#13 Pedro Domingos: The Rise of The Machines

Aug 30, 2016
Overview

In this episode, Professor Pedro Domingos, a leading machine learning researcher at the University of Washington, discusses the nature of AI, the five schools of machine learning, and how algorithms are transforming decision-making. He explores the automation of white-collar jobs, the concept of human-AI collaboration, and the future of self-driving cars.

At a Glance
16 Insights
1h 4m Duration
17 Topics
8 Concepts

Deep Dive Analysis

Defining Artificial Intelligence and its Scope

Sources of Knowledge: Natural vs. Artificial

Machine Learning vs. Traditional Computer Programming

The Concept of a Master Algorithm

Understanding and Trusting Algorithm Outputs

Five Major Schools of Machine Learning

Implications of Algorithm-Generated Patents

Automation: White-Collar vs. Blue-Collar Jobs

Human-Machine Collaboration (Centaurs)

Adversarial Machine Learning and AI Arms Races

First-Mover Advantage in AI and Market Competition

Sociological Limits on Machine Decision-Making

The Singularity and S-Curves of Technology Growth

Deep Blue vs. AlphaGo: Different AI Approaches to Games

Challenges and Future of Self-Driving Cars

Contrasting Self-Driving Car Development Models

Influential Books and Current Reading

Artificial Intelligence (AI)

AI is a subfield of computer science dedicated to making computers perform tasks that typically require human intelligence. This includes capabilities like reasoning, common sense, language understanding, vision, and learning, aiming to create an intelligent artificial entity.

Machine Learning

Machine learning is a paradigm where computers program themselves by learning from data examples, rather than being explicitly programmed for each task. You provide input-output pairs (e.g., X-ray and diagnosis), and the computer figures out the algorithm to transform one into the other, often outperforming humans.

Master Algorithm

A master algorithm is a single learning algorithm capable of solving all different kinds of learning problems, much like a master key opens all doors. The ultimate goal of machine learning research is to develop such a grand unified theory of learning that can adapt to any data or task.

Bayesian Learning

Inspired by first principles, Bayesian learning quantifies belief in various hypotheses using probabilities, starting with a 'prior probability.' As new evidence is observed, beliefs are updated, making hypotheses consistent with the evidence more likely, allowing for decision-making based on weighted probabilities.

Symbolic Learning

This school of machine learning emulates the scientific method, where machines formulate and test hypotheses to explain data, then refine or discard them. It allows computers to discover knowledge much faster than human scientists, as demonstrated by robot scientists discovering new drugs.

Learning by Analogy

Inspired by psychology, this approach involves retrieving similar past situations from memory when faced with a new problem. The solution from the previous situation is then applied or transformed to fit the new context, proving to be a powerful method for learning and reasoning.

The Singularity

The singularity is a concept suggesting that if machines can learn to create increasingly better learning algorithms, intelligence will rapidly accelerate and surpass human capabilities, leading to infinite growth. This idea is often based on extrapolating exponential progress curves into the future.

S-Curves (Technology Growth)

S-curves describe typical technology growth, which appears exponential initially but eventually flattens out due to finite limits in the real world. This contrasts with the singularity's prediction of infinite growth, suggesting that technological advancements will reach new plateaus rather than continuing indefinitely.

?
What is artificial intelligence (AI)?

AI is a subfield of computer science that focuses on enabling computers to perform tasks that typically require human intelligence, such as reasoning, common sense, language understanding, vision, and learning.

?
Where does knowledge come from, both naturally and artificially?

Naturally, knowledge comes from evolution (encoded in DNA), experience (acquired through living), and culture (from communication and books). Artificially, a new source is computers discovering knowledge from data.

?
How does machine learning differ from traditional computer science programming?

In traditional computer science, humans write explicit algorithms for every task. In machine learning, the computer programs itself by analyzing input-output examples and figuring out the underlying algorithm, often performing better than human-designed ones.

?
Can algorithms patent their creations?

Yes, machine learning algorithms have already generated patents for electronic devices and other inventions, which work better than those designed by human engineers.

?
Which types of jobs are easier for AI to automate: blue-collar or white-collar?

Counterintuitively, white-collar jobs (like engineering, law, medicine) are often easier to automate than blue-collar jobs (like construction). This is because white-collar tasks often require abilities that don't come naturally to humans and are learned through education, while blue-collar tasks leverage deeply evolved physical abilities.

?
Can humans and AI work together to achieve better results than either alone?

Yes, in many fields, a combination of human and computer intelligence (referred to as 'centaurs') can outperform either humans or machines individually, leveraging their complementary strengths and weaknesses.

?
Do machine learning algorithms engage in 'arms races' with each other or with humans?

Yes, this is already happening in areas like spam filtering, stock market trading, and online advertising, where algorithms constantly adapt and evolve to exploit or counter the behaviors of other algorithms or human adversaries.

?
Is the 'singularity' (infinite growth of AI intelligence) a realistic prediction?

The concept of the singularity, suggesting infinite exponential growth of AI, is dubious because real-world technology growth curves are typically S-curves, meaning they eventually flatten out due to finite limits, rather than continuing indefinitely.

?
How does AlphaGo's victory in Go compare to Deep Blue's victory in chess?

Deep Blue used classic AI with extensive search and explicit feature evaluation, with no machine learning. AlphaGo, in contrast, extensively used machine learning (deep learning for pattern recognition and self-play) combined with game search, representing a significant leap in AI capability for complex games like Go.

?
What makes self-driving cars challenging to deploy widely?

The primary challenge for self-driving cars is interacting with unpredictable human drivers and pedestrians in uncontrolled environments like city streets. Self-driving cars can easily coordinate with each other, but human behavior introduces complexity.

?
What are the different approaches to developing self-driving cars?

Companies like Tesla and traditional automakers often take an incremental approach, gradually introducing automated features while still requiring human oversight. Google, however, advocates for a fully self-driving approach, arguing that mixed control is problematic due to human inability to quickly take over in complex situations.

1. Adopt Human-AI ‘Centaur’ Teams

Combine human and machine intelligence for optimal performance, leveraging their complementary strengths (e.g., human intuition/common sense with machine consistency/data processing). This approach has been shown to outperform either humans or machines alone in complex tasks like chess and medical diagnosis.

2. Re-evaluate Job Automation Risks

Recognize that white-collar jobs (e.g., engineering, law, medicine) may be more susceptible to automation than often perceived, while some blue-collar jobs are harder to automate. This understanding should inform career planning and skill development strategies.

3. Leverage ML for Scale & Consistency

Employ machine learning for tasks requiring high consistency, accuracy, and the processing of vast quantities of data, such as medical diagnosis. Machines often outperform humans in these areas due to their unwavering consistency and ability to process more information.

4. Anticipate Adversarial ML Behavior

Plan for and adapt to adversarial behavior when deploying machine learning systems, as users or competitors will change their behavior to exploit weaknesses. This requires dynamic and adaptive ML strategies, as seen in spam filters, stock markets, and online advertising.

5. Seek Untapped ML Opportunities

Identify domains or ’low-hanging fruit’ where machine learning has not yet been widely applied. These areas offer significant opportunities for innovation and market leadership, even with less advanced algorithms initially, by simply being the first to apply ML.

6. Combine Diverse ML Paradigms

Integrate different machine learning approaches (e.g., classic AI search with deep learning, or connectionist with evolutionary) to develop more robust and adaptive systems. This allows for solving complex problems that single-paradigm approaches cannot tackle alone.

7. Expand ML Data Inputs

Broaden the scope of data inputs for machine learning algorithms beyond traditional sources to include diverse, real-time, and unconventional data (e.g., social media, traffic, satellite imagery). This provides richer insights and allows algorithms to learn from factors humans might overlook.

8. Prioritize ML Explainability & Interaction

When developing or deploying ML systems, prioritize explainability and allow for rich, natural language interaction. User trust is crucial for adoption, and opaque ‘black boxes’ hinder this, making it important for algorithms to explain their rationale and accept feedback.

9. Delegate ‘How’ Decisions to AI

Delegate ‘how’ decisions (execution, logistics, optimization) to algorithms while retaining human control over ‘what’ decisions (goals, ultimate objectives). This allows humans to set the vision while machines efficiently achieve it.

10. Drive Change as Early Adopter

Don’t wait for established gatekeepers (e.g., doctors, IT departments) to adopt new technologies. Instead, embrace and demonstrate the value of new tools like ML systems as an individual or early adopter to drive broader organizational and societal change.

11. Avoid Mixed-Control Autonomous Systems

Do not implement autonomous systems where humans are expected to take over instantly from machine control. This creates dangerous situations due to human inattention and lack of context, making fully autonomous or fully human-controlled systems safer.

12. Combine Supervised & Reinforcement Learning

For complex learning tasks, combine supervised learning from existing expert data (e.g., human master games) with reinforcement learning through self-play. This strategy, exemplified by AlphaGo, can achieve superior performance.

13. Design for Autonomous-Only Environments

Where possible, design infrastructure and systems to separate human and autonomous agents. Human unpredictability significantly complicates autonomous operation, and autonomous systems can coordinate much more efficiently with each other.

14. Acknowledge Knowledge Uncertainty

Be aware that any knowledge induced from data is inherently uncertain, as generalization may not always be correct. Additionally, recognize the human tendency to be overconfident in knowledge derived from evolution, experience, and culture.

15. Understand Tech Growth as S-Curves

Recognize that technological growth, including AI, follows S-curves (initial exponential growth followed by flattening) rather than infinite exponentials. This implies plateaus and phase transitions rather than continuous, unchecked growth, setting realistic expectations.

16. Employ Meta-Learning Approaches

Utilize meta-learning, where one algorithm learns to combine or optimize the outputs of multiple other learning algorithms. This approach, used by Netflix and IBM Watson, achieves more sophisticated and effective results.

The thing that's exciting today is that there's actually a new source of knowledge on the planet, and that's computers. Computers discovering knowledge from data.

Pedro Domingos

Often, just by taking a basic machine learning algorithm, applying on a database of, for example, x-rays and diagnosis, you actually wind up with something that is better at, for example, you know, pathology, than a highly trained human being would be.

Pedro Domingos

People often think that the easiest jobs to automate are like the blue collar ones. But actually, our experience in AI is that it's actually more the opposite. It's often white collar jobs that are easier to automate.

Pedro Domingos

The best chess players in the world today are what are called centaurs in the community. They're a team of a human and a computer.

Pedro Domingos

The first part of an S curve is actually mathematically indistinguishable from exponential. So it's easy to look at that part and say like, Oh, exponential growth, we're headed for a singularity. Actually, what we have is one of these S curves and we're headed for a phase transition.

Pedro Domingos

What really makes life hard for self-driving cars is us, the humans.

Pedro Domingos
7
Number of directors on a venture fund board One of the directors is an algorithm.
30 million
Number of moves AlphaGo learned from human masters' games This was part of its initial learning phase.
5 years
Estimated time until a lot of self-driving cars are on the road This is an estimate by Pedro Domingos.
10-15 years
Estimated time until most cars are self-driving This is an estimate by Pedro Domingos.