Overview:
This ACM Learning Webinar, part of ACM’s commitment to lifelong learning, presents a Town Hall with Peter Norvig on AI, machine learning, and related topics. Rosemary Paradis, Principal Research Engineer and Secretary-Treasurer of ACM SIG AI, moderates the discussion.

ACM Resources:
ACM offers educational and professional development resources to bolster skills and enhance career opportunities for its over 100,000 computing professionals and student members. Members can access ACM learning resources at atthelearning.acm.org, including educational and professional development opportunities.

Webinar Logistics:
Attendees experiencing technical difficulties can press F5 (Windows), Command R (Mac), refresh their browser (mobile), or close and relaunch the presentation. Additional widgets and resources are available on the bottom panel and right sidebar. Questions can be submitted through the Q&A box. The presentation is being recorded and will be archived. Attendees will receive an email notification when it’s available.

Speaker Introduction:
Peter Norvig, Director of Research at Google, is today’s speaker. He previously led Google’s Core Search Algorithms Group and NASA Ames Computational Sciences Division and was NASA’s Senior Computer Scientist. He received the NASA Exceptional Achievement Award in 2001.

Peter Norvig’s Path to AI:
In high school, Peter Norvig had access to a computer and learned basic programming. He also had a linguistics class that sparked his interest in natural language processing. He pursued both linguistics and computer science in college, leading him to AI.

Peter Norvig’s Current Interests:
Making AI and machine learning more accessible to all programmers. Lowering the barrier to entry for machine learning by creating tools that enable competent programmers to learn and apply machine learning without advanced degrees. Addressing the bottleneck of tasks waiting to be solved due to the limited number of experts in the field.

Textbook Recommendation:
Peter Norvig co-authored the leading textbook in AI, “Artificial Intelligence: A Modern Approach.” The new edition of the book is in progress and expected to be completed during Peter Norvig’s co-author’s sabbatical next year.

Advancements in AI:
AI has seen significant progress in recent years, particularly in benchmark tasks like speech recognition, image recognition, and machine translation. This progress is attributed to the availability of more data, increased computing power, and the development of new techniques such as deep learning.

Emphasis on Fundamentals:
Despite the excitement around new techniques, Peter Norvig emphasizes the importance of understanding and teaching the fundamentals of AI. He cautions against relying solely on a single method, as the best approach for a task can evolve over time.

AI Trends:
Expert AI scholar Peter Norvig predicts that the latest AI methods will not last forever, so it’s crucial to focus on fundamental aspects like representation of the world, handling uncertainty, and reasoning. Researchers should not discard past lessons but adapt to the trending topics of the time.

Communicating Impractical Ideas:
Senior researchers should provide younger researchers with practical milestones and metrics to measure progress. By tracking progress, it’s easier to determine if an idea is impractical or requires further development. Collaboration and agreement on the process are essential for successful implementation.

Societal Changes in AI:
Increased computer power and data availability have led to societal changes that favor AI-based approaches. People’s interests in activities such as reading, sharing pictures, and personalized recommendations drive the demand for AI solutions. AI is not just about following definitive instructions but also about optimization and uncertainty handling.

Essential Resources:
Exploring diverse sources and blogs of renowned AI experts like Andrew Ng, Sebastian Thrun, and Andrew Gelman for insights into AI. Following Chris Ola’s sophisticated animations and visualizations to understand AI concepts. Utilizing Mark Guzdiel’s resources for computer science education. Reading primary sources and preprints from conferences like NIPS to stay updated with the latest research.

Applying Deep Learning:
Understanding the domain, data, biases, and model requirements is crucial. Experimenting with various architectures, learning rates, and hyperparameters to develop intuition. Collaborating with others to gather a large dataset of successful and unsuccessful AI models to extract meta-knowledge.

Choosing a Programming Language:
Selecting a popular language like TensorFlow, with a supportive community, is recommended for beginners. Considering factors like support structure and personal preferences when choosing between different AI platforms.

AI and Wikipedia:
Utilizing structured data from Wikipedia’s attribute-value boxes through available APIs. Recognizing the limitations of structured data in capturing all the information present in article text.

Harnessing External Sources:
More extensive use of APIs will allow access to information and services from external sources, like the Google Knowledge Graph. These enhanced APIs will provide a wealth of knowledge, but also introduce the challenge of managing uncertainty and interpreting probability distributions.

Programming in an Uncertain World:
AI problems are becoming more prevalent in programming, requiring programmers to address uncertainty and deal with probabilistic data. Specialized programming languages, like probabilistic programming languages, may be useful for specific tasks, but traditional programmers will still play a significant role. Programmers will need training and adaptation to effectively handle probability distributions and make informed decisions based on uncertain outcomes.

Achieving Safety in AI Systems:
Provable 100% safety is not attainable in AI, non-AI, or human systems. The focus should be on understanding trade-offs, risks, and the level of risk in terms of percentages. Transparency and understanding of AI mistakes are crucial, as they differ from human errors. Developing a calculus of utilities is necessary to precisely define what we want AI systems to optimize and achieve.

Challenges in Autonomous Systems:
Asymptotically approaching 100% safety in autonomous systems poses unique challenges. Systems that work 50% of the time can be useful with a vigilant driver ready to take over. When systems reach 99% safety, over-reliance and over-confidence can lead to complacency and potential accidents. User interface design is crucial to prevent over-reliance and ensure drivers remain alert and ready to intervene when necessary.

The Role of Utility Functions in Encoding Ethics:
Utility functions serve as a representation of an individual’s ethical beliefs and preferences. These functions define what is considered desirable and fair, guiding the behavior of AI systems.

Challenges in Combining Utilities:
When dealing with a society of diverse individuals with different desires, combining individual utility functions to achieve a collective outcome presents difficulties.

Privacy Concerns:
Ethical considerations in AI development extend to the field of privacy, necessitating careful handling of personal data.

Fairness in AI Systems:
AI systems may inadvertently prioritize the majority population over minority groups when optimizing for overall performance. This can result in unequal attention and benefits for different subgroups, raising questions of fairness and equity.

Addressing Fairness:
To promote fairness, AI developers must modify optimization goals and consider subgroup-specific objectives. Ethical dilemmas arise in determining which subgroups should receive prioritized attention and protection.

Ethical Considerations in AI:
Peter Norvig emphasizes the importance of ensuring fairness in AI systems, particularly for protected subgroups such as different genders, races, and nationalities. He suggests that prioritizing fairness is an ethical question that requires careful consideration and the development of objective functions to maximize fairness.

Security Concerns in Self-Driving Cars:
In the context of self-driving cars, Norvig acknowledges the need to address security concerns and prevent hacking. He highlights the historical trade-offs between security and other desirable features, noting that consumers have often prioritized affordability and accessibility over security.

Prioritizing Security in the Internet of Things:
Norvig expresses concern about the current approach to the Internet of Things (IoT), which prioritizes affordability, ease of use, and a wide range of possibilities over safety and security. He suggests that it is time to reconsider these priorities and invest more in making IoT systems secure.

Challenges in Implementing Security at Scale:
Norvig recognizes the difficulty in implementing security measures at a wide scale, given consumer preferences for inexpensive and accessible technology. He acknowledges the need for breakthroughs that can provide both flexibility and security in IoT systems.

Open Research Questions in AI:
Norvig identifies several open research questions in AI that are in high demand for the industry today and in the future. One key area is small data, where AI systems need to learn and perform well even with limited data availability. Transfer learning, which involves transferring knowledge gained from one domain to another, is a promising approach in this context, but further research is needed.

Challenges in AI:
Difficulty in moving between levels of specificity and generality in reasoning. Combining symbolic AI with neural or sub-symbolic approaches. Logical reasoning over representations that lack definitive definitions.

Understanding the Brain in AI Research:
Importance of understanding the brain as a separate subject from the goal of AI. Benefits of applying neuroscience insights to problem-solving in AI. The Google Brain team’s focus on providing a programming tool for problem-solving.

Knowledge Axiomatic or Structural Approach:
The need for both learning and knowledge-based reasoning approaches in AI. Challenges with writing down knowledge manually: laboriousness and brittleness. The advantage of machine learning in handling exceptions and robustness.

Combining Machine Learning and Hard Reasoning:
Hybrid systems can combine machine learning and hard reasoning techniques to improve performance. In mathematical proof problems, neural nets can guess relevant axioms to make the search for proof faster.

Google’s Hybrid Research Approach:
Google’s research aims to solve applications and serve users, not just publish papers. Research is conducted with the goal of inventing new technologies that can be used to improve existing applications or create new ones.

Benefits of Hybrid Research:
Hybrid research can lead to faster development of new technologies. It ensures that research is focused on practical problems and has a clear purpose. It encourages collaboration between researchers and engineers to solve real-world problems.

Challenges in Natural Conversation:
Engaging in natural conversations with AI remains a significant challenge. The shift towards mobile-first and screenless devices emphasizes the need for conversational assistants. The goal is to create conversations that feel natural and achieve the user’s intended result, rather than simply providing relevant information. Striking a balance between making AI appear like a person and clarifying its limitations is essential.

Chatbots and the Turing Test:
Current chatbots can pass the Turing test in limited domains, mimicking human-like responses. However, these systems often struggle with complex conversations requiring context, reasoning, and multiple levels of action.

User Interface and Usability:
The user interface plays a crucial role in creating usable AI systems. Systems that clearly define their capabilities and limitations are more helpful and practical than those aiming for perfect human-like interaction. Inspiration can be drawn from automated teller machines (ATMs) that perform specific tasks reliably within their defined scope.

Progress in Online AI Education:
Since Peter Norvig’s TED talk on teaching AI to 100K students, there has been ongoing development of tools and materials for online AI education. The focus is on enabling effective teaching of AI to large groups of students. Improvements have also been made for regular classroom teaching of AI.

Tools for Course Creators and Learners:
Tools for course authors and teachers have been developed to simplify the process of creating online courses, such as video recording, editing, and course structure. Improvements have also been made to user interfaces, making online learning systems more interactive and user-friendly.

Nonlinear Learning Paths:
Linear learning paths, where students progress through a course in a fixed order, are being replaced by nonlinear branching networks. This allows for personalized learning experiences, where students can choose their own paths based on their knowledge and interests.

Recommendations and Knowledge Graphs:
Recommendation systems have been developed to suggest relevant content and activities to students based on their individual needs and progress. Knowledge graphs are used to map relationships between concepts and skills, enabling the system to create personalized learning pathways.

Community and Motivation:
The importance of motivation and community in online learning has been recognized. Online learning platforms now include features that foster connections between students and teachers, as well as peer-to-peer interactions. This enhances the learning experience and helps keep students engaged.

Early Stages of Development:
Despite the progress made, online education is still in its early stages of development. There is room for further improvements in personalization, assessment, and the integration of new technologies. Ongoing research and innovation will continue to shape the future of online learning.

Civic Education:
AI can be used to improve the education of the general population in various subjects, leading to a more informed and positive level of discourse.

Environmental Understanding:
AI can analyze satellite data and other sources to provide a better understanding of the physical world, including weather patterns and climate change.

Health:
AI is making progress in understanding diseases, drugs, and personalized medicine, offering significant opportunities for improving healthcare.