Attention mechanisms in ChatGPT for crafting effective responses

In artificial intelligence, OpenAI’s ChatGPT has emerged as a pioneer, captivating users with its ability to engage in human conversations. At the heart of this technological marvel lies a critical component that orchestrates the magic—the attention mechanism. The attention mechanism was first introduced in a publication, “Attention is All You NeedVaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. “Attention is all you need.” Advances in neural information processing systems 30 (2017).,” by Google. In this blog, we delve into the inner workings of the attention mechanism and how ChatGPT utilizes it to generate effective responses.

The attention mechanism and transformer

The attention mechanism is a component in deep learning models that mimics cognitive attention. Like the human ability to selectively pay attention to different elements of information, the attention mechanism allows models to focus on specific parts of the input data when making predictions or decisions. We’ll explain the detailed working of the attention mechanism in the following sections.

Before the attention mechanism, recurrent neural network (RNN) models were used widely to model sequential dependencies for natural language processing (NLP) tasks. These models suffer from vanishing gradients and have limited parallelization because they are sequential. Due to these limitations, extracting meaningful relationships between distant elements in a text was very difficult.

The foundation of ChatGPT’s conversational expertise is laid upon the transformer model. The following figure demonstrates the transformer model originally presented in 2017.

The transformer model mainly consists of the following different steps:

1. Tokenization and input encoding: This step converts each word, called a token, into a vector of fixed length—for instance, 512The vector size can be 512 in transformer models like ChatGPT. While models like GPT-2 and GPT-3 commonly use larger vector sizes, such as 768 or 1024, there is flexibility in choosing the vector size based on the specific requirements of the model. However, for demonstration purposes, the vector size is kept at 512..

2. Positional encoding: This is performed to capture the spatial information of each token. It provides information about each token’s position in the sequence.

3. The attention mechanism: The attention is performed to capture the importance or relevance of each token to other tokens.

4. The feed-forward network: It processes and enhances the information captured by the attention mechanism, applying linear transformations and nonlinear activation functions to capture complex patterns within the input sequence.

Let’s delve deeper into each of the aforementioned steps in relation to ChatGPT in the following sections.

1. Tokenization and input encoding

At the outset, ChatGPT dissects input words into smaller units through tokenization. Each token is then translated into initial embeddings, providing a numerical representation for every input fragment. This step is crucial for enabling the model to work with the intricacies of language.

Consider the sentence, “Mysterious footsteps echoed in the silent forest.” In the given sentence, each word is considered a token. Also, each word is converted to a vector with a dimension $d_{model}=512$.

Note: For illustration purposes, all the numbers in this blog are randomly generated. However, for actual model training, these numbers are generated using predefined algorithms, for example, Universal Sentence EncoderCer, Daniel, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant et al. “Universal sentence encoder.” arXiv preprint arXiv:1803.11175 (2018)..

2. Positional encoding

Positional encoding recognizes the significance of word order and captures the spatial information of each word in a sentence. Without positional encoding, the model might consider different permutations of the same words as equivalent, leading to potential confusion. For example, “The sun sets behind the mountain” and “The mountain sets behind the sun” would have the same representation without positional encoding.

Positional encoding ensures that the model not only comprehends the semantics of words but also understands their positions within the input sequence, preserving the temporal shades of language.

The positional encoding of each word is also a vector of size $d_{model} = 512$, which is added to the corresponding embedding vectors of each token, as illustrated below:

After adding the embedding and position encoding vectors, the result is provided as input to the encoder.

3. The attention mechanism

The attention mechanism in the transformer model is used to capture long-range dependencies and generate a context-aware representation for each token in the sequence based on its relationships with other tokens. For example, consider the following two sentences:

• “She poured milk from the jug into the glass until it was full.”

• “She poured milk from the jug into the glass until it was empty.”

As humans, it is easy to understand that “it” refers to the glass in the first sentence, while in the second, it refers to the jug. However, for machine learning models, this relationship of words is identified using the attention mechanism.

The transformer model used a multi-head self-attention mechanism. However, to understand it, first, we need to have an in-depth understanding of the self-attention mechanism.

Terminology alert:

The attention mechanism operates on a set of queries consolidated into a matrix $Q$. Additionally, the keys and values are grouped together into matrices $K$ and $V$, respectively. The dimension of each of these matrices is $d_{sequence} \times d_{model}$, where $d_{sequence}=7$ and $d_{model}=512$ for the input sentence “Mysterious footsteps echoed in the silent forest.” To understand how these matrices are initially created, you can refer to this Educative answer about what the intuition behind the dot product attention is.

Self-attention

Before discussing multi-head self-attention, it’s necessary to understand the self-attention mechanism. The self-attention mechanism computes the importance of different words in a single sequence concerning each other. We will assume our previous example, where $d_{sequence}=7$ and $d_{model}=512$. Self-attention is computed using the following formulation:

The result of self-attention is a $d_{sequence}\times d_{model}$ matrix that represents how much attention each position in a sequence gives to other positions.

The softmax function generates similarity scores of each word with other words within the range of 0 to 1 (probability values), as depicted below:

Multi-head attention

Multi-head attention enables the model to capture different aspects or patterns in the relationships between words, enhancing its ability to learn diverse and complex dependencies. It extends self-attention by running it in parallel multiple times. The inputs ($Q$, $K$, and $V$) are linearly transformed into multiple subsets. Each input is processed independently through several self-attention blocks called heads. For example, if we consider eight heads ($h$), the input dimension to each head would be ${d_{model} \over h}={512 \over 8} =64$. Let’s denote this value by $d_{k}$.

Let’s understand the working of multi-head attention in different steps:

• The matrices $Q$, $K$, and $V$ are multiplied with their respective weight matrices $W^Q$, $W^K$, and $W^V$.

• Let’s call the resultant matrices $Q^R$, $K^R$, and $V^R$. The dimension of each of these matrices is $d_{sequence} \times d_{model}$. In our case, it is $7\times 512$.

• There are a total of eight attention heads; therefore, each of these matrices is converted into eight subsets of size $7\times64$.

• Each subset is passed through the softmax function and is multiplied with its respective subset of the $V^R$ matrix, according to the following formulation:

  Here, $i$ represents a subset of each matrix bearing the dimension $7\times 64$ .

• The result from each head is combined into a matrix named $C$, with a resulting dimension of $7\times512$.

• The matrix $C$ is multiplied with a weight matrix $W^C$. This completes the process of a single multi-head attention block in the transformer model.

The process is illustrated below:

The purpose of the multi-head attention mechanism in models like ChatGPT is to enhance the capacity of the model to capture diverse patterns, relationships, and context within the input sequence. Instead of depending on a single attention mechanism, multi-head attention enables the model to focus on various parts of the input sequence by utilizing multiple sets of attention weights, each focusing on different aspects.

Let’s suppose we’re using two-head attention for the following sentence:

  “She poured milk from the jug into the glass until it was empty.”

We might expect the following visualization of the output. For the query word “it,” the first head (colored blue) focuses on the words “the jug,” while the second head (colored brown) focuses on the words “was empty.” Therefore, the ultimate context representation will center around the words “the,” “jug,” and “empty,” making it a more advanced representation than the conventional approach.

In simpler terms, ChatGPT’s attention mechanism is like a guiding light that helps it understand and respond coherently in conversations. This technology turns language complexities into something smart algorithms can handle.

4. The feed-forward network

In ChatGPT, the feed-forward network plays a pivotal role in refining the information gathered by the attention mechanism. This network operates independently for each position in the input sequence, applying a series of transformations to enhance the model’s understanding of complex relationships. Starting with a linear transformation, each position’s representation undergoes an activation function, typically ReLU, introducing nonlinearity. Subsequently, another linear transformation is applied, and layer normalization ensures stable learning. This position-wise feed-forward process is repeated across multiple layers and attention heads in ChatGPT’s architecture.

By employing feed-forward networks, ChatGPT enhances its ability to capture intricate patterns within the input sequence. The position-wise and layer-wise operations allow the model to adapt to diverse relationships and nuances, contributing to its overall effectiveness in comprehending and generating contextually relevant responses in natural language conversations. The feed-forward network is a crucial component in the transformer model, empowering ChatGPT to process and distill information from the attention mechanism, ultimately leading to more sophisticated language understanding and generation capabilities.

Reality surpasses simplicity

The reality of the attention mechanism used by ChatGPT for generating effective responses reflects the complexity of understanding and responding to human language. In the intricate landscape of real-world conversations, people convey meaning through context, nuances, and varied expressions. The attention mechanism in ChatGPT attempts to capture this complexity by dynamically focusing on different parts of the input sequence, mirroring the way humans pay attention to relevant details in a conversation. However, the real world is inherently multifaceted, involving cultural nuances, diverse language styles, and contextual subtleties that challenge any language model. While ChatGPT’s attention mechanism is a remarkable tool, it underscores the ongoing journey to navigate the intricacies of human communication, acknowledging that achieving a full understanding and generation of responses reflective of the richness of real-world conversations remains an evolving and complex task.

Next step

If you want to expand your knowledge of deep learning further, the following courses are an excellent starting point for you: