Part 2 — How machines Learn

After discussing Turing's (1950) foundational ideas on machine intelligence in the previous article, we now turn to the next pivotal development in AI: backpropagation, introduced by Rumelhart, Hinton, and Williams (1986). This breakthrough enabled the training of deep neural networks, forming the backbone of modern Large Language Models (LLMs) like GPT and BERT.

What is a Neural Network?

To appreciate backpropagation, let's start with the basics. A neural network is a computer system modeled loosely on the way our brains work. It consists of layers of artificial neurons (or nodes) connected to one another. Each layer processes input (like images, text, or data), performs computations, and passes the result to the next layer. These networks can perform tasks like recognizing objects in images, classifying text, or making predictions.

However, in order for a neural network to get better at these tasks, it needs a way to learn from its mistakes. This is where backpropagation comes in.

What is Backpropagation?

Backpropagation is the learning method that allows a neural network to improve itself by adjusting the internal connections between neurons. Here’s how it works:

• Initial prediction: The network takes an input (say, an image of a handwritten number) and makes a prediction (for example, it thinks the number is "5").
• Error calculation: It compares its prediction to the actual correct answer (the number is really "7") and calculates the error.
• Learning from the error: Backpropagation sends this error backwards through the network, telling each neuron how much it contributed to the mistake. The network then adjusts the weights (the strength of connections between neurons) to reduce future errors.

Simple Example: Throwing Darts

Think of backpropagation like teaching someone to throw darts. If you’re learning to hit the bullseye, each time you miss, you learn something about how far off you were. If the dart goes too high, you adjust your next throw by aiming lower. Over time, you improve by making small corrections based on each miss.

In a neural network, backpropagation works similarly—it takes the error from each wrong prediction and adjusts the network so that next time it "throws," it's more likely to be accurate.

More Examples to Understand Backpropagation

1. Voice Assistants: Consider voice assistants like Siri or Alexa. When you ask them a question and they respond incorrectly, the system learns from the mistake using backpropagation. With every interaction, the model refines its understanding of your voice, accents, or unusual phrases to improve its responses over time.
2. Handwriting Recognition: Imagine training a neural network to recognize handwritten digits. At first, it might confuse "3" with "8." By using backpropagation, the network learns from these mistakes and gets better at distinguishing between similar-looking numbers.
3. Spam Filters: Backpropagation is also used to train email spam filters. When a filter mistakenly allows a spam email into your inbox, it uses the error to adjust its internal rules. Over time, it gets better at identifying patterns that indicate spam versus regular email.

Why Backpropagation Is Crucial for LLMs

Backpropagation made it possible to train deep neural networks—networks with many layers of neurons—effectively. This is especially important for LLMs, which are trained on vast amounts of data and consist of millions (or even billions) of parameters.

For example:

• Language Prediction: When a model like GPT predicts the next word in a sentence, backpropagation helps it adjust if the prediction is wrong. This enables it to learn grammar, context, and even nuances like humor or tone over time.
• Machine Translation: If an LLM translates a sentence inaccurately, it uses backpropagation to refine its understanding of languages. This leads to better translations with continued use.

In essence, backpropagation allows LLMs to learn from huge amounts of text, improve their predictions, and provide us with more accurate and human-like outputs.

Connecting Back to Turing

In Turing's (1950) paper, he raised the question, "Can machines think?" Hinton’s work on backpropagation shows how machines can learn. Turing imagined a future where machines could mimic human intelligence, and backpropagation is the key mechanism that allows AI systems to develop that intelligence by learning from data, much like humans learn from experience.

Read and Explore

• Read the Rumelhart, Hinton, Williams (1986) paper here or access it directly here.
• If you missed Part 1 on Turing (1950): "Computing Machinery and Intelligence", you can find it here.

Stay tuned for more research papers that paved the way for LLMs, and feel free to share your thoughts or notes in the comments below!