ABC TO XYZ OF AI

Building an AI Tool in 5 Minutes: A Quick Demo

Welcome to the AI Engineer Bootcamp, where we will build the foundations of AI and gradually advance our knowledge to eventually create personal, exciting AI-powered projects. Developing an AI tool might seem complicated for newcomers, but with just a few lines of code, we can achieve incredible results entirely from scratch. This is precisely what you will learn to do in this course.

Building a Chatbot in Five Minutes

In this practical lesson, we will build a chatbot that answers questions about the “Introduction to AI” module, the first module you will encounter in this bootcamp. The goal is ambitious: to implement the chatbot’s backend and user interface in only five minutes. We will employ the popular LangChain Python library to implement the backbone of the application and the Streamlit Python library to create an elegant user interface with minimal code. Let’s see if creating an app in five minutes is achievable.

Backend Implementation: Loading and Splitting Documents

Our demonstration begins by implementing the backend of our Q&A chatbot in Jupyter Notebook. We have already prepared the transcript of the “Introduction to AI” module as a PDF, so our first task is to load it. LangChain provides tools to make this straightforward, returning the file as a list of documents. Next, we split these documents further, ensuring each contains no more than 200 tokens, corresponding to roughly 150 words. This helps feed the chatbot the shortest text possible during the Q&A stage, optimizing the application’s cost.

Creating Embeddings and Vector Database

Next, we transform each document into an array of numbers, also known as vectors or embeddings. This transformation allows our chatbot to quickly find the text most relevant to the question asked. We initialize an embedding function by OpenAI to perform this transformation. Using this alongside the list of documents, we create a Chroma vector database that keeps all vectors in a local folder named “intro to AI.”

Retriever Initialization and Testing

Having the vectors at hand, we initialize a retriever. The retriever’s task is to find the document most relevant to the question asked. For example, if we invoke it with the question “What did Alan Turing do?” the retriever is expected to fetch the document most closely related to that topic. Printing the response confirms that the retriever performs flawlessly.

Writing Prompt Templates

Next comes writing the prompts, which is critical in the chatbot creation process because prompts instruct the model about its intended behavior and purpose. Notice the keywords with the curly brackets; they represent input text placeholders, making the prompts reusable. A reusable prompt is called a prompt template. Our templates expect the user’s question and the retriever’s context.

Initializing the Language Model

We initialize the language model by opting for OpenAI’s GPT-4.0 and setting the temperature to zero, ensuring similar answers each run. We also need a string output parser, which guarantees the result will be output as a string.

Implementing the LangChain

The most important LangChain construction is the chain itself. Its purpose is to link components by passing the output of one as an argument to the next. We store the question and context in a dictionary, feed the dictionary to the prompt template, pass the resulting prompt to the language model, and feed the model’s response to the parser, ensuring the final output is a string we can print.

Testing the Chatbot with Streaming Response

Let’s test the chatbot by asking the familiar question “What did Alan Turing do?” To stream its response, we apply the stream function and use a for loop to display each text chunk continuously. The chatbot streams the response successfully, using context drawn from the lecture rather than its general knowledge. The response is concise and formatted differently from a typical ChatGPT answer.

Creating a User-Friendly Interface with Streamlit

Although the backend is impressive, we need to wrap it up in a nice package and give it a user-friendly interface. We introduce Streamlit for this purpose. The source code starts with essential LangChain and Streamlit imports. We instantiate OpenAI chat and embedding models, initialize a string output parser, a Chroma vector store, and a retriever using the local vector store created earlier. We include the system prompt instructing the chatbot about its purpose and the prompt template expecting the question and context. Finally, we include our chain.

Streamlit Application Setup

The Streamlit-specific part begins by setting an intuitive title for our application, “365 Q&A Chatbot.” We add a divider to separate the title from the rest of the elements and create a text input field expecting the user’s question. We then create a button labeled “Ask” responsible for triggering the streaming of the response. If the button is clicked without a question filled in, we display a warning message with an appropriate icon to ensure the user inputs a question.

Generating and Displaying the Response

Assuming all text fields have been filled in, we create a text placeholder and an empty string to store the text. We apply the stream method on the chain object inside a for loop, appending each chunk to the string and using a suitable Streamlit method to display the string generated so far. This completes the implementation.

Final Testing and Reflection

Although we did not quite nail the five-minute mark, the implementation was excellent. After saving the file and refreshing the browser page, asking the chatbot “What did Alan Turing do?” works flawlessly. This project exemplifies what you can build after completing the course. Even if some implementation felt advanced, careful, step-by-step guidance will help you quickly acquire the necessary skills.

Course Overview

The bootcamp will start by exploring AI’s history, key concepts, and common buzzwords. Next, it covers Python programming to ensure you have the coding skills needed. Then, it delves into natural language processing and large language models, covering essentials like text processing, classification, vectorization, the attention mechanism, and transformers. Following that, the course teaches the LangChain Python library for creating AI-powered applications. Later modules include vector databases with semantic searches using Pinecone, speech recognition basics and practical machine learning applications, and applied AI engineering, guiding you through planning and building an AI-powered tool such as an interview simulator for a data scientist position using OpenAI and Streamlit.

Conclusion

This is the game plan for our bootcamp. Let’s jump straight into the world of AI-powered applications with a module you are well familiar with: Introduction to AI. Enjoy!

Key Takeaways

• Building an AI tool can be achieved quickly and efficiently using Python libraries like LangChain and Streamlit.
• Transforming documents into embeddings enables efficient retrieval of relevant information for chatbot responses.
• Prompt templates are essential for instructing language models on their intended behavior and ensuring consistent outputs.
• Combining backend AI logic with a user-friendly interface creates practical and interactive AI applications.

Natural vs Artificial Intelligence

Introduction to Natural Intelligence

Here is someone driving on the highway. Another is solving complex math problems. A man is crafting poetry, and a girl is mastering music. These are just a few examples of natural intelligence demonstrating our inherent capacity to acquire and learn new skills.

We do not come into this world with all these abilities. Learning them takes time. Fortunately, nature has equipped us with the right tools. Our brains are designed to observe and process vast amounts of information. This is how we learn and become more intelligent than yesterday.

The Oxford Dictionary defines intelligence as the ability to acquire and apply knowledge and skills. We often do not appreciate how sophisticated our brains are. The technological innovations created by human intelligence today look like magic compared to the tools available just a few centuries ago.

One of our greatest strengths as humans is that we are tool builders. Throughout history, we have repeatedly found ways to boost our productivity. One such way involved the creation of machines.

At first, such machines looked like this. This is Gutenberg’s printing press, which was invented in 1440. If we had to rank the most incredible machines of all time, the printing press would likely claim the number one spot because it revolutionized how we share knowledge.

But for all its brilliance, we cannot say that this is an intelligent machine, right? This device operated under fixed parameters and could not acquire and apply knowledge and skills.

It took centuries before the notion of intelligent machines came into existence. Computer scientists, inspired by natural intelligence and human brain processes, envisioned imparting such intelligence to machines. The new field of study aimed at achieving this goal is known as artificial intelligence or AI.

Key Takeaways

• Natural intelligence is our inherent capacity to acquire and learn new skills over time.
• Human brains are designed to observe and process vast amounts of information, enabling learning.
• Intelligence is defined as the ability to acquire and apply knowledge and skills.
• Artificial intelligence emerged from the desire to create machines capable of acquiring and applying knowledge, inspired by natural intelligence.

Demystifying AI, Data Science, Machine Learning, and Deep Learning

In this course, we will frequently discuss AI, machine learning, and data science. Sometimes, these terms might be used interchangeably, which can cause confusion. Let’s take a few minutes to clarify their definitions.

Understanding Artificial Intelligence

Artificial Intelligence (AI) aims to make machines intelligent, allowing them to learn and acquire new skills. AI is a broad discipline that encompasses several subfields.

Machine Learning: A Key Subfield of AI

Machine learning, one of AI’s key subfields, harnesses data to predict outcomes by identifying intricate dependencies beyond mere correlations. Essentially, much like a mathematical function, we feed input data to a model, which processes the data in a certain way and produces an output.

Consider this practical example:

• An algorithm can analyze a user’s movie watching history to predict which movies they will likely enjoy next.
• Similarly, a customer’s financial transactions can be used to generate a credit score that forecasts their loan repayment likelihood.

We will explore various types of machine learning models later in the course, but for now, remember this definition and that machine learning is a significant branch of AI.

The Role of Data Science

How does data science fit into the picture? AI and machine learning are essential parts of data science. Every data scientist worth their salt can work with machine learning algorithms.

However, data scientists often use other statistical methods that are more traditional, such as data visualization and statistical inference. Often, they aim to extract insights from data without relying solely on machine learning.

Therefore, data science has an essential intersection with AI and its subfield machine learning, but it also relies on mathematics, statistics, and data visualization to gain business value from data.

A data scientist can develop a sophisticated algorithm to predict future client orders, but they can also plot clients’ orders against store visits and uncover valuable business insights.

Key Takeaways

• Artificial Intelligence (AI) aims to make machines intelligent by enabling them to learn and acquire new skills.
• Machine Learning (ML) is a key subfield of AI that uses data to predict outcomes by identifying complex dependencies beyond simple correlations.
• Data Science encompasses AI and ML but also includes traditional statistical methods such as data visualization and statistical inference to extract insights from data.
• Data scientists use a combination of ML algorithms and traditional techniques to derive business value from data.

Weak vs Strong AI

In our last lesson, we discussed a machine learning algorithm that predicts movie recommendations based on your viewing history.

We described this as a form of artificial intelligence, but we can agree that this is limited intelligence. The machine can perform a specific task that it has been designed to do.

We call this type of artificial intelligence narrow AI.

This type of AI is already integrated into our daily lives and has proven valuable for businesses.

When OpenAI released the ChatGPT and ChatGPT 3.5 models in 2022, we saw one of the first successful attempts at creating semi-strong AI.

Everyone who has tried ChatGPT knows that it can handle a wide range of tasks. It can even generate responses often indistinguishable from those a human might provide, which is what Alan Turing envisioned with his imitation game.

We’ve reached a point where an AI can successfully pass the Turing test.

Today, ChatGPT can write jokes, proofread your text, recommend the best course of action, read and create pictures, create Excel and Python formulas, and solve math tasks.

So at the very least, we can call this a semi-strong AI.

OpenAI, Google, and other top AI research institutions have yet to achieve artificial general intelligence (AGI).

Sam Altman, OpenAI’s CEO, defines AGI, also referred to as strong AI, as the most powerful technology humanity has ever created.

AGI will generally be more intelligent and capable than humans in several tasks.

Many believe that for an AI to be considered AGI, it should be able to create science on its own. This means using all collectively available knowledge to make new discoveries and ideas.

How close are we to achieving strong AI? Perhaps pretty close.

Sam Altman states AGI could be developed in the reasonable, close-ish future.

As we edge closer to this frontier, we must consider the immense potential and ethical implications of creating machines that can think, learn, and even create science better than we do.

Key Takeaways

• Narrow AI performs specific tasks it is designed for and is already integrated into daily life.
• Semi-strong AI, exemplified by ChatGPT, can handle a wide range of tasks and pass the Turing test.
• Artificial General Intelligence (AGI), or strong AI, aims to surpass human capabilities in multiple tasks and create new scientific knowledge.
• The development of AGI is approaching, raising significant potential and ethical considerations.

Structured vs Unstructured Data

Introduction to Data Types

In this part of the course, we will discuss data, the main ingredient needed to create an AI model. There are two main types of data: structured and unstructured.

Structured Data

As the name suggests, structured data is organized into rows and columns. For example, if I want to see how many sales transactions there have been this month, I can create a well-organized Excel spreadsheet with rows and columns and input the data, which then takes the form of structured data.

Unstructured Data

Unstructured data, such as text files, images, videos, and audio files, lacks a defined structure and cannot be organized into rows and columns. In fact, 80 to 90 percent of all the world’s data is unstructured. This means that most data does not have this simple row-column predefined field structure.

Value of Unstructured Data

In the past, structured data was considered more valuable because it was easier to analyze. However, AI advancements today have found ways to turn unstructured data into valuable insights, with companies like Meta and Google excelling in this area.

Business Opportunities

Analyzing unstructured data is opening up enormous opportunities for businesses. They now have billions of photographs, video footage, text messages, and emails. For the first time, they can use this information to gain unprecedented insights.

Looking Ahead

But how have AI scientists made sense of so much unstructured data? We will discuss that in our next lesson.

Key Takeaways

• Data is the main ingredient needed to create an AI model.
• There are two main types of data: structured and unstructured.
• Structured data is organized into rows and columns, such as in spreadsheets.
• Unstructured data includes text files, images, videos, and audio files, making up 80 to 90 percent of all data.
• Advances in AI have enabled the extraction of valuable insights from unstructured data, opening new opportunities for businesses.

How We Collect Data

Introduction to the mNIST Dataset

The mNIST database is the hello world example for students who want to learn machine learning. The dataset includes 70,028 by 28 pixel grayscale images of handwritten digits. This exercise aims to train a machine learning algorithm to recognize the numbers we have in an image, even though they are never the same because people have different handwriting.

If you write the number three and I write the number three, we’ll have similar but slightly different versions. So how does the computer distinguish between images of three, six, or nine? It boils down to the fundamentals of computer science and electrical engineering.

All information can be broken down into zeros and ones. Here’s how this happens. The handwritten number three can be represented in the following way. Each pixel in these images has a value between 0 and 255 that describes its shade of gray, where zero is white and 255 is black.

In computers, all information, including these pixel values, is stored in binary form using combinations of zeros and ones. When we consider the pixel values in mNIST or any digital image, these numbers are converted into a binary format so computers can store and process them.

So we see different pictures of the number three, but they are pictures with similar number sequences for the computer. To solve the mNIST task via machine learning, the computer is trained to identify and differentiate these numerical sequences, enabling it to recognize digits from 0 to 9.

This example is fundamental because it shows how we can turn different types of information from the world around us into data that computers can read. Similarly, videos are thousands of pictures composed of hundreds or even thousands of pixels. Each pixel in turn corresponds to a number represented in binary form with only zeros and ones.

Even sound and written speech can be represented as sequences of zeros and ones. This is how computers receive the necessary structure to learn. They use this information to find patterns, study similarities, and ultimately learn how to execute different types of tasks.

We do not always appreciate how remarkable human brains are. We can see the world around us and collect a lot of information simultaneously. Our minds process countless details, subconsciously using our senses to absorb what we see, hear, and taste. This complex system allows us to understand and respond to the world, helping us to interact meaningfully with everything around us.

AI researchers aspire to provide machines with such capabilities utilizing tools that collect data through sensors, video, audio, texts, social media, satellite images, and internet browsing patterns. They gather information by scraping the web, accessing APIs, and harnessing big data analytics. This wealth of data enables them to train algorithms, refine machine learning models, and enhance artificial intelligence systems to better understand and interact with the world.

Data scientists often say, especially when building AI models, “Garbage in, garbage out.” High quality data input ensures excellent model output.

Key Takeaways

• The mNIST dataset consists of 70,028 grayscale images of handwritten digits, each 28 by 28 pixels.
• Computers represent images as binary data by converting pixel values ranging from 0 to 255 into zeros and ones.
• Machine learning algorithms learn to recognize digits by identifying patterns in these binary representations.
• High-quality data input is essential for producing accurate and reliable AI model outputs.

Labelled and Unlabelled Data

Welcome back to our AI modeling lesson. Two primary methods exist for collecting and preparing data for AI modeling: labeled or unlabeled data.

Imagine you have a dataset of 10,000 photos featuring various animals, including dogs. Before any modeling begins, someone meticulously reviews each photo, classifying them as dog or not a dog. This process creates a labeled dataset.

Of course, this concept can be applied not only to pictures but also to text, audio, video, and other data types.

Now envision a collection of YouTube comments. We can classify comments as positive, negative, or neutral. To train a model that categorizes new platform comments and flags harmful ones, manually labeling is undoubtedly time-consuming and expensive.

So why go through all the effort? The advantage is that in the end, we will be able to train an AI model with high-quality data that can significantly improve its accuracy.

Models typically trained on labeled data are more reliable and perform well in real-world applications. However, a clear trade-off exists between the time and resources spent labeling data versus model performance.

The other option is to work with unlabeled data. Today’s machine learning advancements allow us to work with unstructured data such as images, video, text, and unlabeled audio.

This means feeding the dataset of 10,000 photos to the model without going through every picture. In this case, we leave it to the AI model to learn on its own.

We will explain the learning process in more detail later. For now, you must be able to tell the difference between labeled and unlabeled datasets.

Thank you for watching.

Key Takeaways

• Labeled data involves manual classification, enhancing model accuracy but requiring significant time and resources.
• Labeled datasets can be applied across various data types including images, text, audio, and video.
• Unlabeled data allows models to learn from unstructured inputs without manual labeling.
• There is a trade-off between the effort spent on labeling data and the resulting model performance.

Metadata: Data that Describes Data

The main reason for the advancement of AI in recent years is the rapid digitalization of our lives. The growth of online shops, mobile phones, cameras, social media, sensors, and the Internet of Things devices has led to an exponential increase in data generation.

Today, we enjoy access to not only more data, but also significantly higher quality data. Consider the sharp contrast between photographs you took with your older phone several years ago and those taken with your smartphone today. This surge in high quality data has significantly fueled the development of AI.

But how do people recognize and make sense of the ever growing ocean of digital data? Much of this data is unstructured and too voluminous to label. Metadata proves invaluable here. Every data set should include it.

Metadata is data that describes other data. It summarizes key details like asset type, author, creation date, usage, file size, and more.

Here’s what the file’s metadata for this lesson looks like in our example with the dogs data set. Each photo includes metadata detailing its name, capture date, photographer, file size, and more.

Whether you handle structured or unstructured, labeled or unlabeled data, it’s essential to include metadata, especially in large data sets and complex systems.

Key Takeaways

• The rapid digitalization of our lives has driven the advancement of AI in recent years.
• The exponential increase in data generation is due to online shops, mobile phones, cameras, social media, sensors, and Internet of Things devices.
• Access to higher quality data, such as improved smartphone photographs, has significantly fueled AI development.
• Metadata, which is data that describes other data, is essential for managing large, unstructured, and unlabeled data sets.
• Including metadata with every data set is crucial, especially in complex systems and large data collections.

Machine Learning

Introduction to Machine Learning

In this section of the course, we will begin discussing the different types of machine learning. Before we start, I would like to ask you a favor. Please visit the course dashboard and click on “leave a rating.” This will mean a lot to me and help other students recognize that this course is worth taking. Since this is a large course, many people do not complete all lessons in one sitting and eventually forget to leave a rating. So, please do it now. It should only take a couple of seconds and will truly validate our efforts. Thank you very much.

Recap of AI Basics and Introduction to Machine Learning

So far, we have explored AI basics, its history, and the distinctions between weak and strong AI. You have already learned about the key ingredient to building AI: high-quality, abundant data. We are making excellent progress. Next, we will explore machine learning, the AI subfield pivotal to its recent success.

The Concept of Machine Learning

The idea of machine learning is fascinating because it boils down to designing a system capable of learning and improving through a trial and error process. Consider building a machine learning model similar to a student taught by a teacher. In this analogy, the machine learning model is the student, and the data scientist is the teacher. The teacher’s role is to help the student learn how to solve problems. To do this effectively, the teacher must provide the student with many suitable learning materials, which in machine learning means providing a lot of good data.

With sufficient information, the student carefully studies and absorbs the training materials. Similarly, the machine learning algorithm learns to recognize patterns in the provided training data. The aim is to prepare students to succeed on their final exam, which features unfamiliar questions that test previously covered material. In the machine learning model’s case, we aim to teach it how to solve a particular problem with data it has never seen before.

The more up-to-date and better the training data, the better the model can solve new challenging problems it has never encountered. Without adequate training, a student may underperform compared to a well-taught one, regardless of natural talent. The same applies to machine learning. A simple model with lots of good data might perform better than a more complicated one with less data.

Adopting new teaching methods and tailoring styles to students can significantly enhance their performance. Similarly, in machine learning, models improve prediction and analysis as they encounter new data types and updated algorithms. This continuous learning and adaptation process is crucial to developing robust AI systems that can handle diverse challenges and evolve.

Just as some students excel in problem solving across various fields, certain machine learning models are better suited to specific ones. The data scientist needs to understand this and be able to use the appropriate machine learning model in a given situation. I hope this somewhat abstract example was helpful.

Practical Example: Machine Learning in Real Estate

Let’s consider a practical example of how machine learning can help businesses. A real estate agent plans to develop a mobile app to provide prospective clients with an estimated selling price for their homes. Users can receive this estimate by answering targeted questions within the app. This could significantly boost the real estate agent’s business by enabling them to collect contact details from people interested in selling their property.

To put the plan in motion, the real estate agent contacted a data scientist and asked whether the project was feasible. The data scientist’s first question was about the availability of a sufficiently large list of past transactions. Fortunately, the real estate agent’s company had access to a database with thousands of past transactions over the last few years. Later, the data scientist was pleased to find the data well organized, detailing house prices, sizes, room counts, bedroom numbers, distances from the city center, neighborhoods, and other relevant aspects.

“I can build a machine learning model with that,” said the data scientist. Here’s how the data scientist simplified the model for his client. You are familiar with the classic function notation, where yy is a function of XX, right? Imagine we want to develop a machine learning model to predict house prices in our city by analyzing past real estate transactions in your training data set.

We have plenty of past transactions where both XX and yy are known. We need to train a machine learning algorithm with historical data to discover patterns and learn the best way to predict the future price of a home, yy, based on the data from past transactions and the home’s known characteristics. Clients enter their home’s characteristics into the app. This forms input xx. The machine learning model then generates output yy, predicting the home’s price based on similar historical data.

This is where we will end the story. Spoiler alert: the app was a huge success for the real estate agent’s business. I hope this lesson clarified what machine learning is and how it works. In our next lesson, we will discuss the different types of machine learning models.

Thank you for watching.

Key Takeaways

• Machine learning enables systems to learn and improve through trial and error, similar to a student learning from a teacher.
• High-quality, abundant training data is crucial for effective machine learning model performance.
• Machine learning models can predict outcomes, such as house prices, by learning patterns from historical data.
• Continuous learning and adaptation improve machine learning models’ ability to handle new and diverse challenges.

Supervised, Unsupervised, and Reinforcement Learning

In this video, we will describe three main types of machine learning: supervised, unsupervised, and reinforcement learning.

Then, in the next video, we will explore deep learning, a sophisticated subset of machine learning powered by neural networks and inspired by the human brain.

If this sounds too abstract and complicated, please do not worry. We will discuss the intuition behind these AI techniques without getting into too much detail.

Supervised Learning

Supervised learning excels when we work with labeled data. In the earlier example, the training data set reveals whether a picture contains a dog. This is how the algorithm can learn to classify new pictures as dog or not dog.

We have provided the machine learning model with an extensive training data set containing labeled data, so it knows what to look for. Later, when we provide a new picture, it can identify whether it contains a dog based on the feedback given during training. This is a classification problem solved by supervised machine learning.

Another primary application of supervised learning is predicting an outcome. Remember when we discussed the real estate agent’s mobile app in our last lesson? The data set with past house sales transactions contained both known prices and features of the homes.

We did not merely highlight home features. This is an example of supervised learning used for prediction, also called a regression problem. When a user inputs the characteristics of a new home into the app, the app predicts the new home’s price based on its features and historical information for similar homes.

Simply put, in supervised learning, we explicitly train the model with known outputs to guide its learning.

Unsupervised Learning

In unsupervised learning, we process data without labels. For example, consider a data set of 10,000 animal pictures. Half feature dogs and the other half various animals, but we provide the machine learning model with no labels or indications.

The model will need to scan through all pictures and look for patterns. Ultimately, it will be able to distinguish specific groups and individuate them. We will have a group of dogs, a group of cats, and so on.

However, the machine learning model does not specify the contents of these images. It merely indicates that they share similar characteristics.

Grouping them is unsupervised learning and is helpful in many situations. First, labeling data is costly and not always practical. Moreover, sometimes we do not know what relationships we should look for in the data, so it makes sense to have the algorithm find these interesting relationships first.

In this way, a supermarket chain could find clusters of target customers with different behaviors, or a real estate agent could use an unsupervised model to determine which types of properties are sold most often.

Reinforcement Learning

Another key machine learning type is reinforcement learning, which operates without labeled data and is applied in specific scenarios where the machine figures out how to reach a designated goal.

Reinforcement learning shares with supervised machine learning the understanding of the desired output. We essentially direct the computer to optimize our goal achievement. But unlike supervised learning, we do not provide labeled data to the machine learning model. Instead, we create rules.

The machine learning model learns through trial and error within specific parameters of our created rules.

Reinforcement learning thrives in robotics and online recommendation systems. For example, platforms like Netflix use reinforcement learning to enhance recommendation systems.

Unlike traditional methods that rely on labeled data, which indicates which TV shows a user might like, reinforcement learning employs a system where the model learns from user feedback through interactions.

Initially, Netflix may provide random show recommendations. As a user interacts with these suggestions by watching a show, adding it to their list, or skipping it, the system gathers feedback. Positive interactions, like watching a full episode, signal the model to recommend similar content in the future. Negative responses, like skipping a show, guide it to adjust recommendations.

Over time, the model refines its predictions to better align with individual user preferences, constantly adapting to new viewing patterns and behaviors.

That is the intuition behind supervised, unsupervised, and reinforcement learning.

In the next lesson, we will examine deep learning, a truly fascinating feat of computer science.

Key Takeaways

• Supervised learning uses labeled data to train models for classification and regression tasks.
• Unsupervised learning identifies patterns and groups in unlabeled data, useful when labels are unavailable or costly.
• Reinforcement learning optimizes goal achievement through trial and error within defined rules, without labeled data.
• Applications include image classification, price prediction, customer segmentation, robotics, and recommendation systems.

Deep Learning

Deep learning is a fascinating subset of machine learning inspired by how the human brain works.

Here is an AI-generated image. What do you see? At first glance, it’s a sunny day and a crowded beach, right? Upon closer inspection, we see children playing in the sand around a massive sand castle at the center.

We then examine each person on the beach more closely. We find that I rendered this individual’s face oddly, resulting in a bizarre and slightly unsettling appearance.

Our brain processes information in various phases and at varying depths. Initially, viewing an image provides a raw, broad impression that presents insights into the scene’s context. The more time and attention we devote to details, the more information and subtleties we can process and observe.

In the context of deep learning, a neural network processes information similarly. Here is what a neural network resembles. Please do not be scared; everything will make sense in a second.

Picture the first layer as our input information, similar to observing a sunny, crowded beach day. Then, as the data passes through more network layers, intermediate layers start recognizing more complex features such as shapes or specific objects.

Remember how I noticed the enormous sand castle in the middle? Every intermediate layer of the neural network builds a more detailed understanding of the basic features identified by earlier layers. So there is an incremental increase in the level of detail acquired.

The deeper layers of the network synthesize lower-level features into high-level features, representing more complex aspects of the input data. This is how we can spot the strange face generated by AI. It took some time to process the information and reach this conclusion.

Honestly, I think I subconsciously wanted the picture to reflect that because I was curious if the AI had produced a quality example. Eventually, my brain focused on this little detail.

Okay, perfect. So that’s the intuition behind deep learning. It is a complex process that allows machines to learn by processing input information in stages.

Let’s explore some technical details to deepen our understanding. We call this neural network an artificial neural network, or ANN. It is inspired by biological neural networks, but they work much differently.

Like our senses and ANNs, the input layer sends raw data to the brain. Then the intermediate layers, or hidden layers, process the input information. Neural networks can have one or multiple hidden layers. Increasing layers enhances complexity.

Adding more layers to a network increases its learning capacity, but the downside is that it needs to be carefully managed to ensure effective learning. We also have an output layer that generates the final result.

Every layer of the artificial neural net is made of neurons, or nodes, responsible for processing and transforming the information received.

If you recall, I previously referenced the MNIST example in one of our lessons. We noted that training a model to recognize handwritten digits involves supplying it with thousands of pre-labeled examples.

How does this training happen? While the process may initially appear complex, let’s simplify it to a more manageable form.

The process starts with the input layer of the ANN receiving an image of a handwritten digit. Each pixel of this image serves as an input node.

In the case of MNIST, images are 28 by 28 pixels, so the input layer typically has 784 input nodes. Imagine that the 784 input nodes are stored as a vector inside the neural net to form the input layer.

Each of the 784 input nodes contains a number based on how bright or dark it is. Here, zero represents white, while any value greater than zero indicates a color other than white.

We call this number activation. The higher the number, the darker the content inside a given node.

We describe the 784 input nodes in the input layer as the neural net’s width and its number of layers as its depth. Here, the network comprises three hidden layers, along with the input and output layers, totaling a depth of five.

We have numerous connections between nodes because each layer’s nodes are linked to every node in the subsequent layer. This extensive network of connections is essential for learning from input data, serving as mathematical transformations.

These transformations occur through a mix of weights and non-linear operations. With optimal weight combinations across all nodes, learning is enabled. This method translates the input via various layers, refining the information before it generates a result.

Learning occurs by designing a system that identifies optimal weights and biases to solve a specific problem, involving thousands of repetitions to discover the best combination.

So why do we need several layers? What kind of transformations can we expect in each layer to end up with a system capable of recognizing the digits in a photo?

Well, when we see an image in the background, our brain connects various components of what it is made of. For instance, the number three comprises two elements: a rounded top and a curved bottom.

The first hidden layer in our neural network might learn to recognize these features from the input data. Then, in the second hidden layer, we continue to build on the information processed by the previous one.

We’ve identified edges and curves, and now we can use them to discern more complex shapes like loops and intersections, bringing us closer to recognizing numerals.

When the information reaches the third layer, the neural network has learned to recognize the overall shape and configuration representing the number three.

Finally, the output layer takes the processed data from the last hidden layer and determines whether the digit is a three.

This is how a neural network learns to perform what appears to be a simple task: recognizing a number. The underlying process, however, involves intricate mathematical manipulations which we did not describe fully.

In essence, the layers of an ANN, through their depth and breadth, create a robust system of pattern recognition and data interpretation that mirrors some aspects of human cognitive processes, albeit more structured.

This capability to analyze large, high-dimensional data sets and recognize complex patterns with high accuracy makes deep learning a revolutionary tool in AI. It is what has made today’s incredible AI advancement possible.

Key Takeaways

• Deep learning is inspired by the human brain’s layered processing of information.
• Artificial Neural Networks (ANNs) consist of input, hidden, and output layers that progressively extract complex features.
• The depth and width of a neural network determine its capacity to learn and recognize patterns.
• Deep learning enables machines to analyze high-dimensional data and recognize complex patterns with high accuracy.

Robotics

Tales of mechanical beings date back to ancient times.

For instance, the myth of Talos depicts a colossal bronze man built by Hephaestus, the Greek god of invention and blacksmithing, tasked with safeguarding Crete.

This bronze figure patrolled the island three times a day, hurling boulders at enemy ships nearing the coast.

In the medieval European and Islamic world, devices like al-Jazari’s automata, including water clocks and programmable machines, showcased early engineering prowess.

One of the greatest pioneers of the Italian Renaissance, Leonardo da Vinci, sketched designs of a mechanical knight capable of limited movements, foreshadowing today’s humanoid robots and a mechanical lion.

So the idea of robots and the robotics field is not new. People have been fascinated with them for a long time.

It is a true phenomenon in pop culture that has captured the imagination of previous generations.

Modern Robotics and Technological Advances

So what is different today? Rapid technological advancements and mainly the incredible growth of AI are making possible the creation of intelligent machines, which we could only imagine and depict in movies in the past.

Robotics is the branch of technology that deals with the design, construction, operation, and use of robots—machines that can perform tasks automatically or with human-like capabilities.

Let’s break this down.

• Mechanical engineers are needed to construct robots’ physical structures. They are responsible for designing and constructing robots, including their mobility mechanisms.
• Electronics and electrical engineers design systems that enable the robot to operate and control its actions.
• While electronics power the system, various types of AI drive the robot’s decision making and behavior.
• Robots are equipped with various sensors and cameras to collect necessary data and perceive their surroundings.
• This is how AI developers and engineers pitch in.

So this truly interdisciplinary field requires specialized skills in several distinct areas.

Like other types of AI, the robotics domain is inspired by human capabilities. Robots are designed to mimic and augment human abilities.

To achieve that, researchers use a multidisciplinary approach combining different AI technologies.

Typically, we should consider a system with multiple models instead of a single model that does it all.

For example, an automata robot might use computer vision for object detection and environment understanding, simultaneous localization and mapping for navigation and mapping, reinforcement learning for decision making, and a natural language processing model for understanding and generating human language.

Combining all these models allows a robot to perceive the environment, make decisions, communicate with people, and act accordingly.

Use Cases of Robotics

Let’s point out several intriguing use cases.

You probably know that under the direction of Elon Musk, Tesla is building a Tesla Bot, which is described as a general purpose robotic humanoid.

The goal is to create a robot that can relieve humans of dangerous, repetitive, and boring tasks in logistics.

For example, a Tesla Bot could help with factory and warehouse work, moving and stacking boxes, counting and auditing inventory, and lifting weights.

And it is not just Tesla. Many companies develop their own autonomous robots and contribute with valuable research.

Medical robots are a fascinating use case already in the mass adoption phase. These robots can perform incredibly accurate medical interventions and even complex surgeries which save lives.

There are so many amazing innovations that are becoming mainstream:

• Self-driving cars
• Harvesting robots
• Cleaning robots
• Space robots
• Search and rescue robots
• Security and surveillance robots

The future is here, and AI is at the forefront of all of these innovations.

Key Takeaways

• Robotics has ancient origins, with myths and early automata illustrating humanity’s long-standing fascination.
• Modern robotics integrates mechanical, electrical, and AI engineering to create intelligent machines.
• Robots combine multiple AI models such as computer vision, reinforcement learning, and natural language processing to perform complex tasks.
• Current applications include humanoid robots like Tesla Bot, medical robots, autonomous vehicles, and various specialized robots across industries.

Running out of data

When we see diagrams like this one, it is natural to expect that the next version of a language model (LM) will be significantly larger than the previous one.

However, one major issue is likely to hinder this process.

If GPT-4 has already read and processed a significant part of the publicly available data on the internet, where would the new data for GPT-5 come from?

AI developers risk running out of data and face a potential slowdown in the growth of model capabilities.

Moreover, now that ChatGPT has gained popularity, there is a significant risk that many of the new articles and materials posted online will be written using AI.

This makes learning new or nuanced information challenging for future LLMs, because they frequently encounter repetitive or derivative content.

Future models will simply echo what previous models have written, thereby increasing the risks of hallucination, biases, and inaccuracy.

Another obstacle to data availability is that, following the hype surrounding generative AI, several large organizations worldwide filed lawsuits and prohibited data scraping of their websites.

Some of the first litigations came from The New York Times, Shutterstock, and best-selling author John Grisham.

Platforms like Reddit and Quora changed their policies to make it illegal for AI developers like OpenAI to scrape their platform data.

To mitigate this issue, OpenAI has started a content licensing program and has already signed agreements with several organizations that own some of the largest proprietary data sets in the world.

OpenAI offered publishers between 1 and 5 million dollars a year to access archives and train its generative AI models.

They signed deals with organizations like Shutterstock, Axel Springer, The Associated Press, Le Monde, and Prisa Media.

In the future, competition between big tech firms developing LLMs will likely significantly increase the price of large proprietary datasets.

Key Takeaways

• The availability of new, quality data is a critical challenge for developing next-generation language models.
• Increasing AI-generated content online risks causing future models to learn repetitive or derivative information.
• Legal restrictions and lawsuits have limited data scraping from major platforms, impacting data collection.
• OpenAI has initiated content licensing agreements with major data owners to secure proprietary datasets for training.