Artificial Intelligence: A Guide for Thinking Humans

Name: Artificial Intelligence: A Guide for Thinking Humans
Uploaded: 2019-11-13T04:24:36.000Z
Duration: 2 h 40 min 42 s

Welcome and Introduction

Opening Remarks

The speaker welcomes the audience, expressing gratitude for their attendance despite unfavorable weather conditions.

Acknowledges the Institute and Lensic Theatre as exceptional venues for hosting events.

Alan Turing's Predictions

Turing's Influence on AI Perception

Reference to Alan Turing’s belief that by the end of the 20th century, discussions about machines thinking would be commonplace without contradiction.

Mention of Ian McEwan's novel "Machines Like Me," where Turing survives past his historical death and revolutionizes AI.

The Protagonist's Journey

Charlie Friend and His Robot Adam

In the novel, protagonist Charlie acquires an AI robot named Adam, inspired by Turing’s contributions.

Charlie expresses disdain for algorithms that intrude into personal life decisions, highlighting concerns about machine autonomy.

Exploring Machine Intelligence

Questions About Superintelligence

The speaker raises questions regarding whether superintelligent machines are mere simulations of human intelligence or fundamentally different entities.

Introduction of Mellie Mitchell

Background on Guest Speaker

Introduction of Mellie Mitchell as a distinguished professor at Santa Fe Institute with a strong academic background in complex systems and computer science.

Mention of her notable works including "Artificial Intelligence: A Guide for Thinking Humans" which is currently sold out but can be ordered from bookstores.

History of Artificial Intelligence

Early Developments in AI

Discussion begins on the origins of artificial intelligence dating back to 1958 with Frank Rosenblatt’s invention of the perceptron, a precursor to modern neural networks.

The perceptron was capable of recognizing handwritten letters and was funded by the Office of Naval Research.

Predictions About AI Evolution

Historical Expectations vs. Reality

Early predictions from figures like Claude Shannon suggested rapid advancements in AI capabilities within decades.

Herbert Simon claimed machines would perform any human task within 20 years; however, these predictions have not materialized as expected.

Contemporary Views on AI Potential

Current Perspectives on AI Impact

Andrew Ng likens AI to electricity in its transformative potential over time.

Elon Musk warns that AI poses significant existential risks, describing it as "summoning the demon."

Future Considerations in AI Development

Ongoing Questions About Superintelligence

The speaker reflects on media portrayals and public discourse surrounding potential future scenarios involving superintelligent robots and their implications for humanity.

What is the Current State of AI?

Introduction to AI

The speaker investigates the current state of artificial intelligence (AI), leading to the writing of a book on the subject.

Discussion includes defining AI, which remains ambiguous, and highlights advancements like chess-playing machines, speech recognition, GPS navigation, virtual assistants, machine translation, facial recognition, and self-driving cars.

Defining Artificial Intelligence

Nils Nilsson's definition: AI involves building machines that perform tasks typically requiring human intelligence; however, determining which tasks require intelligence is complex.

Historical perspective: Chess was once seen as a pinnacle of intelligence until brute-force algorithms allowed machines to outperform humans in the game.

Evolving Definitions and Approaches

John McCarthy defined AI as studying common sense and goal achievement in systems; this definition lacks clarity.

A recent definition describes AI as an "anarchy of methods," indicating its diverse approaches and evolving nature.

Methods in Artificial Intelligence

Three main methods identified:

Logic: Programming for logical deductions but often brittle when faced with new scenarios.

Statistics: Learning from data became prominent after logic's limitations were recognized.

Brain Emulation: Simulating brain processes has gained traction recently despite historical challenges.

Machine Learning and Deep Learning Revolution

From the 1950s to the 1980s, machine learning was a minor aspect of AI; it focuses on data-driven learning rather than rule-based programming.

In recent years, deep learning has surged ahead within machine learning due to its effectiveness in various applications like speech recognition and facial recognition.

Understanding Deep Learning Systems

Deep learning systems are inspired by brain functions; they process visual information through layered structures similar to those found in biological brains.

The speaker humorously notes potential criticism from neuroscientists regarding oversimplifications about how deep learning mimics brain activity.

Understanding Deep Neural Networks and Their Evolution

The Structure of Neural Networks

The architecture of neural networks consists of multiple layers where neurons in one layer feed into the next, allowing for the separation of light into features like edges, which are then built up to recognize shapes and objects.

This model is inspired by early concepts of brain function from around 1950, with convolutional neural networks (CNNs) simulating this layered processing to classify images effectively.

The term "deep" refers to having many layers of simulated neurons, contrasting with "shallow" networks that have only one or two layers. These layers help in classifying objects accurately.

Learning occurs through adjusting the strengths between neuron connections, similar to how deep neural networks modify simulated connections based on classification accuracy.

Achievements in Image Recognition

In recent years, deep neural networks have significantly advanced image recognition capabilities; for instance, Google Image Search can identify dog breeds accurately from photos without prior labeling.

Google Photos exemplifies this technology by allowing users to search untagged images based on content recognition, demonstrating the system's ability to understand context without explicit labels.

Facebook utilizes facial recognition algorithms powered by deep neural networks that learn from millions of labeled faces uploaded by users, enhancing their tagging suggestions.

The Role of ImageNet in Training Models

A crucial dataset for training these models is ImageNet, comprising 1.5 million labeled images sourced from the internet through crowdsourcing efforts.

Competitions using ImageNet have driven advancements in machine learning and computer vision as researchers strive to improve object identification accuracy over time.

Progress Over Time

Historical data shows a gradual decrease in error rates among programs identifying objects within ImageNet; initial attempts had about 28% error rates before significant improvements were made with deep learning techniques introduced in 2012.

The introduction of deep neural networks led to a dramatic drop in error rates due to increased data availability and powerful parallel computing resources.

Human vs. Machine Performance

Current models often outperform humans in object recognition tasks; however, reported human performance metrics were based on tests conducted by individual researchers rather than a comprehensive assessment across multiple subjects.

AI Misconceptions and Developments

Understanding AI's Limitations

A researcher reported a 5% error rate in AI image recognition, highlighting the need for skepticism regarding media claims about AI capabilities.

The media often exaggerates AI advancements, particularly in self-driving cars, which are still reliant on human oversight despite ongoing promises of full autonomy.

Self-Driving Cars: Current State and Future Promises

Experimental self-driving cars equipped with cameras and sensors can navigate highways but require human intervention for safety.

By 2020, there were high expectations for millions of self-driving cars on the road, with Elon Musk claiming significant advancements in Tesla's Autopilot feature.

Game Playing as an AI Benchmark

Since the inception of AI, game playing has been a focal point; DeepMind utilized deep neural networks to teach machines to play classic Atari games from the 1970s.

The learning process involved simulating gameplay where the machine learned through trial and error by adjusting its strategies based on rewards and penalties.

Learning Mechanisms in Machines vs. Humans

Contrary to popular belief, machines do not learn like humans; they require vast amounts of labeled data (e.g., millions of images) to classify effectively.

Unlike children who learn contextually without explicit labeling (e.g., parents naming objects), machines depend heavily on structured training data to achieve proficiency.

Understanding Machine Learning Limitations

The Role of Human Design in Neural Networks

Humans must meticulously design machine learning systems, particularly deep neural networks, which consist of various layers and neurons.

Current machines lack the ability to learn independently; human expertise is essential for creating effective neural network architectures.

Challenges with Edge Cases in Self-Driving Cars

The "long tail problem" highlights issues faced by self-driving cars when encountering rare scenarios not covered during training.

An example illustrates how Tesla's autopilot struggled with snowstorm conditions due to a lack of prior exposure to such edge cases.

Unpredictable Behavior of Autonomous Vehicles

Instances where Teslas collided with stopped fire trucks reveal gaps in the system's understanding of obstacles on the road.

Many accidents involving self-driving cars occur because they stop unexpectedly, leading to rear-end collisions caused by human drivers.

Statistical Long Tail Problem Explained

The long tail distribution indicates that while some situations are common, many unlikely scenarios exist that can still pose risks for machine learning systems.

Addressing the long tail problem requires incorporating a form of common sense into machine learning models, something humans naturally possess.

Limitations in Machine Learning Understanding

Despite complex architectures and extensive training data, machines often fail to grasp what they have learned accurately.

A graduate student's experiment revealed that a neural network classified images based on background blur rather than identifying animals as intended.

Misinterpretation and Lack of Generalization

Machines may latch onto superficial features instead of meaningful characteristics; this leads to misclassification under altered conditions (e.g., photoshopped images).

When tested with slight variations (like paddle position), trained machines struggle significantly compared to humans who adapt easily.

Understanding Machine Learning Limitations and Adversarial Attacks

The Nature of Machine Learning Understanding

The way we describe machine learning processes, such as moving a paddle or hitting a ball, reflects our human concepts rather than the actual understanding of these systems. They may learn differently from us, leading to potential unreliability.

Adversarial Attacks on Neural Networks

In 2013, researchers at Google discovered adversarial attacks on deep neural networks trained on the ImageNet dataset. These networks could confidently identify objects but were vulnerable to manipulation.

Researchers demonstrated that by adding subtle image noise to an image (e.g., a school bus), they could trick the neural network into misclassifying it as something entirely different, like an ostrich.

This manipulation resulted in the neural network being convinced that distorted images were actually different objects, highlighting a significant gap between human perception and machine interpretation.

The term "adversarial example" was coined to describe this phenomenon where adversaries can exploit vulnerabilities in neural networks for malicious purposes.

Implications of Adversarial Examples

A paper titled "Intriguing Properties of Neural Networks" underscored the importance of understanding these vulnerabilities. It sparked a subfield focused on defending against adversarial examples.

Researchers created glasses with patterns designed to confuse facial recognition systems, demonstrating how easily machines can be fooled and raising concerns about their reliability in security contexts.

Vulnerabilities in Autonomous Systems

A notable experiment showed that stickers placed on stop signs could deceive self-driving cars into misinterpreting them as speed limit signs. This highlights critical safety issues regarding autonomous vehicles' reliance on visual data processing.

Philosophical Considerations of AI Understanding

Mathematician Giancarlo Rota questioned whether AI could ever truly grasp meaning beyond mere data processing. This raises fundamental concerns about the nature of understanding in machines compared to humans.

Defining terms like intelligence and understanding remains challenging; current AI lacks comprehension akin to human experience despite performing tasks effectively.

Pursuing Common Sense in AI

Efforts are underway to imbue machines with common sense—an everyday knowledge base that humans possess but machines lack. Paul Allen's AI institute focuses on bridging this gap for more reliable interactions between humans and machines.

For instance, while humans understand that a plastic bag poses no threat when encountered by a vehicle, machines currently lack this contextual awareness necessary for safe operation in real-world scenarios.

Understanding Common Sense in AI

The Challenge of Teaching Machines Common Sense

The speaker discusses the difficulty machines face in understanding context, using the example of birds crossing a road versus pieces of glass. Unlike humans, machines lack intuitive knowledge about such scenarios.

The Winograd Schema Challenge is introduced as a method to assess common sense reasoning in AI through pairs of sentences that require contextual understanding.

An example from the challenge illustrates how humans can easily determine what "full" or "empty" refers to based on experience, while machines struggle with this task.

Another example highlights how machines often fail to identify what shatters when given different contexts involving steel and glass, indicating their reliance on word associations rather than true comprehension.

Current AI performance is noted at around 60% accuracy for these challenges, which is not significantly better than random guessing (50%), emphasizing the gap between human and machine understanding.

Insights from Research and Development

Orinazioni from the Allen Institute for AI comments on the limitations of current AI systems regarding common sense, questioning their potential to take over complex tasks without basic contextual understanding.

DARPA's Machine Common Sense program aims to develop machines with an understanding equivalent to that of an 18-month-old child, highlighting a paradox where advanced AIs excel in games but lack fundamental common sense skills.

This disparity showcases that while machines can outperform humans in specific tasks like chess or Go, they struggle with simple everyday reasoning that comes naturally to humans.

Intuitive Knowledge Required for Understanding

The speaker emphasizes the need for intuitive knowledge—understanding physics, biology, and psychology—to navigate real-world situations effectively.

Examples illustrate how humans intuitively know cause-and-effect relationships (e.g., pulling a dog on a leash), which machines currently cannot grasp without explicit programming or learning.

Observations about social interactions are discussed; for instance, recognizing distractions (like someone being on their phone), which informs behavior in driving scenarios—a skill lacking in current AI models.

Concept Formation and Abstraction Challenges

The discussion shifts towards concept formation—how both humans and machines create abstractions and analogies.

Walking a dog serves as an example; while most people understand this concept easily through various representations, it poses challenges for machines trying to categorize similar activities accurately.

The complexity increases when considering variations like running instead of walking or unconventional scenarios (e.g., dogs riding skateboards), further complicating machine learning processes related to concepts.

Understanding Human-Like Intelligence in AI

The Nature of Concepts and Analogies

The speaker discusses the richness of human concepts, emphasizing our unique ability to create analogies and abstract ideas that build on existing knowledge.

Doug Hofstadter's assertion is highlighted: "without concepts, there can be no thought; without analogies, there can be no concepts," suggesting a foundational role for these elements in AI development.

The speaker identifies the formation and fluid use of concepts as a critical open problem in AI, indicating its importance for future advancements.

Current Challenges with Artificial Intelligence

A question arises about whether machines can achieve intelligence comparable to humans, with the speaker noting that current technology is far from this goal.

Concerns are raised regarding how smartphones and machines replace human functions like memory but also expose users to corporate interests behind data management.

The speaker reflects on the potential dangers of relinquishing control over our cognitive processes to machines that lack true understanding.

Trustworthiness and Control in AI Systems

The discussion shifts to the implications of trusting machines that do not possess genuine intelligence or understanding, raising questions about their reliability.

A quote from Pedro Domingo emphasizes that the issue lies not in machines becoming too intelligent but rather too unintelligent while gaining control over significant aspects of life.

Geopolitical Concerns Surrounding AI Development

A question is posed regarding national security concerns related to foreign investments in AI technologies, particularly by China.

The speaker acknowledges worries about China's use of technology for surveillance and population control, drawing parallels with similar trends occurring domestically.

Ethical Considerations in Technology Use

There’s an emphasis on the need for vigilance regarding how data is used by governments and corporations, highlighting ethical considerations surrounding privacy and decision-making.

Understanding the Role of the Right Brain in Creativity and AI

The Importance of the Right Brain

Historically, the right brain was undervalued due to its inability to label objects, but it plays a crucial role in understanding metaphor, analogy, and musicality essential for poetry.

The right brain is linked to embodied functions such as emotional processing and gut instincts. This raises questions about whether machines can replicate these functions without a physical body.

Rethinking Brain Functionality

Neuroscientists are reconsidering the strict left-brain/right-brain dichotomy, suggesting that our bodily interactions with the world significantly shape our understanding.

Current AI systems lack embodiment; they are isolated from real-world experiences which may hinder their ability to develop complex concepts akin to human intelligence.

Embodiment in AI Development

There is an emerging field focused on "embodied AI," where robots interact with their environment. This area is still developing but shows promise for enhancing machine learning.

The complexity of human cognition and emotion integration suggests that achieving true intelligence in machines may require more than just advanced algorithms.

Self-Driving Cars: Current Limitations and Future Predictions

Definitions of self-driving cars are evolving; full autonomy remains elusive due to numerous unpredictable driving situations.

Instead of fully autonomous vehicles, future developments may involve adapting urban infrastructure to assist self-driving technology.

Geofencing and Adaptation Strategies

The concept of geofencing will likely be implemented, allowing self-driving cars to operate only within designated areas equipped with necessary sensors and mapping.

AI in Medicine: Progress and Challenges

Recent studies indicate that AI is improving predictions related to heart attacks and disease susceptibility through enhanced analysis of medical data.

Data collected from vehicles encountering various situations can be shared across networks, enabling continuous improvement in driving algorithms.

Limitations of AI Learning

While cars can learn from experience, there’s uncertainty regarding their ability to handle unexpected scenarios effectively or achieve full autonomy soon.

Emotional Intelligence in Robotics

A question arises about whether robots could have physical correlates for emotions similar to humans (e.g., neurotransmitters like serotonin), reflecting on child development insights.

Understanding AI Development Through Child Psychology

The Intersection of Child Development and AI Models

The speaker highlights a 30-year-old book discussing how infants develop neural connections, particularly recognizing that their mother will return after leaving the room around eight months old.

Some researchers are exploring child development in AI, referencing a DARPA program aimed at creating an AI with the cognitive abilities of an 18-month-old baby by collaborating with developmental psychologists.

There is skepticism about the success of recreating human-like developmental trajectories in machines due to their lack of embodiment compared to babies.

Simulating Emotion in AI

While no one is attempting to replicate the complex chemical processes of human emotions in machines, there is ongoing research into simulating emotional responses within AI systems.

The discussion transitions to concerns about distinguishing between human interactions and those with AI, especially as both become more prevalent in various platforms.

Identifying Human vs. AI Interactions

A two-part question arises regarding how individuals can identify whether they are interacting with a person or an AI, especially given the rise of bots on social media platforms like Facebook.

The speaker notes that limited discussions may obscure whether one is speaking to an AI; however, broader conversations often reveal inconsistencies typical of current AI capabilities.

Ethical Considerations and Regulation

There are significant ethical questions surrounding how companies should regulate their interactions with users through AI systems, particularly concerning transparency about whether users are engaging with machines or humans.

Google’s demonstration of a conversational system for making restaurant reservations raised debates about requiring such systems to disclose their non-human nature during interactions.

Challenges Ahead: Deepfakes and Misinformation

As technology advances, it becomes increasingly difficult to differentiate between real and artificially generated content (deepfakes), raising concerns about trust and authenticity online.

The speaker expresses fear over the growing challenge of discerning genuine interactions from those generated by computers as both text and media become indistinguishable from reality.