Specialization, Redundancy, and Head Ablation

Research has found both specialized-looking and removable heads in particular trained models. Separate observations from causal interventions, define zeroing and pruning precisely, and account for head permutation and network compensation.

Multiple heads can form different attention patterns, but the architecture does not assign a job to each head. Specialization is a property to test in a trained model, not a promise made by the equation.

Separate Three Levels of Claim

Architectural possibility: heads use separate projections and can form different score and weight matrices.
Observed pattern: in a specified checkpoint and dataset, one head may repeatedly attend to a nearby position, rare token, or syntactic relation.
Causal importance: changing or removing that head changes a stated model output or evaluation measure under a controlled intervention.

A heat map can support the second level. It cannot establish the third without an intervention and a measured effect.

Head Zeroing, Pruning, and Retraining Are Different Experiments

Experiment	What changes	What it can test
zero one head for one input	one $Z^{(r)}$ becomes zero before $W_O$	local output sensitivity
remove a head at evaluation	one trained component is absent across a dataset	test-time dependence under the pruning rule
fine-tune after pruning	remaining parameters may adapt	recoverable dependence
train fewer heads from the start	architecture and optimization path change	whether another model can learn with fewer heads

These experiments need not give the same answer. A head that can be removed after training may still have influenced optimization. A model trained with fewer heads may learn a different division of work.

Empirical Evidence Is Setup-Bounded

Michel, Levy, and Neubig examined trained multi-head models and found that many heads could be removed at test time with little measured performance change in their evaluated settings; sensitivity varied across layers and attention types. This is evidence of redundancy in those models, not a universal head count rule.

Voita and colleagues studied a Transformer machine-translation encoder. In their experiments, some important heads showed consistent positional, syntactic, or rare-word patterns, while many heads could be pruned with little translation-quality loss under the reported procedure. The result supports the coexistence of specialization and redundancy in that setup.

Neither study establishes that every model has the same prunable fraction, that every interpretable pattern is causally necessary, or that a smaller model trained from scratch will match a pruned checkpoint.

Head Identity Can Move

Head indices can be permuted if the matching blocks of $W_O$ are permuted with them. During independent training runs, similar functions may therefore appear at different indices. Other heads may also compensate after pruning or fine-tuning.

A claim such as “head 7 detects syntax” should include at least:

the exact checkpoint and layer;
the inputs or dataset used to identify the pattern;
the measurement that defines detects;
an intervention or performance test if causal importance is claimed;
alternative explanations and uncertainty.

Q1. Interpret a pruning result

In one trained translation model, removing 10 of 12 heads at evaluation changes the reported score by less than the study's chosen tolerance. Which conclusion is supported?

Choose the supported claim

Those heads were not individually necessary for that measured result under the stated pruning setup

Every Transformer should always be built with only two heads

The removed heads had no influence at any point during training

Select one choice, then check.

Hint

Test-time removal does not reconstruct the model's training history.

Solution

The bounded conclusion is supported. The result does not determine an optimal universal head count or prove that the removed heads never affected optimization.

Record Observation, Intervention, and Metric Separately

Useful head analysis does not stop at naming a visual pattern. It states what was observed, what was changed, what output or metric was measured, and which claims remain open.

References

Paul Michel, Omer Levy, and Graham Neubig, Are Sixteen Heads Really Better than One? (2019). The study reports layer- and attention-type-dependent test-time head pruning results in its evaluated models.
Elena Voita et al., Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned (ACL 2019). The paper studies specialization and pruning in Transformer machine-translation encoders under a stated experimental procedure.