Multi-Head Attention Mechanism

The multi-head attention mechanism is a key component of the Transformer architecture, introduced in the seminal paper "Attention Is All You Need" by Vaswani et al. in 2017. It plays a crucial role in enhancing the ability of models to focus on different parts of an input sequence simultaneously, making it particularly effective for tasks such as machine translation, text generation and more.

Understanding Attention Mechanism

Before diving into multi-head attention, let’s first understand the standard self-attention mechanism, also known as scaled dot-product attention.

Given a set of input vectors, self-attention computes attention scores to determine how much focus each element in the sequence should have on the others. This is done using three key matrices:

• Query (Q): Represents the current word's relationship with others.
• Key (K): Represents the words that are being compared against.
• Value (V): Contains the actual word representations.

The self-attention is computed as:

What is Multi-Head Attention?

Multi-head attention extends self-attention by splitting the input into multiple heads, enabling the model to capture diverse relationships and patterns.

Instead of using a single set of matrices, the input embeddings are projected into multiple sets (heads), each with its own :

1. Linear Transformation: The input is projected into multiple smaller-dimensional subspaces using different weight matrices.
   
   where i denotes the head index.
2. Independent Attention Computation: Each head independently computes its own self-attention using the scaled dot-product formula.
3. Concatenation: The outputs from all heads are concatenated.
4. Final Linear Transformation: A final weight matrix is applied to transform the concatenated output into the desired dimension.

Mathematically, multi-head attention is expressed as:

where:

is a final weight matrix to project the concatenated output back into the model’s required dimensions.

Why Use Multiple Attention Heads?

Multi-head attention provides several advantages:

• Captures different relationships: Different heads attend to different aspects of the input.
• Improves learning efficiency: By operating in parallel, multiple heads allow for better learning of dependencies.
• Enhances robustness: The model doesn’t rely on a single attention pattern, reducing overfitting.

Multi-Head Attention in Transformers

Multi-head attention is used in several places within a Transformer model:

1. Encoder Self-Attention: This allows the encoder to learn contextual relationships within the input sequence. Each word (or token) in the input attends to every other word, helping the model to understand dependencies regardless of their distance in the sequence.

2. Decoder Self-Attention: In the decoder, self-attention ensures that each position in the output sequence can attend only to previous positions (with a masking mechanism), preventing the decoder from “seeing” future tokens during training. This helps in generating sequences in an autoregressive manner while focusing on relevant parts of what has been generated so far.

3. Encoder-Decoder Attention: This layer lets the decoder attend over the encoder's output. It helps the decoder to align and focus on the appropriate input tokens when generating each output token, enabling sequence-to-sequence tasks like translation.

Implementing Multi-head Attention using PyTorch

Step 1: Imports

Importing all necessary libraries for tensor manipulations and neural network building.

Step 2: Scaled Dot-Product Attention Function

This is the core of self-attention:

• Q, K, V are queries, keys, values derived from the same source in self-attention.
• It results in values i.e the weighted sum for each position and head.
• Softmax ensures the attention weights sum to 1.
• If masking, irrelevant positions (like future tokens or padding) get large negative values in logits, so after softmax attention there is 0.

Step 3: Multi-Head Attention Class

Every step mimics the original Transformer:

• Project to QKV,
• Reshape for multiple heads,
• Split into Q, K, V,
• Compute attention,
• Concatenate heads,
• Linear output.

4. Example: Run With Printouts

Output:

Notebook link : Multi Head Self Attention

Applications of Multi-Head Attention

Multi-head attention is widely used in various domains:

1. Natural Language Processing

• Machine translation (e.g., Google Translate)
• Text summarization
• Chatbots and conversational AI

2. Computer Vision: Vision Transformers (ViTs) for image recognition

3. Speech Processing: Speech-to-text models (e.g., Whisper by OpenAI)

The multi-head attention mechanism is one of the most powerful innovations in deep learning. By attending to multiple aspects of the input sequence in parallel, it enables better representation learning, enhanced contextual understanding and improved performance across NLP, vision and speech tasks.