Section 20.3 reports that batch normalisation was first explained by internal covariate shift, that this explanation has not held up, and that what survives is a measured effect: the scale is controlled and the optimisation surface is smoother. The field has no settled account of which mechanism deserves the credit, and the answer may depend on the architecture, the batch size and the activation.
Take a position on this question: what experiment would separate the candidate mechanisms, and what result would change your mind?
In your answer: name at least two candidate mechanisms; specify the control and the quantity you would measure; say what pattern of results would count for and against each mechanism; state one way the experiment could mislead you; and say what the chapter's advice to log gradient norms and change one thing at a time implies for your design.
Rubric. Tick each criterion your answer meets, then save the self-assessment. The platform does not grade it.
Worked solution
There is no single correct answer, so this is a model of the shape a complete answer takes, not a key. A level-3 answer fixes the architecture, the data, the optimizer and the number of steps before it compares anything. It names two candidates. The first is scale control: normalisation keeps each layer's output in a fixed range, so one learning rate has the same effective size at every depth. The second is smoothing: normalisation changes the loss surface, so a larger step still decreases the loss. A third candidate is batch noise, since the mean and variance of a small batch are themselves random.
The answer then chooses a measurement that separates them. If scale control is the mechanism, the ratio of the gradient norm in the first layer to the gradient norm in the last should stay near 1 as depth grows, and the benefit should shrink when the learning rate is tuned separately for each layer. If smoothing is the mechanism, the largest stable learning rate should rise with normalisation even when every layer's scale is already matched by a fixed rescaling, and the loss along the path should curve less sharply. If batch noise is the mechanism, changing the batch size while holding the normalisation fixed should change the benefit, and a batch of one, which has no spread, should lose it; layer normalisation, which standardises each example alone, is the control that keeps the scale effect while removing the batch statistic.
The answer states what would refute it: if the benefit survives with layer normalisation, fixed per-layer rescaling and batch size one, then neither batch noise nor a simple scale story explains it. It names a confound: batch normalisation adds two learned parameters per layer, so a comparison that does not count them, or does not give the baseline the same number of parameters, can attribute their effect to normalisation. It also notes that a change in batch size changes both the noise and the quality of the statistics, so the two cannot be separated by that knob alone. Finally, it costs the experiment: each configuration is a full training run, and the answer says how many runs and how many examples it needs before the comparison can separate the candidates.