In case you’re seeking to upskill in Generative AI (GenAI), there’s a Generative AI Learning Path in Google Cloud Skills Boost. It currently consists of 11 courses and provides a very good foundation on the idea behind Generative AI and what tools and services Google provides in GenAI. The very best part is that it’s completely free!

As I went through these courses myself, I took notes, as I learn best once I write things down. On this part 1 of the blog series, I would like to share my notes for courses 1 to six, in case you need to quickly read summaries of those courses.

I highly recommend you undergo the courses in full, each course is simply 10–20 minutes long they usually’re full of more information than I present here.

Let’s start.

In this primary course of the educational path, you study Generative AI, how it really works, different GenAI model types and various tools Google provides for GenAI.

enables computer systems to find a way to perform tasks normally requiring human intelligence.

is a subfield of AI and provides computers the flexibility to learn without explicitly programming.

ML has two flavors: and learning where the info is labeled within the latter.

is a subset of ML. Deep learning uses Artificial Neural Networks allowing them to capture complex patterns.

(GenAI) is a subset of Deep Learning. It creates latest content based on what it has learned from existing content. Learning from existing content known as training and ends in the creation of a statistical model. When given a prompt, GenAI uses a statistical model to predict the response.

(LLMs) are also a subset of Deep Learning.

Deep Learning Model Types are classified into two: and .

• is used to categorise (eg. a picture is predicted as dog)/
• generates latest data just like data it was trained on (eg. a dog image is generated from one other dog image). Generative language models reminiscent of, , study patterns in language through training data and given some text, they predict what comes next.

The ability of Generative AI comes from models. In Transformers, there’s an and .

are generated words/phrases which can be nonsensical or grammatically incorrect. Normally happens when the model shouldn’t be trained on enough data or trained on dirty data or possibly the model shouldn’t be given enough context or constraints.

A is a big AI model pre-trained on an enormous quantity of information, designed to be adapted and fine-tuned to a big selection of tasks reminiscent of sentiment evaluation, image captioning, object recognition.

offers for foundation models. For instance, and for Language and for Vision.

permits you to quickly explore and customize GenAI models on your applications. It helps developers create and deploy GenAI models by providing quite a lot of tools and resources. There’s a library of pre-trained models, a tool for fine-tuning models, and a tool for deploying models to production.

creates Gen AI Apps without writing any code with its visual editor, drag-and-drop interface.

permits you to test and experiment with Google’s LLMs. Provide developer access to models optimized for various use cases.

is an approachable solution to start prototyping and constructing generative AI applications. Iterate on prompts, augment your dataset, tune custom models.

On this course, you study Large Language Models (LLMs) and explain how they could be tuned for various use cases.

are a subset of .

are large, general purpose language models that could be after which for specific purposes.

Advantages of using LLMs:

1. A single model could be used for various tasks.
2. The fine-tuning process requires minimal field data.
3. The performance is repeatedly growing.

Pathways Language Model ( ) is an example of LLM and a with 540 billion parameters.

A consists of encoder and decoder. The encodes the input sequence and passes to the decoder. The learns to decode the representations for a relevant task.

LLM development vs. traditional ML development:

• In LLM development, when using pre-trained APIs, no ML expertise needed, no training examples, no must train a model. All it’s essential take into consideration .
• In traditional ML development, ML expertise needed, training examples needed, training a model needed with compute time and hardware.

There are 3 sorts of LLM:

• Generic language: Predict the subsequent word basen on the language within the training data.
• Instruction tuned: Trained to predict a response to the instructions given within the input.
• Dialog tuned: Trained to have a dialog by predicting the subsequent response.

provides task specific foundation models. e.g. Language sentiment evaluation, vision occupancy evaluation.

is the strategy of adapting a model to a latest domain (eg. legal or medical domain) by training the model on latest data.

is bringing your personal dataset and retrain the model by tuning every weight in LLM. This requires an enormous training job and hosting your personal fine-tuned model. Very expensive.

Parameter-efficient tuning methods ( ) is a technique for tuning LLM in your custom data without duplicating the model. The bottom model shouldn’t be altered, as an alternative small add-on layers are tuned.

On this course, you study Google’s AI principles:

1. AI ought to be socially helpful.
2. AI should avoid creating or reinforcing unfair bias.
3. AI ought to be built and tested for safety.
4. AI ought to be accountable to people.
5. AI should incorporate privacy design principles.
6. AI should uphold high standards of scientific excellence.
7. AI ought to be made available for uses that accord with these principles.

Moreover, Google is not going to design or deploy AI in 4 areas:

1. Technologies that cause or are more likely to cause overall harm.
2. Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people.
3. Technologies that gather or use information for surveillance that violates internationally accepted norms.
4. Technologies whose purpose contravenes widely accepted principles of international law and human rights.

That is only a placeholder course with a quiz to make sure that you complete the primary 3 courses.

The course is principally about diffusion models, a family of models which have recently shown tremendous promise within the image generation space.

Image Generation Model Families:

• Variational Autoencoders (VAEs): Encode images to a compressed size, then decode back to original size, while learning the distribution of the info.
• Generative Adversarial Models (GANs): Pit two neural networks against one another.
• Autoregressive Models: Generate images by treating a picture as a sequence of pixels.

: Recent trend of generative models.

• Unconditioned generation: Models don’t have any additional input or instruction. They’ll train on images of faces and generate latest faces from that or improve the resolution of images.
• Conditioned generation: These give things like text-to-image (Mona lisa with cat face), or image-inpainting (Remove the lady from the image) or text-guided image-to-image (disco dancer with colourful lights).

The essential idea of diffusion models is to systematically and slowly destroy structure in an information distribution through an after which that restores the structure in data, yielding a highly flexible and tractable generative model of the info.

On this course, you learn concerning the architecture that’s on the core of LLMs.

is a sequence-to-sequence architecture: it takes a sequence of words as input and outputs one other sequence.

Typically, it has 2 stages:

1. An encoder stage that produces a vector representation of the input sentence.
2. A decoder stage that creates the sequence output.

To coach a model, you wish a dataset, a group of input/output pairs. You’ll be able to then feed this dataset to the model, which corrects its weights during training based on the error it produces on a given input within the dataset.

A subtle point is that the decoder generates at each step the probability that every token in your vocabulary is the subsequent one. Using these probabilities, you select a word.

There are several approaches to choosing a word. The best one, the , is to generate the token that has the very best probability. A greater approach known as where the decoder evaluates the probability of sentence chunks, moderately than individual words.

In LLMs, the easy RNN network throughout the encoder-decoder is replaced by transformer blocks that are based on the eye mechanism.

There’s also a hands-on Encoder-Decoder Architecture: Lab Walkthrough that accompanies the course.