Webinar: Artificial Intelligence: A Guide for Thinking Humans - 22 June 2020

Key Concepts

• Artificial Intelligence (AI)
• Deep Learning
• Neural Networks (Convolutional Neural Networks)
• Machine Learning
• Object Recognition
• Image Captioning
• Reinforcement Learning
• Adversarial Examples
• Barrier of Meaning
• Concepts
• Analogies
• Abstraction
• Statistical Correlations
• Bias in AI

1. Introduction and Historical Context

• AI has experienced cycles of optimism and pessimism.
• Early optimism: Frank Rosenblatt's perceptron (1950s) was predicted to "walk, talk, see, write, reproduce itself, and be conscious."
• Other early predictions: Claude Shannon (1961) predicted science fiction robots within 15 years; Herbert Simon (1965) predicted machines doing any work a man could do within 20 years; Marvin Minsky (1967) predicted AI substantially solved within a generation.
• Current optimism: Fueled by big data, parallel computing, and deep learning.
• Counter voices: Elon Musk called AI our "biggest existential threat" (2014); Douglas Hofstadter expressed terror at the blind rush in AI development.
• Mitchell's motivation: Hofstadter's views spurred her to investigate how close AI is to human intelligence, leading to her book.

2. What is Artificial Intelligence?

• AI encompasses various applications (object recognition, game playing, etc.).
• Definitions:
  • Nils Nilsson: "Building machines that perform tasks normally requiring human intelligence." (Problematic because tasks like chess don't necessarily require human-like intelligence.)
  • John McCarthy: "The study of the Common Sense world and how a system can find out how to achieve its goals."
  • Committee definition: "A branch of computer science that studies the property of intelligence by synthesizing intelligence." (Circular definition)
  • Recent paper: "AI is an anarchy of methods." (Reflects the diverse approaches used.)
• Historical approaches:
  • Logic: Encoding knowledge and reasoning using logical sentences. (Brittle, doesn't scale well.)
  • Statistics: Fitting models to data. (Successful in many domains, but limited.)
  • Biologically inspired methods: Inspired by the brain. (Currently dominant.)

3. The Deep Learning Revolution

• Inspired by the visual system in the brain (hierarchical layers of neurons in the visual cortex).
• Deep neural networks: Mimic the layered structure of the visual cortex.
• Convolutional networks: Input goes through hierarchical layers, mimicking the visual cortex. Layers compute increasingly complex features (edges, shapes, objects).
• Classification module: Outputs confidences (e.g., dog or cat).
• Object recognition: A major success of deep learning.
• Example: Google Image Search identifies a Great Pyrenees in a photo.
• Example: Google Photos can search for "fountains" in uploaded images.
• Facebook can find faces in images.
• ImageNet: A large, labeled dataset of over a million images used to train object recognition systems.
• ImageNet competition: Deep neural networks dramatically reduced error rates starting in 2012.
• Human performance: Andrej Karpathy estimated human error rate on ImageNet at 5% (based on his own limited testing).
• Media reports: Often incorrectly state that computers are better than humans at object recognition.
• Image captioning: Convolutional network feeds into a network that outputs words.
• Example: System captions an image as "A group of young people playing a game of Frisbee."
• Self-driving cars: Use deep neural networks to recognize objects on the road.
• Predictions: Many predicted fully self-driving cars by 2020 (not yet realized).
• Video games: Deep reinforcement learning used to play Atari games.
• Example: DeepMind's system achieved 1327% of human level on Breakout.
• Google acquired DeepMind in 2015 for over half a billion dollars.
• AlphaGo: System that beat the best humans at Go.

4. A Closer Look at Machines That Learn

• Overstatement: Deep learning systems "learn on their own" and "teach themselves."
• Reality: Systems need millions of labeled photos and hand-tuning.
• Challenge: Difficult to determine exactly what these complex systems have learned.
• Example: Student trained a network to distinguish animals from non-animals. The system learned to use blurry background vs. clear background as a shortcut.
• Firetruck example: A network trained to recognize firetrucks with 99% confidence misclassifies a firetruck in a different pose as a school bus, fire boat, or bobsled.
• Image captioning errors: "A horse is standing in the middle of the road" (no horse present). "A group of people sitting at a bus stop" (incorrect identification).
• Video game limitations: A network trained to play Breakout couldn't transfer its knowledge to a version where the paddle was shifted up a few pixels.
• Self-driving car accidents: Humans rear-end them because they stop suddenly due to misidentifying objects.
• Adversarial examples: Small, engineered noise added to an image can fool a network into misclassifying it (e.g., school bus becomes an ostrich).
• Face recognition hacks: Eyeglass frames with specific patterns can fool facial recognition systems.
• Traffic sign manipulation: Stickers on stop signs can fool self-driving car systems into thinking they are speed limit signs.

5. The Barrier of Meaning

• John Carlo Rota: "I wonder whether and/or when AI will ever crash the barrier of meaning."
• Machines don't extract the same meanings from images and language as humans do.
• This is the source of errors and brittleness in AI systems.
• Challenge: How to get machines to learn what we mean.
• Andre Karpathy's blog post: Highlights the vast difference between computer vision and human-level vision.
• Obama photo example: Understanding the photo requires intuitive physics, knowledge of social dynamics, and mental models of cause and effect.
• Concepts: Needed for abstraction and analogy.
• Example: The concept of "walking a dog" is flexible and can encompass various scenarios (multiple dogs, running, cats, etc.).
• Hofstadter and Sander: "Without concepts there can be no thought, and without analogies there can be no concepts."
• Forming and flexibly using concepts is the most important open problem in AI.

6. Unsolved Problems in AI

• Nearly all problems in AI are still unsolved.
• 1956 Dartmouth workshop proposal: Aims to make machines use language, form abstractions and concepts, solve problems reserved for humans, and improve themselves.
• Eric Horvitz: The 1956 proposal could be resubmitted today and still receive funding.
• All these problems are still open and ready for innovation.

7. Questions and Speculations

• Creativity is extremely important in intelligence, involving the ability to generate new things rather than just recognize them.
• Artificial neurons are very different from real neurons.
• Intelligence involves more than just learning; it also includes applying knowledge in different situations, abstraction, analogy, and reasoning.
• Bias in AI systems is a significant issue, stemming from skewed training data and other factors.
• Zero-shot learning (recognizing new objects without prior training) is possible in some cases but remains a challenge.
• AI includes rule-based systems and expert systems, not just machine learning.
• Machine learning is primarily inductive, while logical deduction is less used now.
• Reinforcement learning involves intermittent rewards from the environment, while unsupervised learning has no feedback.
• Most learning is done offline, but parallel hardware is speeding up real-time learning.
• Speech processing has become very good, but natural language understanding and translation still face challenges.
• AI is being used in investing, but much of the work is secretive.
• If a task can be done using statistics, it may be preferable to using more complex AI techniques.
• Emotional intelligence is an important area in human psychology but less researched in AI.
• The Turing test's relevance is debated, with some finding it a good measure of intelligence and others finding it easily fooled.
• International leadership in AI is a concern, particularly with advances made by China.

8. Conclusion

• AI has made significant progress, particularly in areas like object recognition and game playing.
• However, AI systems are not always reliable, can be easily fooled, and don't extract the same meaning as humans.
• The barrier of meaning remains a major challenge.
• Forming and flexibly using concepts is the most important open problem in AI.
• Many fundamental problems in AI remain unsolved, offering opportunities for innovation.