6

GenAI (Generative AI)

 1 year ago
source link: https://wilsonmar.github.io/genai/
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.
neoserver,ios ssh client

Generate new text, images, audio, and video rather than discrete numbers, classes, and probabilities.

This article is a work currently in progress.

This article introduces Generative AI (GenAI) on several cloud platforms:

AI is a discipline (of theories to make machines act like people, such as learning)

At microsoft

Microsoft has ownership interest in OpenAI, whose ChatGPT exploded in popularity in 2023.

  • “Azure OpenAI” became an offering March, 2023

Microsoft’s GitHub also unveiled its CoPilot series for developers on Visual Studio IDEs.

Many of Microsoft 365 SaaS offerings (Word, Excel, PowerPoint, etc.) have been upgraded with AI features.

Microsoft Bing Search

  • https://www.linkedin.com/learning/generative-ai-the-evolution-of-thoughtful-online-search by Ashley Kennedy (Managing Staff Instructor at LinkedIn Learning)

Search: Crawling, Indexing, ranking

https://lnkd.in/eCDjW4EW

ChatGPT made available to the public Nov 2022 reached 1 million users in less than a week.

Limitations:

  • Biased databases input
  • Point-in-time data (frozen in time)
  • Lack of common sense
  • Lack of creativity
  • No understanding of generated text
  • normalization of mediocrity

Tutorials

The “What is Generative AI” course at LinkedIn Learning by Dr. Pinar Seyhan Demirdag (Senior Data Scientist at Microsoft) is 1 hour and 15 minutes long and has 5 modules:

  • Introduction to Generative AI
  • Generative AI in action
  • Generative AI in the real world
  • Generative AI in the future
  • Next steps

The learning path for Generate artificial intelligence has 5 modules:

Microsoft has a Microsoft AI Fairness initiative.

  • https://www.linkedin.com/learning/what-is-generative-ai
  • https://www.linkedin.com/learning/generative-ai-imaging-what-creative-pros-need-to-know

  • https://www.linkedin.com/learning/generative-ai-the-evolution-of-thoughtful-online-search

  • https://www.linkedin.com/learning/generative-ai-for-business-leaders
  • https://www.linkedin.com/learning/ai-accountability-essential-training-16769302
  • https://www.linkedin.com/learning/responsible-ai-principles-and-practical-applications
  • https://www.linkedin.com/learning/foundations-of-responsible-ai
  • https://www.linkedin.com/learning/introduction-to-responsible-ai-algorithm-design
  • https://www.linkedin.com/learning/tech-on-the-go-ethics-in-ai

  • https://www.linkedin.com/learning/what-is-generative-ai/how-generative-ai-workspace by Pinar Seyhan Demirdag

  • https://www.linkedin.com/learning/streamlining-your-work-with-microsoft-bing-chat/understand-how-chat-ai-works by Jess Stratton (LinkedIn Learning Staff Author, Tech Consultant)

GitHub CoPilot

https://github.com/features/copilot

At Google

Dr. Gwendolyn Stripling, AI Technical Curriculum Developer at Google Cloud created courses at several sites:

A. https://www.coursera.org/learn/introduction-to-generative-ai/lecture/TJ28r/introduction-to-generative-ai

B. Google created a Generative AI learning path FREE 1-day courses with FREE quizzes (but one HANDS-ON lab in Vertex AI):

  1. Introduction to Generative AI
    • Hallucinations
    • Text-to-image using Stable Diffusion
  2. Introduction to Large Language Models - Google’s Bard AI https://bard.google.com/ is Google’s answer to OpenAI’s GPT series of large language models to generate images, audio, and videos.
    • https://www.techrepublic.com/article/google-bard-cheat-sheet/

  3. Introduction to Responsible AI 1 day

  4. Generative AI Fundamentals

  5. Introduction To Image Generation with diffusion models.

  6. Encoder-Decoder Architecture
  7. Attention Mechanism 2015 for Tensorflow
    • https://www.youtube.com/watch?v=fjJOgb-E41w to improve text translation by giving each hidden state a soft-maxed score
  8. Transformer Models & BERT Models 2017-18 for NLP
    • VIDEO: Overview added context to words
    • BERT (Bidirectional Encoder Representations from Transformers) developed by Google in 2018, trained using Wikimedia & Books in two variations: base (12 layers with 768 hidden units and 12 attention heads) and large (24 layers with 1024 hidden units and 16 attention heads).
    • 15% is what Google found to be the optimal balance in Masking (randomly replacing words with [MASK] tokens) and 85% Next Sentence Prediction (NSP) (predicting whether two sentences are adjacent or not).

    • BERT input embeddings: Token, Segment, Position, with [SEP]

    • Lab resource: classify_text_with_bert.ipynb from github
    • VIDEO: HANDS-ON walk-through of running “asl-gup.ipyr” notebook for Sentiment Analysis classifier_model.fit using Vertex AI Tensorflow Karas with GPU accessing the 25,000-record imdb database (trainable=true), optimized for binary accuracy. Run model saved from Google Cloud bucket uploaded to Vertex AI Registry. Deploy to endpoint (5-10 minutes). Test. Delete.
  9. Create Image Captioning Models with a CNN and RNN
  10. Introduction to Generative AI Studio for language, Vision, Speech. It has a “Model Garden”. Reflection Cards.

Other courses:

  • Generative AI with Vertex AI: Text Prompt Design for language, Vision, Speech. It has a “Model Garden”.

  • https://www.coursera.org/learn/introduction-to-large-language-models On Coursera: Google Cloud - Introduction to Large Language Models

GenAI Summary

Generative AI is abbreviated as GenAI.

Generative AI differs from other types of AI, such as “discriminative AI” and “predictive AI,” in that it doesn’t try to predict an outcome based on grouping/classification and regression.

  • Text classification, Translation among languages, Summarization, Question Answering, Grammar correction

Generative AI is a type of artificial intelligence (AI) that generate new text, images, audio, and video rather than discrete numbers, classes, and probabilities.

Output from GenAI include:

  • Text Generation
  • Image Generation (“Deep Fakes”), Image Editing
  • Video generation
  • Speech Generation: (Text to speech)
  • Decision Making: Recommandations, Play games
  • Explain code line by line
  • Code Generation

GenAI learns from existing data and then creates new content that is similar to the data it was trained on.

GenAI doesn’t require a large amount of labeled data to train on. Instead, it uses a technique called self-supervised learning, which allows it to learn from unlabeled data. This is a huge advantage because it means that generative AI can be used in a wide variety of applications, even when there isn’t a lot of data available.

A foundation model is a large AI model pre-trained on a vast quantity of data that was “designed to be adapted” (or fine-tuned) to a wide range of downstream tasks, such as sentiment analysis, image captioning, and object recognition.

Large Language Models (LLMs) are a subset of Deep Learning, a subset of Machine Learning, a subset of Artificial Intelligence. Machine Learning generate models containing algorithms based on data instead of being explicitly programmed by humans.

NLP (Natural Language Processing) vendors include:

  • Abnormal
  • Horizon3.ai
  • Darktrace identifies phishing emails using ML
  • SentinelOne

LLM creators:

  • OpenAI - now closed source
  • Google
  • NVIDIA

  • Meta (Facebook PyTorch) - open source

  • UC Berkeley
  • LMU Munich
  • Seyhan Lee | Artist | Generative AI Expert | AI Art Lab works on Hollywood movies

One way models are created from binary files (images, audio, and video) is “diffusion”, which draws inspiration from physics and thermodynamics. The process involves iteratively adding (Gaussian) noise for GAN (Generative Adversarial Networks) and VAE (Variational Autoencoders) algorithms to recognize until images look more realistic. This process is also called creating “Denoising Diffusion Probabilistic Models” (DDPM).

The models generated are “large” because they are the result of being trained on large amounts of data and also because they have a large number of parameters (weights) that are used for a wide variety of tasks, such as text classification, translation, summarization, question answering, grammar correction, and text generation.

The performance of large language models (LLMs) generally improves as more data and parameters are added.

  • The Pathways Language Model (PaLM) has 540 billion parameters, trained on Google’s 1.6 trillion parameter Switch Transformer model.
  • Facebook’s LaMDA has 1.2 trillion parameters.
  • OpenAI’s GPT-3 has 175 billion parameters.

Such large LLMs require a lot of compute power to train, so are expensive to create. Thus, LLMs currently are created only by large companies like Google, Facebook, and OpenAI.

LLMs are also called “General” Language Models because they can be used for a wide variety of tasks and contexts.

LLMs are also called “Transformer” Language Models because they use a type of neural network called a Transformer used for language translation, summarization, and question answering. Transformers are a type of neural network that uses “attention mechanisms” to learn the relationships between words in a sentence. They are called “Transformers” because they transform one sequence of words into another sequence of words rather than more traditional “Encoder-Decoder” models that focus on the “hidden state” between individual words.

Attention models use a RNN “self-attention” decoder mechanism that allows the model to learn the relationships between words in a sentence. VIDEO CS25: Encoder-decoders generate text using either “greedy search” or “beam search”. Greedy search always selects the word with the highest probability, whereas beam search considers multiple possible words and selects the one with the highest combined probability.

LLMs are also called “Autoregressive” Language Models because they generate text one word at a time, based on the previous word. They are called “Autoregressive” because they are a type of neural network that uses a type of neural network called a Transformer. Transformers are a type of neural network that uses attention mechanisms to learn the relationships between words in a sentence. They are called “Transformers” because they transform one sequence of words into another sequence of words.

It uses a neural network to learn from a large dataset.

After being developed, they only change when they are fed new data, called “fine-tuning” the model.

LLMs are also called “Universal” Language Models because they can be used for a wide variety of human written/spoken languages in prompts and outputs.

Prompt Engineering

A prompt is a short piece of text that is given to the large language model as input, and it can be used to control the output of the model in many ways.

Internally, when given a prompt (a request) GenAI uses its model to predict what an expected response might be, and thus generates new content.

OpenAI charges money to use GPT-4 with a longer prompt than GPT-3.5.

“Dialog-tuned” prompts are generate a response that is similar to a human response in a conversation with requests framed as questions to the chatbot in the context of a back-and-forth conversation.

Parameter-Efficient Tuning Methods (PETM) are methods for tuning an LLM on custom data, without duplicating the model. This is done by adding a small number of parameters to the model, which are then used to fine-tune the model on the custom data. This is done by adding a small number of parameters to the model, which are then used to fine-tune the model on the custom data.

Checklist for Prompt Engineering:

  • Details about content,
  • context (provide an example of answer),
  • use clear language
  • tone,
  • asthetic
  • role (imagine you’re the product manager for a brand-new smartphone company. What are ten potential innovative features that could be added within the next five years?)
  • analogies
  • debate-style questions (for and against)

References on prompt engineering:

Limitations

QUESTION: Detect emerging security vulnerabilities?

GenAI output is not based on human creativity, but rather on the data that it was trained on.

So GenAI is currently not built to do forecasting.

But many consider GenAI output as (currently) “creative” because GenAI can seem to generate content that is difficult to distinguishable from human-generated content, such as fake news, fake reviews, and fake social media posts.

Whatever biases were in inputs would be reflected in GenAI outputs.

Concerns

GenAI currently were not designed to be “sentient” in that it does not have a sense of self-awareness, consciousness, or emotions. More importantly, GenAI currently are not designed to have a sense of morality, in that it can generally recognize whether prompts and generated content is offensive, hateful, or harmful.

Developing responsible AI requires an understanding of the possible issues, limitations, or unintended consequences from AI use. Principles include Transparency, Fairness, accountability, scientific excellence. NOTE: “Explainability” is not a principle because it is not always possible to explain how an AI model works. “Inclusion” is not a principle because it is not always possible to include everyone in the development of AI models.

“ChatGPT 3.5 has all of the knowledge and confidence of a 14-year-old who has access to Google.” –Christopher Prewitt

“GPT-3 is a powerful tool, but it is not a mind reader. It is not a general intelligence system. It is not a self-aware being. It is not a robot. It is not a search engine. It is not a database. It is not a knowledge base. It is not a chatbot. It is not a question answering system. It is not a conversational AI. It is not a personal assistant. It is not a virtual assistant. It is not a personal knowledge base. It is not a personal knowledge guru.

Hallucinations (Making Stuff Up)

“Hallucinations” in output are made-up by the model and not based on reality. This can happen due to several causes: * input data is not representative of the real world * input data contains noisy or dirty data

* not trained on enough data
* not given enough context (in prompts)
* not given enough constraints

* prompt does not provide enough context

Their source of data (corpus) is kept confidential because that can be controversial due to licensing, privacy, and reliability issues.

  • Use of content from books and publications may have copyright concerns.
  • Use of content from websites would have licensing concerns even though they are publicly contributed
  • Use of Wikipedia (9 billion documents), Stack Overflow, Reddit, Quora, etc. have concerns about the usefulness that data

To ensure that AI is used responsibly, Google recommends “seeking participation from a diverse range of people”.

Google Bard code generation

  • explain code line by line
  • debug lines of source code
  • translate code from one language to another
  • generate documentation and tutorials about source code

Google AI Studio

Without writing any code:

  • Fine-tune models on custom data
  • Deploy models to production
  • Create chatbots using Google’s PaLM API for Chat
  • Image Generation (generate new images or generate new areas of an existing image)

GenAI Studio from PaLM API:

  • Fine-tune models
  • Deploy models to production
  • Create chatbots
  • Image generation
  • Write text prompts to generate

Google MakerSuite

Google’s MakerSuite is a suite of GUI tools for prototyping and building generative AI models by iterating on prompts, augment datasets with synthetic data, and deploy models to production, and monitor models in use.

  • Text to Image (generate new images or generate new areas of an existing image)

Generative AI App Builder

Generative AI App Builder creates apps for generating images.

GenAI API

Text to Image generation

midjourney (like Apple: a closed API, art-centric approach)

DALL-e (Open API released by a corporation - technical over design)

Stable Diffusion

  • https://github.com/CompVis/stable-diffusion uses Python Colab Notebooks
  • https://www.gofundme.com/c/help-changes-everything

Users: Stitchfix.com recommends styles.

https://prisma-ai.com/lensa

https://avatarmaker.com

Resources:

  • OpenAI
  • ChatGPT Discord server
  • Prompt Engineering Guide
  • PromptVine
  • Learn Prompting
  • PromptPapers
  • PromptHub

Video generation

https://www.unite.ai/synthesys-review/

Anomaly Detection

Variational Autoencoders (VAE)

Use cases: Find financial fraud, flaws in manufacturing, Network security breaches,

References

https://www.unite.ai/zero-to-advanced-prompt-engineering-with-langchain-in-python/

At AWS (coming soon)

Others must know: please click to share:

GenAI (Generative AI) was published on August 08, 2023.


report

Recommend

About Joyk


Aggregate valuable and interesting links.
Joyk means Joy of geeK