Lecture
MIT Sloan: Intro to Machine Learning (in 360/VR)
- Future progress in machine learning is expected to rely on the shift from supervised learning with costly human annotations to unsupervised, semi-supervised, and reinforcement learning to reduce human involvement in data annotation.
- Continued advancement depends heavily on increased compute power, addressing the leveling off of Moore's Law through massive parallelization and super-efficient power implementations.
- Research aims to transition from narrow task-specific intelligence to general intelligence capable of learning from minimal data, though the current extent of machine learning's limitations remains unknown.
- Deep learning is projected to close the gap between AI, robotics, and machine learning by automatically discovering features across inputs like images, text, audio, and physical world data without expert encoding.
- Practical applications such as image segmentation, video colorization, translation, and end-to-end driving commands are currently feasible, yet the technology struggles with reasoning, planning, and closing the loop from sensors to effectors.
- Significant risks persist regarding robustness in real-world environments, where systems face challenges with occlusion, sensor spoofing, lighting variations, and susceptibility to noise that causes confident misclassification.
- Current data constraints involve sets ranging from hundreds of thousands to tens of millions of labeled examples, which are insufficient for billion-scale real-world systems, creating a barrier between pattern memorization and true understanding.
- Challenges in deploying autonomous systems include designing safe objective functions for urban environments, managing degrees of freedom in actuation, and achieving near-perfect accuracy required for life-critical tasks like pedestrian detection.
- The community envisions a future where neural networks operate with greater autonomy, potentially solving complex problems like traffic optimization and human coexistence, though policy frameworks for regulation and ethical coexistence are currently undeveloped.
- Philosophical inquiries regarding the nature of human intelligence and the limits of modeling human cognition within machines are expected to continue alongside technical development.