental, and social states using video, audio, and text (e.g., face detection, emotion recognition, body pose estimation).
• Collaborative Interaction: The human-AI relationship must shift from passive service to active, continuous collaboration where the AI communicates limitations and seeks help.
• Safety via Disagreement: Using "arguing machines" (ensembles of independent AI systems) where disagreement between models signals uncertainty, triggering human review (reducing error rates from ~8-10% to 2.8% in ImageNet tests).

Specific Perception Challenges

• Face Recognition:
  • Challenge: Requires invariance to pose, lighting, aging, makeup, and accessories while distinguishing highly similar identities.
  • Current State: DeepFace and FaceNet achieve near-human performance using embeddings to minimize Euclidean distance between matching identities.
  • Grand Challenge: Achieving 95% binary accuracy on whether a person "wants to be left alone" using long-term temporal data over 30 days.
  • Open Problems: Handling non-frontal views, interpretability of black-box decisions, and privacy preservation.
• Activity Recognition:
  • Challenge: Quantifying motion dynamics and resolving ambiguous start/end points of activities.
  • Current State: C3D (3D ConvNets) and two-stream networks (RGB + Optical Flow) detect actions in video sequences.
  • Open Problems: Localization of activities within long videos, handling multi-label concurrent activities, and integrating environmental context.
• Body Pose Estimation:
  • Challenge: Mapping 2D pixels to 3D joint orientations amidst occlusion.
  • Current State: Bottom-up approaches (detecting all joints then grouping them via parts affinity fields) allow real-time multi-person pose estimation.
  • Open Problems: Incorporating temporal dynamics and deformable body models beyond static skeleton joints.
• Emotion Recognition:
  • Current Reality: Systems detect caricatured expressions (smiling, frowning) rather than genuine emotional states.
  • Scientific Consensus: True emotion recognition remains a significant open problem far from current capabilities.

Societal and Ethical Implications

• Congressional Automation (Thought Experiment): An AI system could automate legislative representation by analyzing the beliefs, expertise levels, and issue prioritization of the population.
• Autonomous Vehicle Milestones:
  • Tesla has accumulated over 1 billion miles under autopilot.
  • Waymo has surpassed 10 million autonomous miles.
• The Grand Challenge of Scale: Ensuring safety and trust as systems transition from current miles to hundreds of billions of miles, involving diverse demographics (teens, elderly, non-tech users).
• Turing Test Evolution: The "Alexa Prize" challenge aims to create social bots capable of prolonged, natural language conversation that rivals human interaction.
• Symbiosis vs. Parasitism: Successful AI development requires shifting from parasitic human annotation (costly, offline) to symbiotic learning (learning naturally through human-AI interaction).