Skip to content

module8-notes.md

Latest commit

 

History

History
373 lines (280 loc) · 25.1 KB

module8-notes.md

File metadata and controls

373 lines (280 loc) · 25.1 KB

Module 8: Emerging Topics in AI

Introduction to Emerging Topics in AI

8.1.1: Introduction

  • The course's final module focuses on cutting-edge AI research areas and technologies.
  • It provides a high-level overview of key concepts in AI rather than technical details.
  • AI is evolving rapidly, impacting various fields and industries.
  • Examples of AI applications include generative systems in art, AI-generated music and entertainment content, medical Q&A, legal AI bots, industrial monitoring systems, and more.
  • Pre-trained transformers significantly drive innovation in AI, including computer vision and large language models.
  • The module covers areas of active research in AI, such as one-shot algorithms, speed improvement, mobile deployment, and more.
  • The rapid evolution of AI technology raises both potential benefits and challenges.
  • Ethical and regulatory concerns about AI in the legal industry are discussed.

8.1.2: Getting Started

  • The module focuses on fundamental concepts related to emerging AI topics.
  • No new software installation or usage is required for this module.

AI the Creator

8.2.1: Introduction

  • Previous modules covered transformers and their applications in generative AI.
  • Transformers are potent models that process data, learn from it, and make predictions or generate content using self-attention.
  • Generative AI utilizes transformers to interpret text instructions and create various output types.
  • Transformer technology can generate text, images, music, or code.

8.2.2: Interactive Text Generation

  • ChatGPT is a sophisticated natural language processing model based on GPT-4.
  • It can suggest income tax deductions and performs well on various tasks.
  • ChatGPT is trained using reinforcement learning with three steps: supervised fine-tuning, mimicking human preferences, and Proximal Policy Optimization (PPO).
  • This reinforcement learning strategy improves the naturalness of ChatGPT's responses.
  • Google's Bard focuses on recognizing inaccurate information and may admit limitations.
  • Microsoft's Copilot is integrated into its Office suite for various text-related tasks.

8.2.3: Image Generation
Generative AI for Images:

  • Image-generating AI systems like Stable Diffusion use deep learning techniques.
  • They map textual descriptions to images or parts of images.
  • They interpret user prompts describing desired images and create matching images.

Role of Noise in Image Generation:

  • Noise refers to random pixels that disturb images.
  • Image generation models recognize and predict how to clear up noisy images.
  • The training process involves adding noise to an image and predicting the original image from the noisy version.
  • Image generation in reverse: a noisy image is generated based on text descriptors, and the model clarifies it.

Efficiency in Image Generation:

  • Image generation is computationally demanding but can be optimized.
  • Stable Diffusion compresses images to smaller sizes for diffusion and reverse diffusion processes.
  • This optimization reduces the computational resources required.

Other Image Generators:

  • DALL-E 2 by OpenAI generates images based on text prompts using contrastive learning.
  • Artbreeder allows users to manipulate images with sliders to create original artwork.
  • DeepArt.io recreates user images in the styles of art movements like Pop Art or Impressionism.

8.2.4: Music Generation
AI in Music Generation:

  • Music generators use complex transformers to associate prompts with sound data qualities.
  • Transformers can analyze existing music data and categorize it based on labels from user prompts.
  • This process identifies styles, genres, keys, transitions, tempos, and other musical elements.
  • Producers can describe their desired song to the algorithm to create a customized piece.

Example: Aimi:

  • Music-generating model Aimi can create a new sound file that producers can modify without composing or recording from scratch.

8.2.5: Problems and Possibilities
Issues with Generative AI Training Data:

  • Generative AI models require large datasets for training.
  • Lack of transparency regarding the data used for training.
  • Concerns about using copyrighted works for training without proper compensation.

Example: Stable Diffusion and Getty Images:

  • Stable Diffusion models were accused of generating images based on copyrighted work, leading to a lawsuit involving Getty Images.

Challenges in AI Training:

  • Lawsuits related to AI models trained on public data repositories like GitHub's Copilot.
  • Creators of publicly available works do not anticipate their intellectual property being used by AI.

Opportunities in Content Creation:

  • Generative models offer opportunities for speeding up content production.
  • Potential for innovative and experimental content creation.
  • Chaining together multiple AI components for complex tasks without additional user direction.

8.2.6: References
Abdullah, U. 2023. How does Stable Diffusion work? Available: https://www.pcguide.com/apps/how-does-stable-diffusion-work/ [2023, April 25].

Baio, A. 2022. *Exploring 12 million of the 2.3 billion images used to train Stable Diffusion’s image generator. Available: https://waxy.org/2022/08/exploring-12-million-of-the-images-used-to-train-stable-diffusions-image-generator/ [2023, April 25].

Black, D. AI music generator apps - DJ’s worst nightmare or the ultimate tool? Available: https://cybernews.com/editorial/ai-music-generator-apps/ [2023, April 25].

Bullas, J. 2023. The ultimate list of AI image generator tools - create powerful visuals for your next digital marketing campaign. Available: https://www.jeffbullas.com/ai-image-generator/ [2023, April 25].

Chan, K. 2023. What can ChatGPT maker’s new AI model GPT-4 do? Available: https://abcnews.go.com/Business/wireStory/chatgpt-makers-new-ai-model-gpt-4-97881871 [2023, April 25].

Isaacs-Thomas, B. 2023. How AI turns text into images. Available: https://www.pbs.org/newshour/science/how-ai-makes-images-based-on-a-few-words [2023, April 25].

Metz, C. 2023. *What Google Bard can do (and what it can’t). The New York Times. Available: https://www.nytimes.com/2023/03/21/technology/google-bard-guide-test.html [2023, April 4].

OpenAI. 2022. Introducing ChatGPT. Available: https://openai.com/blog/chatgpt [2023, April 25].

Ramponi, M. 2022. How ChatGPT actually works. Available: https://www.assemblyai.com/blog/how-chatgpt-actually-works/ [2023, April 25].

Reuters, T. 2023. Microsoft will add AI tech to office suite programs like Word, Excel. Available: https://www.cbc.ca/news/business/microsoft-ai-copilot-office-suit-1.6780630 [2023, April 25].

Vanian, J. 2023. Microsoft adds OpenAI technology to Word and Excel. Available: https://www.cnbc.com/2023/03/16/microsoft-to-improve-office-365-with-chatgpt-like-generative-ai-tech-.html [2023, April 25].

Vincent, J. 2022. The lawsuit that could rewrite the rules of AI copyright. Available: https://www.theverge.com/2022/11/8/23446821/microsoft-openai-github-copilot-class-action-lawsuit-ai-copyright-violation-training-data [2023, April 25].

Vincent, J. 2023. Getty Images sues AI art generator Stable Diffusion in the US for copyright infringement. Available: https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-lawsuit-getty-images-stable-diffusion [2023, April 25].

Xiang, C. 2023. Developers are connecting multiple AI agents to make more ‘autonomous’ AI. Available: https://www.vice.com/en/article/epvdme/developers-are-connecting-multiple-ai-agents-to-make-more-autonomous-ai [2023, April 25].

AI Outside the Computer

8.3.1: Introduction
Expanding Applications of AI Beyond Computers:

  • AI technologies moving beyond computers into real-world devices and robots.
  • Aim to solve problems, improve safety, overcome human limitations, and enhance convenience.

Areas of Focus in AI Expansion:

  • Autonomous vehicles, robots, and devices interacting with humans and the environment.
  • Key AI models and research areas: computer vision, generative adversarial networks, and tinyML.

8.3.2: Autonomous Vehicles
Autonomous Vehicles and AI:

  • Autonomous vehicles, or self-driving cars, are an AI implementation gaining public attention.
  • Potential applications beyond personal transportation in industries like public transport and shipping.
  • Self-driving vehicle development relies heavily on data processing for training and decision-making.

Challenges in Self-Driving Vehicle Development:

  • Complex challenges include other vehicles, objects, adverse weather, and more.
  • Requires real-time perception and reaction to the environment.
  • Reliance on computer vision technology.

Computer Vision in Self-Driving Vehicles:

  • Computer vision allows machines to "see" and analyze images.
  • Requires training on diverse data to recognize objects from different angles and perspectives.
  • Real-time data processing, pixel analysis, and labeling for decision-making.
  • Innovations in 3D modeling and image analysis improving the process.

Development Progress and Challenges:

  • Progress in autonomous driving technology, but challenges remain.
  • Predictions for fully autonomous driving pushed further into the future.
  • Companies like Tesla, Mercedes, BMW, and Ford working on automated driving features.

8.3.3: Robots
Applications of Computer Vision in Robotics:

  • Robotics is an advanced field that commonly uses AI technologies.
  • Advancements in AI fields like autonomous vehicles can be applied to robots.
  • Robots, like autonomous vehicles, rely on computer vision for interaction with the physical world.

Role of Large Language Models:

  • Large language models assist robots in interpreting human language into instructions.
  • They help robots plan and execute tasks and provide verbal feedback.

Google's PaLM-E Visual-Language Model:

  • In March 2023, Google demonstrated a robot with the PaLM-E visual-language model.
  • The robot successfully interpreted instructions to fetch snacks and clean a spill.

Specialized Robotics Applications:

  • Japan has developed specialized robots for elder-care facilities.
  • These robots assist with physically demanding tasks, facilitate recreational activities, and provide companionship.
  • These projects show the potential of AI technologies in diverse applications.

8.3.4: The Internet of Things
Internet of Things (IoT) Overview:

  • IoT involves a network of connected devices equipped with sensors, controllers, processors, and software for data collection and processing.
  • Connected devices exchange data and perform networked tasks.
  • Examples of IoT applications include smart fridges, home automation, and data gathering for companies.
  • IoT devices have been in use since the early to mid-2010s, and AI is now a standard design element in IoT devices.

TinyML in IoT:

  • TinyML refers to machine learning applications on low-power devices, handling more minor data and algorithms.
  • Applications include voice recognition, predictive maintenance, and more.
  • It allows AI functionality on resource-constrained IoT devices.

Generative Adversarial Networks (GANs) in IoT:

  • GANs refine computer vision in IoT devices.
  • GANs consist of generative and discriminator models in competition.
  • GANs use actual data from IoT devices to train models, improving functionality.
  • They are valuable for generating realistic failure data in contexts with data scarcity, aiding in predictive maintenance and failure prediction.

8.3.5: Mobile Deployment

  • AI is already in use in daily life, including smartphones.
  • AI powers features like autocomplete, camera improvements, and face recognition for security.
  • Machine learning in products allows for personalized user experiences.
  • Algorithms keep users engaged on social media platforms.
  • Mobile developers aim for seamless user experiences and task automation.
  • The growth of AI in mobile technology offers more personalization and data for algorithm improvement.
  • Increased customization can enhance security through sensitive verification systems.

8.3.6: Problems and Possibilities

  • AI automation raises questions of accountability.
  • Responsibility for AI decisions is unclear.
  • Self-driving cars pose challenges in determining responsibility for accidents.
  • Bias in AI decision-making due to biased data.
  • Debate on human oversight in AI decision-making.
  • Concerns about economic disruption due to job displacement by robots.
  • Potential for AI to create new opportunities and industries.
  • There is a need to reorganize work and society to address AI's impact.
  • Opportunities for AI in TinyML and assistive technologies.
  • Consider ethical implications and inclusion of people with disabilities in AI design.
  • The importance of thinking about emerging technologies' broader impacts and implications.

8.3.7: References
Arun. 2020. An introduction to TinyML. Available: https://towardsdatascience.com/an-introduction-to-tinyml-4617f314aa79 [2023, April 26].

awg. 2023. Auto-GPT: Open-sourced disaster? [LessWrong forum post, 26 April] Available: https://www.lesswrong.com/posts/s9JWqgnv7xT2mxmE7/auto-gpt-open-sourced-disaster [2023, May 2].

Benanav, A. 2020. Automation and the future of work. New York: Verso Books. ISBN: 1839761296

Bloomberg. 2022. Tesla makes automated driving available to all owners in N. America. Available: https://europe.autonews.com/automakers/tesla-makes-automated-driving-available-all-owners-n-america [2023, April 25].

Collins, E. 2023. A comprehensive checklist for IoT project success. Forbes. Available: https://www.forbes.com/sites/forbesbusinesscouncil/2023/03/16/a-comprehensive-checklist-for-iot-project-success/ [2023, April 26].

Clearbridge. n.d. ML & AI mobile apps: how to integrate the tech. Available: https://clearbridgemobile.com/ml-ai-mobile-apps/ [2023, May 2].

De Freitas, J. 2023. Will we blame self-driving cars? Wall Street Journal. Available: https://www.wsj.com/articles/will-we-blame-self-driving-cars-11674745636 [2023, April 3].

Doyle, N. 2022. Artificial intelligence is dangerous for disabled people at work: 4 takeaways for developers and buyers. Forbes. Available: https://www.forbes.com/sites/drnancydoyle/2022/10/11/artificial-intelligence-is-dangerous-for-disabled-people-at-work-4-takeaways-for-developers-and-buyers/ [2023, May 2].

Edwards, B. 2023. Google’s PaLM-E is a generalist robot brain that takes commands. Available: https://arstechnica.com/information-technology/2023/03/embodied-ai-googles-palm-e-allows-robot-control-with-natural-commands/ [2023, April 26].

Galloway, S. 2023. Luddites and AI. Available: https://medium.com/@profgalloway/luddites-and-ai-d4a491f2aa88 [2023, May 26].

IBM. n.d. What is computer vision? Available: https://www.ibm.com/topics/computer-vision [2023, April 25].

Jaokar, A. 2018. AI technologies used in Robotics. Available: https://www.datasciencecentral.com/ai-technologies-used-in-robotics/ [2023, April 26th].

Knight, W. 2022. A new trick lets artificial intelligence see in 3D. Available: https://www.wired.com/story/new-way-ai-see-3d/ [2023, April 6].

Martinez, C. 2021. Artificial intelligence and accessibility: examples of a technology that serves people with disabilities. Available: https://www.inclusivecitymaker.com/artificial-intelligence-accessibility-examples-technology-serves-people-disabilities/ [2023, May 2].

Pazzanese, C. 2020. Ethical concerns mount as AI takes bigger decision-making role in more industries. Available: https://news.harvard.edu/gazette/story/2020/10/ethical-concerns-mount-as-ai-takes-bigger-decision-making-role/ [2023, May 2].

Proulx, S. n.d. How AI needs to be redesigned for people With disabilities Available: https://makeitfable.com/article/ai-and-analytics-people-with-disabilities/ [2023, May 2].

Smith, Z.S. 2022. Self-driving car users shouldn’t be held responsible for crashes, U.K. report says. Forbes. Available: https://www.forbes.com/sites/zacharysmith/2022/01/25/self-driving-car-users-shouldnt-be-held-responsible-for-crashes-uk-report-says/ [2023, May 2].

Team ACV. 2022. Consumers are geeking out over these best self-driving cars of 2022. [Blog, 31 August]. Available: https://www.acvauctions.com/blog/best-self-driving-cars-2022 [2023, April 25].

Thomson, J. 2022 Whose ethics should be programmed into the robots of tomorrow?. Available: https://bigthink.com/thinking/robots-ethical-decisions/ [2023, May 2].

Wright, J, 2023. Inside Japan’s long experiment in automating elder care. Available: https://www.technologyreview.com/2023/01/09/1065135/japan-automating-eldercare-robots/ [2023, April 26].

Additional Areas of Active Research

8.4.1: Introduction

  • Active research areas in AI beyond transformers.
  • One-shot learning as a strategy for image recognition.
  • 3D environments created using AI for various applications.
  • Research on improving algorithm speed and computational resources.
  • Ethical and regulatory considerations in AI implementation.

8.4.2: One-Shot Learning

  • Understanding the challenge of training a model to identify novel images it has never seen before.
  • Introduction to one-shot learning as a solution for comparing two novel images.
  • One-shot learning involves using a Siamese neural network (SNN) to measure feature similarity between images.
  • Training the SNN with a triplet loss function that includes an anchor, positive, and negative image.
  • One-shot learning applications include facial recognition, signature authentication, computer vision, and neural activity identification.
  • Mention zero-shot and few-shot learning techniques for situations with limited or no labeled data.
  • Advantages and limitations of one-shot learning compared to traditional machine learning algorithms.

8.4.3: Creating 3D Environments

  • Traditional 3D environment creation is a manual and time-consuming process.
  • Importance of accurately simulated 3D environments for entertainment and training purposes.
  • Introduction of Neuro Radiance Field (NeRF) by Nvidia as an innovation in 3D modeling.
  • NeRF's ability to predict and generate complex 3D environments from a few images.
  • The neural network in NeRF processes images to create 3D points, predict light passage, and populate virtual spaces.
  • Ongoing improvements and variations of NeRF, including RegNeRF, pixelNeRF, Mega-NeRF, and LOLNeRF.
  • Applications of NeRF in video game development for more immersive experiences and in training self-driving vehicles and robots.
  • Ethical considerations of using NeRF to create realistic 3D training environments without endangering human lives.

8.4.4: Algorithm Speed and Computational Resource Management

  • AI systems are computationally resource-intensive, requiring significant processing power.
  • Cloud-computing data centers consume substantial energy, equivalent to small industrialized nations.
  • Environmental impact due to power consumption, primarily when relying on fossil fuels.
  • Example of BLOOM model emitting a large amount of carbon.
  • Improving AI algorithms for faster performance and more efficient use of power.
  • AI's potential to provide solutions to power consumption and emissions problems.
  • Use AI reinforcement learning to optimize chip design and reduce energy consumption.
  • Introduction of reinforcement learning models like BCOOLER to optimize energy usage.
  • Uncertainty about whether AI optimizations will entirely offset data center costs and reduce emissions, requiring further research and testing.

8.4.5: Ethics and Regulations
AI Ethics Issues:

  • AI systems can potentially harm people before issues are detected; some issues may never be detected.
  • AI ethics intersects philosophy, technology, and ethics.
  • Researchers analyze and audit AI systems to identify harmful impacts and propose technical solutions.
  • Ethics research takes place in academic research institutions and tech companies.

AI Ethics in Academic Institutions:

  • Academic institutions involved in AI ethics research include Harvard University's Berkman Klein Center, the University of Oxford's Centre for the Governance of AI, the Data & Society Research Institute, the Alan Turing Institute, the Ada Lovelace Institute, and the Centre for Human-Compatible Artificial Intelligence.
  • Difficulty accessing technical details of commercial AI products limits researchers' ability to audit and propose solutions.

AI Ethics in Tech Companies:

  • Some tech companies, including Accenture Labs, IBM, Google, Samsung, and Sony, have committed to ethical AI principles.
  • Corporate research into AI ethics may affect technical research progress and profits.

Government Regulations:

  • Governments recognize the need to regulate AI to prevent harm.
  • Regulations should be based on AI ethics research.
  • Limited AI regulations exist globally, with the EU's GDPR being an example.
  • Ongoing research and collaboration are essential for proactive AI regulation aligned with ethical principles.

8.4.6: Problems and Possibilities
AI's Potential and Concerns:

  • AI can bring efficiency and solve resource issues.
  • Concerns include contributions to climate change due to computational resource demands.
  • Many data centers are located in developing nations, posing challenges.

Positive Developments:

  • Companies are motivated to reduce overhead and develop resource-efficient technologies.
  • Accountability and ethical incentives can drive positive social impacts.

Ultimate Goal of AI Research:

  • The ultimate goal is artificial general intelligence (AGI).
  • AGI development may take years, but various forms of AI are already part of daily life.

Rapidly Evolving AI:

  • AI is evolving rapidly with opportunities and challenges.
  • The choices made in AI development and implementation have broad consequences.

Ethical Responsibility:

  • Approaching AI development and implementation with ethical responsibility is crucial.
  • The potential for progress and innovation is vast, but risks and implications are significant.

8.4.7: References
aiethicist.org. n.d. AI organizations. Available: https://www.aiethicist.org/ai-organizations [2023, May 2].

Allyn, B. 2020. Ousted Black Google researcher: 'they wanted to have my presence, but not me exactly'. Available: https://www.npr.org/2020/12/17/947719354/ousted-black-google-researcher-they-wanted-to-have-my-presence-but-not-me-exactl [2023, May 2].

Datagen. n.d. Neural Radiance Field (NeRF): A Gentle Introduction. Available: https://datagen.tech/guides/synthetic-data/neural-radiance-field-nerf/ [2023, May 2].

De Laat, P.B. 2021. *Companies committed to responsible AI: from principles towards implementation and regulation? Philosophy and Technology. 34. pp. 1135-1193.

Dickson, B. 2020. What is one-shot learning?. Available: https://bdtechtalks.com/2020/08/12/what-is-one-shot-learning/ [2023, May 2].

Gillam, W. 2021. Accelerating AI computing to the speed of light. Available: https://www.ece.uw.edu/spotlight/ai-computing/ [2023, May 2].

Jones, E. & Easterday, B. 2022. Artificial intelligence’s environmental costs and promise. Available: https://www.cfr.org/blog/artificial-intelligences-environmental-costs-and-promise [2023, May 2].

Knight, W. 2022. A new trick lets artificial intelligence see in 3D. Available: https://www.wired.com/story/new-way-ai-see-3d/ [2023, April 6].

Logunova, I. 2022. A guide to one-shot learning. Available: https://serokell.io/blog/nn-and-one-shot-learning [2023, May 2].

Luccioni, A.S., Viguier, S., & Ligozat, A.L. 2022. Estimating the carbon footprint of BLOOM, a 176B parameter language model. 10.48550/arXiv.2211.02001.

Martinovs, D. 2019. 10 organizations leading the way in ethical AI. [Blog, 22 August]. Available: https://ocean.sagepub.com/blog/10-organizations-leading-the-way-in-ethical-ai [2023, May 2].

Maslej, N. 2023. 2023 AI index: a year of technical achievement, newfound public scrutiny. Available: https://hai.stanford.edu/news/2023-ai-index-year-technical-achievement-newfound-public-scrutiny [2023, May 2].

Rahtz, M., Varma, V., Kumar, R., Kenton, Z., Legg, S., and Leike, J. 2022. Avoiding unsafe states in 3D environments using human feedback. Available: https://deepmindsafetyresearch.medium.com/avoiding-unsafe-states-in-3d-environments-using-human-feedback-5869ed9fb94c [2023, May 2].

Rouse, M. 2023. Zero-shot, one-shot, few-shot learning. Available: https://www.techopedia.com/definition/34949/zero-shot-one-shot-few-shot-learning [2023, May 2].

Salian, I. 2022. NVIDIA research turns 2D photos into 3D scenes in the blink of an AI. [Blog, 25 March]. Available: https://blogs.nvidia.com/blog/2022/03/25/instant-nerf-research-3d-ai/ [2023, May 2].

Spyridaki, K. n.d. GDPR and AI: friends, foes or something in between? Available: https://www.sas.com/en_us/insights/articles/data-management/gdpr-and-ai--friends--foes-or-something-in-between-.html [2023, May 2].

Vincent, J. 2023. AI is entering an era of corporate control Available: https://www.theverge.com/23667752/ai-progress-2023-report-stanford-corporate-control [2023, May 2].

Course Wrap-Up

8.5.1: Congratulations!

  • Congratulations on completing the course!
  • Reflect on the accomplishments in your educational and professional journey.
  • Key learnings from the course:
    • Understanding AI and its impact on our lives.
    • Applying machine learning techniques.
    • Using unsupervised and supervised machine learning.
    • Evaluating and optimizing machine learning models.
    • Neural networks and deep learning.
    • NLP and transformers for text-related tasks.
    • Recent innovations in AI.
    • Future applications of AI.
  • Please consider applying your knowledge in your current or future career.
  • Encouragement to continue learning in AI.