Decoding Transformers: A StepbyStep Explanation

The AI boom, fueled by tools like ChatGPT, DALLE, and Midjourney, is built on a revolutionary technology: the Transformer. But what exactly is a Transformer? This article breaks down the inner workings of these powerful neural networks, offering a visuallydriven explanation to help you understand how they operate.

What Does GPT Mean?

Let's start with the basics: GPT stands for Generative Pretrained Transformer.

• Generative: These models generate new text.
• Pretrained: They learn from massive amounts of data.
• Transformer: The core neural network architecture.

How Transformers Work: A HighLevel Overview

At a high level, here's what happens when a Transformer generates text:

1. Tokenization: The input text is broken down into small pieces called tokens (words, parts of words, or even character combinations). For images or sound, tokens are small patches or chunks.
2. Embedding: Each token is converted into a vector – a list of numbers representing its meaning. Similar words have vectors that are close to each other in a highdimensional space.
3. Attention Block: The vectors interact with each other, passing information back and forth. This allows the model to understand the context of each word. For example, differentiating between "machine learning model" and "fashion model".
4. MultiLayer Perceptron (MLP) / FeedForward Layer: The vectors undergo a parallel operation, akin to answering questions about each vector and updating them based on the answers.
5. Repetition: The process iterates between attention blocks and MLP blocks.
6. Prediction: The final vector is used to generate a probability distribution over all possible next tokens.

This process of repeated prediction and sampling is what allows Transformers like ChatGPT to generate coherent text one word at a time.

From Prediction to Generation: The Power of Sampling

The core of a Transformer is a prediction model. Give it a snippet of text, and it predicts the next word. To generate longer text, the model:

1. Receives an initial snippet (the seed text).
2. Samples a word from the probability distribution it generates.
3. Appends the sampled word to the text.
4. Repeats the process, using the new text as input.

Larger models, like GPT3, are significantly better at generating coherent and sensible stories compared to smaller models like GPT2. This highlights the importance of scale in achieving advanced AI capabilities.

Deep Learning Fundamentals: A Quick Recap

Transformers are a type of deep learning model. Deep learning is a branch of machine learning where models learn from data to perform tasks like image recognition or text prediction. Instead of explicitly programming a procedure, you provide the model with examples and let it learn the patterns itself.

Key principles of deep learning relevant to understanding Transformers:

• Input as Arrays of Numbers: All inputs must be formatted as arrays of real numbers (tensors).
• Layered Transformations: Data is transformed through multiple layers, each structured as an array of numbers.
• Weighted Sums: Parameters (weights) interact with data through weighted sums, often packaged as matrixvector products.
• Backpropagation: The training algorithm used to adjust the model's weights.

Word Embeddings: Mapping Meaning to Vectors

A crucial step in processing text is converting words into vectors – a process called word embedding. Each word is associated with a vector that represents its meaning in a highdimensional space.

This allows the model to capture semantic relationships between words. For example:

• Words with similar meanings are located close together in the vector space.
• Directions in the space can represent semantic concepts (e.g., gender, nationality).
• Mathematical operations on word vectors can reveal surprising relationships (e.g., king man + woman ≈ queen).

The embedding matrix defines the vector representation for each word in the vocabulary. In GPT3, the vocabulary size is 50,257 tokens, and the embedding dimension is 12,288, resulting in approximately 617 million weights.

The Role of Context and Context Size

The vectors not only represent individual words but also have the capacity to absorb context from the surrounding text. The network processes a fixed number of vectors at a time, known as the context size. GPT3 was trained with a context size of 2048, meaning it can incorporate 2048 tokens of text when making a prediction. This context size limits how much information the Transformer can consider, potentially leading to the bot losing track in very long conversations.

Predicting the Next Word: The Unembedding Matrix and Softmax

The final step involves predicting the next word based on the processed vector. This is achieved through two key components:

• Unembedding Matrix: This matrix maps the final vector to a list of values, one for each token in the vocabulary.
• Softmax Function: This function normalizes the values into a probability distribution, ensuring that each value is between 0 and 1 and that they all add up to 1.

The Unembedding matrix in GPT3 contributes another 617 million parameters to the network, bringing the total to over a billion.

Softmax in Detail

Softmax transforms arbitrary numbers into a valid probability distribution. It does this by first raising 'e' to the power of each number and then dividing each term by the sum of all those positive values. This ensures the output is a normalized list that adds up to 1.

Temperature

The temperature parameter in the softmax function controls the randomness of the output. A higher temperature leads to a more uniform distribution, while a lower temperature makes the model more deterministic and predictable.

Conclusion: Building a Foundation for Understanding Attention

Understanding word embeddings, softmax, dot products, and the overall framework of deep learning is crucial for grasping the attention mechanism, the cornerstone of modern AI. With this foundational knowledge, you're wellequipped to dive deeper into the inner workings of Transformers and explore the next level of AI innovation.