Analysis of What Different Heads Learn

The introduction of multiple attention heads raises a significant question: Do these parallel attention mechanisms actually learn different things? If each head simply learned the same patterns, the added computational complexity would offer little advantage over a single, larger attention head. Fortunately, empirical evidence and analysis suggest that different heads often specialize, learning to focus on distinct types of relationships within the input sequence.

The core mechanism enabling this potential specialization lies in the independent linear projections applied to the Queries ($Q$), Keys ($K$), and Values ($V$) for each head. Recall the projection for head $i$: $Q_i = Q W^Q_i$ $K_i = K W^K_i$ $V_i = V W^V_i$ where $W^Q_i \in \mathbb{R}^{d_{model} \times d_k}$, $W^K_i \in \mathbb{R}^{d_{model} \times d_k}$, and $W^V_i \in \mathbb{R}^{d_{model} \times d_v}$ are the learnable weight matrices for head $i$. Since these matrices are initialized independently and updated via backpropagation, each head has the capacity to project the input embeddings into a subspace where a particular kind of relationship is more apparent or useful for the model's objective.

Observing Head Specialization

Understanding precisely what each head learns is an active area of research, often referred to as "interpretability." A common technique involves visualizing the attention weights produced by different heads for given input sequences. By examining which tokens attend strongly to which other tokens within a specific head, we can infer the patterns it prioritizes.

For instance, consider the sentence: "The quick brown fox jumps over the lazy dog." We might observe patterns like:

• Head 1 (Local Context): Attention weights might be highest between adjacent words, acting similarly to a bigram model. "quick" might attend strongly to "The" and "brown".
• Head 2 (Syntactic Dependency): Attention might focus on syntactically related words, even if they are distant. "jumps" might attend strongly to its subject "fox".
• Head 3 (Specific Tokens): A head might consistently attend strongly to punctuation or special tokens if they carry significant information for the task.
• Head 4 (Content Similarity): Attention could link words with related meanings or roles, like "fox" and "dog".

Illustration of how different heads might attend to different relationships in the sentence "The quick brown fox jumps over the lazy dog". Head 1 (blue, dashed) focuses locally, Head 2 (pink, solid) connects verb to subject, and Head 4 (green, dotted) connects related nouns.

Visualizations often use heatmaps where rows and columns represent token positions, and the color intensity indicates the attention score $softmax(\frac{Q_i K_i^T}{\sqrt{d_k}})$. Different heads yield distinct heatmaps, highlighting their varied focus.

More rigorous analysis involves "probing." This means training simple linear classifiers or other small models on the output representations ($Z_i = Attention(Q_i, K_i, V_i)$) of individual heads to see how well they can predict specific linguistic properties (e.g., part-of-speech tags, syntactic dependencies). Success in predicting a certain property suggests the head encodes information relevant to it.

Examples of Learned Patterns

Studies analyzing trained Transformers have identified several common types of specialization among heads:

1. Positional/Local Focus: Some heads learn to attend primarily to tokens within a small window around the current token, effectively capturing local context. This can resemble the behavior of convolutional networks or n-gram models.
2. Syntactic Dependencies: Certain heads become adept at tracking grammatical relationships. They might learn to connect verbs to their subjects or objects, adjectives to the nouns they modify, or prepositions to their objects, sometimes across long distances.
3. Coreference/Related Entities: Heads may link mentions of the same entity within a text or connect semantically related concepts.
4. Attending to Delimiters/Special Tokens: In models using special tokens (like [CLS], [SEP]), some heads often develop a strong focus on these tokens, potentially using them as aggregation points for sequence-level information.
5. Rare Word Handling: Some heads might specialize in attending from rare words to more common, contextually informative words, helping the model understand infrequent terms.
6. Identity/Copying (Less Common): Occasionally, a head might learn to attend strongly to the current token's own position, effectively copying its representation.

Benefits of Head Diversity

The ability of different heads to specialize provides several advantages:

• Capturing Diverse Information: Sequences contain multiple layers of information (syntax, semantics, position). Multi-head attention allows the model to capture these different facets simultaneously without forcing a single mechanism to average potentially conflicting signals.
• Richer Representations: By concatenating the outputs of these specialized heads ($Z = Concat(Z_1, ..., Z_h)W^O$), the model constructs a richer, multi-faceted representation of each token, incorporating insights from various relational perspectives.
• Improved Model Capacity: The parallel subspaces allow the model to express more complex functions and dependencies than a single attention mechanism with the same total dimensionality.

Considerations and Caveats

While the concept of head specialization is appealing and supported by evidence, some points require consideration:

• Interpretability is Challenging: Assigning a single, clear function to each head is often difficult. Heads might perform multiple roles, or their behavior might be complex and context-dependent. Visualization provides clues but not definitive answers.
• Redundancy: Not all heads necessarily learn unique patterns. Some degree of redundancy might exist, or certain heads might contribute less significantly to the final output. Pruning less important heads is an area of research for model compression.
• Layer Dependence: The patterns learned by heads can vary depending on their depth within the Transformer stack. Heads in lower layers might focus more on local syntax, while heads in higher layers might capture more complex semantic or long-range dependencies.

In summary, the multi-head structure is not merely about parallel computation; it's a design that encourages functional specialization. By allowing different heads to attend to information in different representation subspaces, the model can integrate diverse relational patterns, leading to more effective sequence representations. The final linear projection ($W^O$) learns how to best combine these specialized perspectives for downstream processing by the feed-forward network and subsequent layers.