Introduction to Generative AI

It is GenAI when y is ...

- natural language -image -audio

Generative Deep Learning Model

-generates new data that is similar to data it was trained on (Generates new content) -understands the distribution of data and how likely a given example is -predict new words in a sequence

Not GenAI when y is a ...

-number -discrete -class -probability

Machine Learning

-subfield of AI -trains model by input data, can make predictions drawn from training data -gives computers the ability to learn without explicit programming A type of artificial intelligence that leverages massive amounts of data so that computers can improve the accuracy of actions and predictions on their own without additional programming.

Discriminative Deep Learning Model

-used to classify or predict -typically trained on a dataset of labeled data -learns the relationship between the features of the data points and the labels

What is generative AI?

1. AI that creates new content based on what has been learned from existing content 2. the process of learning from existing content is called training and results in the creation of a statistical model 3. when given a prompt, Gen AI uses this statistical model to predict what an expected response might be and this generates new content -learns underlying structure of data and can create new examples

deep learning model types

1. discriminative 2. generative

what are some benefits of prompt tuning?

1. enables large language models (LLMs) to adapt to a wide range of tasks 2. helps LLMs generate more accurate responses

what are three examples of tasks that Bard code generation can perform?

1. explain your code to you line by line 2. debug your lines of source code 3. translate code from one language to another

Types of Generative AI Based on Data

1. input: image output: text, image, video 2. input: text output: text, image, audio, decisions

Google's approach to responsible AI is based on which of the following commitments?

1. it's built for everyone 2. it respects privacy 3. it's driven by scientific excellence 4. it's accountable and safe

Model Types

1. text-to-image 2. text-to-video, text-to-3D 3. text-to-task

supervised learning

A type of machine learning where algorithms are trained by providing explicit examples of results sought -learn from past examples to predict future values (already looking for something specific in the data)

Generative AI

AI technology that can create various types of new content based on what it has learned from existing content -the process of learning from existing content is called training and result results in the creation of a statistical model -when given a prompt, Gen AI -text -imagery -audio -synthetic data >a subset of deep learning

what was the process of traditional programming?

You had to hard code the rules for distinguishing something For example: In order to identify a cat the person had to describe each characteristics that could distinguish a cat from a dog or any other animal

downstream tasks

a dependent task that cannot run until an upstream task reaches a specified state

deep learning

a subset of machine learning that uses artificial neural networks , allowing them to process more complex patterns than traditional machine learning -generative AI and large language modes (LLMs) are a subset of deep learning

What is a prompt?

a text to query or instruct a large language model (LLM) to generate desired output

unsupervised learning

a type of machine learning where algorithms are looking at the raw (unlabeled) data, and seeing if it naturally falls into groups

What is recommended by Google as most important in establishing AI governance?

building practices around ethical decision-making

Unsupervised ML models

data comes with no tags (labels)

Supervised ML Models

has tags (label): names, numbers

neural networks

interconnected neural cells. With experience, networks can learn to perform tasks by processing data and making predictions -use "semi-supervised learning"

foundation model

large AI model pre-trained on lots of data designed to be adapted to downstream task

semi-supervised learning

neurons trained on small amounts of labeled data to help learn basic concepts of a task and large amounts of unlabeled data to help generalize to new examples

what action does Google recommend organizations take to ensure that AI is used responsibly?

seek participation from a diverse range of people

intelligent agents

sophisticated software programs that use collaborative filtering technologies to learn from past user behavior in order to recommend new purchases -systems that can reason and act autonomously

upstream tasks

task must reach a specified state before a dependent task can run

prompt design

the quality of the input determines the quality of the output

AI

theory and development of computer systems able to perform tasks normally requiring human intelligence -discipline -creation of intelligent agents: systems that can reason and act autonomously

In transformer, hallucinations are...

words or phrases that are generated by the model that are often nonsensical or grammatically incorrect, it can happen due to 1. The model not trained with enough data 2. noisy or dirty data 3. model is not given enough context 4. model not given enough constraints

y= f (x )

y= model output f= model x= input data

How did a wave of neural networks distinguish a cat from a dog?

you could show it a picture, and with that, they are able to predict which animal it was