AI for Science
Schmidhuber: "build an optimal scientist, then retire"
Perception
Creativity
Curiosity
Logic
Creativity
Understanding
Exploration
Perception
Data analysis
Generalization
Inference
Communication
Attention
Memory
Action
Curiosity
Logic
Creativity
Understanding
Exploration
Perception
Data analysis
Generalization
Inference/learning
Communication
Attention
Memory
Human
Machine
Action
AI Scientist
Logic & memory
- Numerical computing (linear algebra, ODEs, optimization)
- Computer algebra, computational [insert math area]
- Proof assistants
- Automatic verification
- Formal systems, symbolic logic, etc.
What computers originally did...
- Databases
- Search (recall)
Huge capacity
Almost incorruptible
Data analysis, perception, attention
- Standard data analysis: statistics
- Sensors
- Machine vision, machine learning, classification
- Image processing, signal processing
- Unsupervised learning, finding structure in data (knowledge discovery)
- Machine learning with basic attention, identifying interesting parts of the data
Inference, learning
- Standard data fitting
- Supervised/ predictive learning. neural networks
- Reinforcement learning (needs to act on the world)
- Probabilistic programming
-
Program induction: neural Turing machines, differentiable neural computers. infering algorithms
Acting
- Robotics
- Actuation
- Lab robots. High-throughput, super precise, fast
- Reinforcement learning
"In the evolution of man, it is the hand that drives the subsequent evolution of the brain" ~ Jacob Brownoski
Communicating
- Natural language processing
- Speech synthesis/recognition
- We need to understand AI, and it needs to understand us:
- Data visualization
- Knowledge extraction (from papers)
- Summarizing
- Literature exploration, etc.
- Intelligent UIs
- AI+human conversations
Generalization, creativity
- One-shot and zero-shot learning
- Generative models
- Creativity ~ intelligent search
Human-level concept learning through probabilistic program induction
Understanding
- A result of integrating all the features of intelligence?
- Large scale integration?
- Need logic, generalization, language understanding, etc.
- Google knowledge graph / brain ?
- AGI
- Need to learn more from cognitive science / neuroscience?
- Spiking neural networks, neuromorphic computing
Consciousness
the ghost
Curiosity, exploration
Does AI need emotions?
Intuition?
Active learning
Current areas of application
Medicine
Literature exploration
Materials science
Physics
Biology
deck
By Guillermo Valle
