Training: Introductie tot Generatieve AI

Generative artificial intelligence (AI) focuses on creating models that can generate content such as text, images, or even multimedia. Unlike discriminative models that classify or label existing data, generative models operate by learning patterns from the provided data and producing novel outputs. You'll begin this course with an overview of generative. You will explore some notable examples of generative models, including OpenAI's ChatGPT and Google Bard. Next, you will look at the use of prompt engineering when interacting with AI chatbots. Then, you will then delve into the history and evolution of generative AI models including important milestones that culminated in the conversational agents that we work with today. Finally, you will explore the risks and ethical considerations associated with generative AI, such as unintentional use of copyrighted data, the use of personal data for training, and the creation of malicious deepfakes using AI. You will also learn how you can mitigate some of these risks while working with generative technologies.

Autoencoders are a class of artificial neural networks employed in unsupervised learning tasks, primarily focused on data compression and feature learning. Begin this course off by exploring autoencoders, learning about the functions of the encoder and the decoder in the model. Next, you will learn how to create and train an autoencoder, using the Google Colab environment. Then you will use PyTorch to create the neural networks for the autoencoder, and you will train the model to reconstruct high-dimensional, grayscale images. You will also use convolutional autoencoders to work with multichannel color images. Finally, you will make use of the denoising autoencoder, a type of model that takes in a corrupted image with Gaussian noise, and attempts to reconstruct the original clean image, thus learning better representations of the input data. In conclusion, this course will provide you with a solid understanding of basic autoencoders and their use cases.

Variational autoencoders (VAEs) represent a powerful variant of traditional autoencoders, designed to address the challenge of generating new and diverse samples from the learned latent space. VAEs introduce probabilistic components, incorporating a probabilistic encoder that maps input data to a distribution in the latent space and a decoder that reconstructs data from samples drawn from this distribution. Begin this course by discovering how variational autoencoders can be used for generating images. Next, you will create and train VAEs in Python and the Google Colab environment. Then you will construct the encoder and decoder. Finally, you will train the VAE on multichannel color images. Upon course completion, you will have a solid understanding of variational autoencoders and their use in generating images.

Generative adversarial networks (GANs) represent a revolutionary approach to generative modeling within the realm of artificial intelligence. Begin this course by discovering GANs, including the basic architecture of a GAN, which involves two neural networks competing in a zero-sum game - the generator and the discriminator. Next, you will explore how to construct and train a GAN using the PyTorch framework to create and train the models. You will define the generator and discriminator separately, and then kick off the model training. Finally, you will focus the deep convolutional GAN, which uses deep convolutional neural networks (CNNs) rather than regular neural networks. CNNs are optimized for working with grid-like data, such as images and these can generate better-quality images than GANs built using dense neural networks. In conclusion, this course will provide you with a strong understanding of generative adversarial networks, their architecture, and their usage scenarios.

The OpenAI Playground is a dynamic and user-friendly platform that allows individuals to engage with OpenAI's cutting-edge language models, such as GPT-3.5 and GPT-4. The Playground enables users to experiment with natural language processing (NLP) capabilities and parameters to tweak the responses of models. You will start this course by exploring the fundamentals of OpenAI models. Next, you will log into the OpenAI Playground and input basic prompts, observing the responses. You will work with multiple application programming interfaces (APIs), including the recommended chat completions API and the legacy completions API, all of which are accessible via the playground. Finally, you will work with the Assistants API which has access to tools for data retrieval, code interpretation, and function calling and can leverage these to respond to user queries. You will utilize the code interpreter to read and visualize CSV data, generating Python code for charts using libraries like Matplotlib and Seaborn.

OpenAI application programming interfaces (APIs) represent a groundbreaking leap in the accessibility of state-of-the-art natural language processing (NLP) capabilities. These APIs provide developers with a powerful toolset to integrate advanced language models seamlessly into their applications, products, and services. You will start this course by engaging with OpenAI through the command-line, utilizing the OpenAI APIs. You will learn how to authenticate yourself using API keys when programmatically accessing API endpoints using cURL commands. You will explore how to configure context for past interactions with the model and access both chat completions and legacy completions APIs via their respective endpoints. Moving onto Python, you will install the OpenAI library to create a client object for endpoint access. You will configure the API key and send requests to the chat completions endpoint with prompts in the JSON format. You will also explore the legacy completions API using the same client object. You will be introduced to the diverse range of model offerings from OpenAI and learn how to use those models. Finally, you will configure model parameters to adjust the response from the model. You will learn about the seed parameter to receive deterministic responses and how the system fingerprint helps track infrastructure changes on the server. You will explore various parameters, including Top P and Temperature for controlling creativity, max length, and stop sequences for response length, and frequency and presence penalty for word and topic repetition.

DALL-E and Whisper are OpenAI's image and audio-based model offerings. DALL-E, an image generation model, demonstrates the ability to create visually striking images based on textual prompts. Whisper represents a state-of-the-art automatic speech recognition (ASR) system. With its high accuracy in transcribing spoken words, Whisper finds utility in various applications, from voice assistants to transcription services. You will begin this course by generating images using OpenAI's DALL-E model. You will generate images using text prompts, create variations of existing images, and perform image inpainting using natural language. Then, you will work with the Whisper model, which caters to speech transcription and translation. You will transcribe and translate audio in different languages and accents, and you will evaluate the performance of these models.

Fine-tuning models is a critical aspect of leveraging pre-trained artificial intelligence models to suit specific tasks or domains. OpenAI allows developers to fine-tune models like GPT-3 and 4, enabling customization for particular applications. You will begin this course by creating prompt-completion pairs for fine-tuning, running a fine-tuning job, and observing the model's performance. You will send prompts based on the training data and examine the model's attempt to answer questions. Next, you will dive into connecting with the Assistants API programmatically. You will create an assistant by providing a role description and model, and you will initiate a thread to simulate user-assistant conversations. You will also upload files and query the assistant based on information contained in the files. Finally, you will explore creating and comparing text embeddings, efficient numerical representations of text that capture meaning and semantics of the text. You will learn how embeddings of similar words are numerically close to one another and how embeddings can be used as a preprocessing technique to represent text for other machine learning applications such as clustering and classification.

Final Exam: Generative AI Introduction and Overview will test your knowledge and application of the topics presented throughout the Generative AI Introduction and Overview journey.