Skip to main content
Back to timeline
arXivSource publication:

MambaVision block ablation mostly measures positional fragility: with size-matched nulls, per-block content effects fall to a fraction of a point

Synopsis

In hybrid vision backbones such as MambaVision, which interleave Mamba state-space mixers with self-attention, removing one block and reading the accuracy drop has served as the measure of that block's contribution. Across all 50,000 ImageNet validation images, each ablation was paired with a size-matched null, the same perturbation with channels shuffled or signs flipped, and real and null damage were nearly indistinguishable in top-1 accuracy and true-class rank. The same test detects content effects in DeiT-Small and Swin-T, so single-block ablation in MambaVision mainly reflects how sensitive the residual stream is to where and how large a perturbation is, not what the block computes.

Source-provided article image: Size-Matched Nulls Reveal What Block Ablation Measures in MambaVision
arXiv

Interpretation

Removing a single mixer block in MambaVision lowers accuracy by almost as much as a same-size perturbation at the same place that carries none of the block's information. Block ablation is normally read as the size of that block's contribution, on the assumption that the damage comes from losing what the block computed. On all 50,000 ImageNet validation images, the pooled content effect over the twelve blocks of MambaVision-T is 0.016 percentage points, with a 90% interval of [0.049, 0.024] inside the pre-set equivalence margin of plus or minus 0.1 points; the same test finds significant content effects of up to 8.7 points in DeiT-Small and 1.25 points in Swin-T.

Scaled by damage, MambaVision's content share in accuracy is comparable to the hierarchical Swin-T and lower than both controls in true-class rank. The small absolute content effect partly reflects how little damage its blocks do, 0.34, 0.21 and 0.11 points on average, against 1.66 for Swin-T and 2.10 for DeiT-Small. The accuracy content share is 0.20 for MambaVision-T, 0.42 for -S and 0.25 for -B, against 0.24 for Swin-T and 0.91 for DeiT-Small; the rank shares are 0.06, 0.16 and 0.39 for the three MambaVision models, below Swin-T's matched stages at 0.94 and DeiT-Small at 0.97, with T and S also below Swin-T as a whole at 0.37.

Removing all blocks of one type compounds damage far beyond the sum of single effects, and size-matched nulls fail to reproduce this even when they keep the alignment between the blocks' perturbations. If blocks were independent, removing several would cost the sum of removing each. Removing all six Mamba or all six attention mixers of MambaVision-T costs 3.7 to 8.4 times the sum of single effects, while group nulls with a shared permutation that preserves alignment compound much less, and the real group ablation's content effect against the shared null is about 80% of its effect against the independent null.

Input-dependent selection is real but only shows up under joint freezing, and it lives mainly in the B and C projections. Freezing the scan parameters of one Mamba mixer at a time costs at most 0.23 points, which would suggest the model barely uses selection. Frozen in all six Mamba mixers at once, image-dependence alone accounts for 20% to 32% of the Mamba contribution, 29.1%, 32.1% and 28.2% across the three model sizes; the same freezing code drops copy accuracy in mamba-130m from 19.06% to zero.

Perspective

The conclusions rest on ImageNet-1k classification with the released checkpoints of MambaVision-T, -S and -B, with DeiT-Small and Swin-T as controls, all evaluated on the fixed 50,000 validation images. Within each MambaVision stage every Mamba mixer precedes every attention mixer, so a difference between mixer types cannot be separated from a difference in depth.

Only one checkpoint per model size has been released, so variation between training runs cannot be measured, and the equivalence margin is anchored to the spread of the null instead. A single permutation can mislead for individual blocks, such as L2B6, L2B7 and L3B0 in MambaVision-B and S0B0 in Swin-T, while pooled conclusions stayed stable. At 384 pixels, a setting outside the model's training regime, the selectivity share falls to 15.6%.