Artificial intelligence (AI) systems begin with data and evolve through various learning techniques to perform complex tasks that mimic human intelligence. From classifying images to generating human-like responses, AI models are built, trained, and refined using machine learning and its advanced subset, deep learning.
Key Insights
- Machine learning, a core component of AI, enables systems to learn from structured data using techniques like supervised, unsupervised, and reinforcement learning.
- Deep learning uses multi-layered neural networks to handle complex data such as images, speech, and language, found in tools like voice assistants and email spam filters.
- Generative AI builds on deep learning to produce original content by learning patterns from large datasets, enabling applications such as chat-based AI assistants that can write, summarize, and provide tailored responses.
This lesson is a preview from our AI Fundamentals for Government Employees Course. Enroll in a course for detailed lessons, live instructor support, and project-based training.
So someone has a problem that requires human-like intelligence. If it's really large or complex, like imagine the time it would have taken to classify all those thousands of cat or non-cat pictures, they may decide that an AI assistant would be helpful. So just like us humans, modern artificial intelligence creation starts with data.
To build an AI system, the relevant data for the problem is first gathered. That data is processed by algorithms, which are like step-by-step rules or mathematical instructions. Those algorithms are used to train neural networks, which are a term we use for a model composed of interconnected layers with learned parameters.
It's given that name because it's inspired by the way that the human brain works. So again, just like us humans, as the AI learns basic rules, those layers can be interconnected and start to look really smart and eventually be applied to new situations. To build an AI model, we need to choose an approach within machine learning based on tasks, data, and constraints.
So now let's talk about that machine learning. Machine learning is a subset of AI that enables machines to learn from data and improve over time without explicit programming. Machine learning is particularly good for structured data.
Machine learning may be supervised, which means it's learning from labeled data, like predicting house prices. It may be unsupervised, so that's like finding hidden patterns in unlabeled data, like customer segmentation, or through reinforcement learning, which is just basically trial and error with rewards and punishments. A common example of machine learning comes to us every day in our email inbox.
So like every time we get an email, that client, that email client is at work looking at our emails. It's checking the subject, it's checking the body, it's looking at the sender's address, it's looking at patterns of past spam to decide if this is legitimate or not, or maybe important or not, or promotion, and so on. Can you imagine the amount of time it would have taken the developers of that email client to code for every possible instance of spam or an important message? It's just not practical, right? So they employed AI, and over time, that filtering improves because it learns from the examples of you and other users.
So that's machine learning in action. It's using past data to make better predictions about the future. But what about more complex data? What if it doesn't come with a pretty structured bow on top? Deep learning is a subset of machine learning made for just such a situation.
It uses deep neural networks to handle complex tasks like image recognition, speech-to-text, and language understanding. So a common example of this that you probably see every day is, say, asking your phone, what's the weather like today? Well, that system is using deep learning to recognize your voice, that's speech to text, to understand the meaning of your words, that's natural language processing or NLP, and it's got to generate a response back to you, right? That sounds natural. It's got to turn that text to speech and make it sound organic.
So the interconnectivity of all those processes is powered by deep learning networks trained on huge amounts of data so that your AI assistant can keep improving at recognizing different accents, tones, and ways of asking the same question. So we've stated our problem, we've collected the necessary data, we've decided what type of process we want to use, and is it will have algorithms or machine learning, deep learning, those types of things? Well, using that chosen approach and data, the model is then trained by adjusting its internal parameters to reduce errors. The trained model is then evaluated using a validation or test set and then fine-tuned even more by adjusting its architecture to enhance its performance.
So that trained model is then deployed into a production environment where it can be used to make predictions and decisions. And that AI system keeps improving with more data, better processes, and subject matter experts and ordinary users like us making corrections. And this brings us to generative AI.
This type of AI, powered by deep learning, is even more specific in that it can take all the impressive components from the sets before it and create new content. It has learned from large data sets and produces original outputs that resemble what it learned. And that resembling what it learned is a key point that's going to come up later.
So even more impressive than our cat detector, generative AI could dream up a new animal altogether. So when you're using ChatGPT and systems like it, you're using generative AI. We'll dive deeper into GPTs later on in this course.
So generative AI systems of all names, like ChatGPT, are what I'll refer to in this course as AI assistants because that's really what they are. True to the definition of AI, generative AI assistants are so smart that they can help you with all kinds of things at home or at work. AI can answer questions for you.
You can tell it your problem, and it'll give you a solution. It'll help you with drafting, give you feedback, or rewrite. It's especially good at summarizing long content.
Just think of the possibilities if you were to really harness this generative AI.
