How AI Learns: Making Machines Learn and Do Their Jobs

On the internet, the algorithms are all around you. You are watching this video because an algorithm brought it to you (among others) to click on, which you did, and the algorithm took note.

When you open the TweetBook, the algorithm decides what you see. When you search through your photos, the algorithm finds them for you and might even create a little movie. When you buy something, the algorithm sets the price, and at your bank, an AI algorithm monitors transactions for fraud. The stock market is full of algorithms trading with algorithms.

Given this, you might want to know how these little algorithmic bots, which shape your world, actually work, especially when they don’t. In this article, we’ll go into detail about how AI learns. Keep reading!

What Is the Purpose of AI and Its Algorithms?

The principal aim of AI and its algorithms is to imitate, enhance, and sometimes even outperform human intelligence in performing particular tasks. AI seeks to simplify complex processes, amplify efficiency, and make informed decisions based on data, in ways that humans either can’t or would find painfully time-consuming and prone to error. Today, AI already plays a huge role in many aspects of life, revolutionizes industries, enhances customer experiences, and amplifies problem-solving skills.

Key Capabilities of AI Today

AI is no longer just a futuristic concept—it is actively shaping industries and improving daily life. Here are some of the things AI can do today:

• Automation of Routine Tasks

  • AI-powered systems automate tasks like sorting emails, answering customer queries, or managing supply chains. For example, Amazon’s warehouse robots process millions of items daily, improving operational efficiency.

  • In customer service, 80% of customer interactions are predicted to be managed by AI chatbots by 2025, according to a report by Gartner.

• Natural Language Processing (NLP)

  • NLP powers voice assistants like Siri, Alexa, and Google Assistant, which have become ubiquitous in everyday life. By the end of 20214, Demandsage reported that over 8.4 billion digital voice assistants were in use worldwide.

  • AI enables real-time language translation, such as Google Translate, which supports over 100 languages.

  • AI models like ChatGPT are capable of generating human-like text and engaging in complex conversations, showcasing advancements in NLP.

• Image and Speech Recognition

  • AI is used for facial recognition, an artificial intelligence technology that is already employed in security systems, such as in airports and smartphones. The global facial recognition market is projected to reach $8.44 billion by 2030, as reported by Statista.

  • AI is widely used in medical imaging to detect conditions such as cancer. For instance, AI-based systems can identify tumors in mammograms with accuracy rates surpassing human radiologists in some cases. A study in Health’s National Library of Medicine found that AI models could diagnose breast cancer with 93% accuracy.

• Data Analysis and Predictions

  • AI analyzes massive datasets to identify patterns and make predictions. For instance, IBM Watson is used in healthcare to analyze medical records and recommend personalized treatment plans. AI-driven systems like DeepMind have been used to predict protein folding with a high level of accuracy, aiding drug discovery.

  • In finance, AI algorithms help detect fraud and assess risks. Visa’s AI system prevents an estimated $25 billion in annual fraud, identifying fraudulent activities with an error rate lower than 0.1%.

• Personalization

  • AI is at the heart of recommendation systems on platforms like Netflix, Spotify, and Amazon, offering personalized recommendations based on user behavior. For example, Netflix claims that over 75% of the content watched on the platform is driven by AI-powered recommendations.

  • AI helps businesses target ads more effectively. According to McKinsey, companies that use AI for marketing and sales see an average increase in sales of 10% to 20%.

• Advanced Problem-Solving

  • AI is used in healthcare to analyze patient data and assist in diagnosing conditions like cancer, heart disease, and diabetes. In radiology, AI models have demonstrated accuracy rates of up to 85% in detecting early signs of lung cancer.

  • In logistics, AI helps optimize delivery routes and reduce fuel consumption. For instance, UPS uses AI-powered systems to optimize delivery routes, which saves the company an estimated 10 million gallons of fuel annually.

• Creative Tasks

  • AI can generate art, music, and even text, as seen with platforms like OpenAI’s DALL·E (image generation) and Jukedeck (music composition). If you’d like to build a similar AI solution, learn how in our article.

  • In journalism, AI algorithms help write news stories, particularly in areas like sports and finance. The Associated Press uses AI to generate automated reports for earnings reports and sports results, processing thousands of articles each year.

The Broader Purpose of AI

AI is designed to augment human capability, lighten the workload, and solve problems that are otherwise impossible to handle manually. Be it by detecting diseases at an early stage to save lives, simplifying routine tasks, or opening ways to revolutionary inventions, AI is gradually changing the way we live, work, and interact with each other.

The benefits are already very noticeable and will continue to increase as artificial intelligence technology advances, finding applications in almost all aspects of life: healthcare, transportation, finance, and entertainment, among others. While AI has already achieved incredible feats, its true potential is still emerging, and the capacity of the technology to revolutionize our world is only just beginning to unfold.

The Evolution of Algorithmic Bots

People initially built algorithmic bots by giving them instructions that humans could explain: “If this, then that.” However, many problems are simply too complex for humans to write straightforward instructions for.

Consider these challenges:

• There are a gazillion financial transactions happening every second. Which ones are fraudulent?

• NetMeTube hosts millions of videos. Which eight should be recommended to a user? Which ones shouldn’t be allowed on the site at all?

• For this airline seat, what is the maximum price a user is willing to pay right now?

Algorithmic bots provide answers to these questions. While the answers aren’t perfect, they’re far better than what a human could achieve. However, no one fully understands how these bots work—not even the humans who built them. Or perhaps, more accurately, “built them,” as we’ll explore further.

Companies that use algorithmic bots often refuse to disclose how they work because these bots are incredibly valuable—essentially functioning as their most productive employees. The intricate structure of these bots’ “brains” is a fiercely guarded trade secret.

Currently, the cutting-edge development of these bots likely relies heavily on concepts like linear algebra. But the exact methods powering the latest breakthroughs, and how the bots truly function, remain a mystery. This is something that will probably always be the case.

Why Are Bots So Secretive?

• Valuable Employees: Bots perform essential tasks and hold significant value for companies.

• Competitive Edge: Revealing bot mechanisms could compromise a company’s advantage.

• Complexity: Even experts struggle to explain the intricacies of bot functionality.

How Does AI Learn: Building Algorithmic Bots?

So, let’s explore one of the simpler yet more comprehensible ways bots can be “built” without fully understanding how their “brains” work.

Imagine you want to create a bot that can identify objects in a picture. For example, is it a bee or a three? Humans can easily make this distinction (even young children), but it’s impossible to directly program a bot to do the same.

Why? Because humans intuitively know that’s a bee and that’s a three. We can describe the differences in words, but bots don’t understand words. Their ability to process this distinction lies entirely in the intricate wiring of their artificial “brains,” which mimics the way ours work—but only to an extent.

The Role of Builder and Teacher Bots

While the function of an individual neuron might be understood, and the general purpose of clusters of neurons roughly grasped, the overall system remains beyond complete comprehension. Yet, it works.

To create a bot capable of sorting tasks like recognizing a bee versus a three, you don’t build it manually. Instead, you create two other bots: a builder bot that constructs bots and a teacher bot that evaluates them. These foundational bots have simpler “brains” that a skilled human programmer can design.

The builder bot’s job is to assemble new bots, but it’s not very good at it initially. It starts by randomly connecting wires and modules in the “brains” of the bots it creates. This results in some very “special” student bots, which are then sent to the teacher bot for evaluation.

The teacher bot, however, doesn’t have the ability to recognize a bee from a three either—if it could, the problem would already be solved. Instead, the teacher bot is equipped with a set of labeled “bee” and “three” images, along with an answer key. Its role isn’t to teach but to test the student bots.

The Challenge of Teaching Bots

• Human Intuition vs. Bot Logic: Distinguishing a bee from the number three is intuitive for humans but challenging to explain in bot-readable terms.

• Bots Don’t Understand Words: Unlike humans, bots need specific wiring and logic to “recognize” objects.

The Solution: A System of Builders, Teachers, and Students

To teach bots complex tasks, developers use:

1. Builder Bots: Create student bots with random configurations.

2. Teacher Bots: Test the student bots without teaching them.

3. Student Bots: Attempt tasks and get graded based on their performance.

The Student Bots’ Learning Process

The adorkable student bots stick out their tongues and try very hard, but they’re not good at what they’re supposed to do. And it’s not their fault—they were built that way. After the test, they return to the builder bot, holding their grades. The best-performing bots are set aside, while the others are recycled.

The builder bot is still not great at building bots, but now it takes the surviving ones and creates new copies with slight variations. Back to school, they go. The teacher bot tests them again, and the builder bot builds again, over and over.

Iterative Improvement Through Testing

Now, a builder that works at random, a teacher that doesn’t teach but only tests, and students who can’t learn—they just are what they are—shouldn’t work in theory. But in practice, it does. This is partly because, in every iteration, the builder bot’s “slaughterhouse” keeps the best bots and discards the rest. It’s also because the teacher bot isn’t managing an old-timey, one-room schoolhouse with a dozen students; instead, it oversees an infinite warehouse with thousands of students.

The test isn’t just ten questions, but a million. How many times does the build-test loop repeat? As many times as necessary. At first, the students that survive are just lucky, but by combining enough of these lucky bots, keeping only what works, and randomly experimenting with new copies of them, eventually, a student bot will emerge that isn’t just lucky. This bot can perhaps barely tell bees from threes.

As this bot is copied and altered, the average test score gradually rises, and the grade required to survive the next round gets higher. With enough iterations, eventually, from the infinite warehouse (or “slaughterhouse”), a student bot will emerge that can tell a bee from a tree in a photo it’s never seen before with impressive accuracy.

Iterative Building and Testing

1. Random Construction: Builder bots randomly design student bots’ internal wiring.

2. Testing Phase: Teacher bots test student bots using a database of examples and answer keys.

3. Survival of the Fittest: Top-performing student bots are modified and re-tested, while the rest are discarded.

4. Repeat Cycle: The process is repeated until a bot performs the task with high accuracy.

Scaling Up the Process

• Massive Data: Tests involve millions of examples to ensure bots learn effectively.

• Incremental Improvements: Each iteration improves performance, even though initial successes are often random.

The Complexity of Bot Logic

However, how the student bot achieves its results is something neither the teacher bot, the builder bot, nor the human overseer can truly understand. Not even the student bot itself knows how it works.

After incorporating countless useful random changes, the wiring in its “brain” becomes incredibly complex. While an individual line of code might be understood, and clusters of code can be vaguely identified by their general purpose, the entire system is beyond human comprehension. Yet, it works. This can be both impressive and frustrating because the student bot excels only at the specific types of questions it was trained for. It might perform well with photos, but fail with videos or struggle if the photos are flipped upside down. Sometimes, it confidently misidentifies things that are clearly not bees as bees.

Since the teacher bot cannot actually teach, the human overseer can only provide more questions to extend the test. The goal is to include the types of questions where the best-performing bots tend to fail. This is a critical point to understand—it explains why companies are so focused on collecting vast amounts of data. More data allows for longer, more comprehensive tests, which ultimately result in better-performing bots.

Why Bots Are Hard to Understand

• Complex Wiring: After countless iterations, bots’ internal logic becomes so intricate that even developers can’t fully explain it.

• Task Specialization: Bots excel at specific tasks (e.g., recognizing photos) but struggle with variations (e.g., upside-down images).

The Role of Data

• More Data, Better Bots: Longer and more comprehensive tests improve bot accuracy.

• Human Input: CAPTCHA tests and other user activities contribute to creating better test datasets for bots.

Human Interaction with Bots

When you encounter an “Are you human?” test on a website, you’re not only confirming your humanity (hopefully), but you’re also contributing to building tests that help bots learn to read, count, or differentiate between lakes and mountains—or horses and humans. Have you noticed an uptick in questions about driving lately? Hmm… What could those tests be preparing bots for?

Teaching bots to recognize objects in photos, interpret signs, or filter videos relies on humans creating accurate and extensive tests. However, there’s another type of test that doesn’t require a human setup—tests conducted on humans themselves.

Hypothetical Case Study

For example, let’s consider a platform aiming to keep users watching as long as possible. Measuring how long a user stays on the site is straightforward. The teacher bot assigns each student bot a group of the platform’s users to monitor. The student bots observe what their users watch, analyze their preferences, and attempt to recommend videos that will maximize user engagement. The longer the users stay on the platform, the higher the student bot’s test score.

Build, test, repeat. After a million iterations, a student bot emerges that is relatively successful at keeping users watching—better than anything a human programmer could design.

But when people ask, “How does the platform’s algorithm choose videos?” the answer is far from simple. Essentially, the algorithm’s behavior is based on:

• The bot itself.

• The user data it analyzed.

• The scoring system created by human overseers to guide the teacher bot.

The student bot’s survival hinges on excelling at the specific test it was designed to pass. However, what the bot is “thinking” or how it arrives at its decisions remains a mystery—even to its creators. All we know is that this student bot became “the algorithm” because it performed just slightly better—perhaps 0.1%—than its predecessor in the test designed by humans.

Real-World Applications of Bots

CAPTCHA Tests

• Validate users as human.

• Help bots learn to recognize text, objects, or scenes.

Streaming Platforms

• Bots are optimized to increase user engagement by recommending videos.

• Success is measured by how long users stay on the platform.

Internet Algorithms

• Bots optimize for specific goals, such as:

  • Increasing user interaction.

  • Setting prices to maximize revenue.

  • Selecting content users are likely to engage with.

Bots’ Role Across the Internet

Everywhere on the internet, behind the scenes, algorithms are constantly at work. They run tests to:

• Increase user interaction.

• Set prices optimally to maximize revenue.

• Select posts from your friends that you’re most likely to enjoy.

• Recommend articles that people are most likely to share.

If something can be tested, it can be taught—well, “taught.” A successful student bot will eventually graduate from the endless cycles of testing and iteration, becoming the dominant algorithm in its domain—at least, until it’s outperformed by the next iteration.

Wrapping Up

We’re accustomed to the idea that the tools we use, even if we don’t fully understand them, are understood by someone else. However, with machines that learn, we’re increasingly in a situation where we rely on tools—or are influenced by tools—that no one, not even their creators, fully comprehends.

Our only option is to guide these tools through the tests we design. It’s a reality we must adapt to because our algorithmic bot companions are everywhere—and they’re here to stay.

If you want to know more about AI, and its algorithms, or aim to create your own AI, feel free to contact us! Our AI department will answer all your questions.