LIT-731: SmoothGrad: removing noise by adding noise

Smilkov, Thorat, Kim, Viégas and Wattenberg (2017) — ARXIV-1706.03825

Key takeaways

• The method is one line. Average the vanilla gradient over n copies of the input perturbed with N(0, σ²). Two hyper-parameters: the noise level σ, reported as a fraction σ/(x_max − x_min), and the sample count n.

• The numbers, and how they were chosen. 10–20% noise “seems to balance the sharpness of sensitivity map and maintain the structure of the original image”, and n > 50 shows “little apparent change” — diminishing returns at fifty. Inception on ImageNet, with a matching MNIST figure, and the paper adds that “the ideal noise level depends on the input”.

• The evidence is qualitative throughout, and the paper states why. “Since quantitative evaluation of a map remains an unsolved problem, we again focus on qualitative evaluation.” Both of its assessments — visual coherence and discriminativity — are judged by eye, the latter by differencing two classes’ maps onto a diverging colour scale.

• It asked a question the record answered yesterday. “It remains an open question to understand which properties affect the discriminativity of a given method — e.g. understanding why Guided BackProp seems to show the weakest discriminativity.” LIT-730 answers it: guided backprop is doing partial image recovery, which is class-insensitive by construction.

• §3.1 is a visualization-practices section and it is the most interesting part. Three lessons, each of which the record has since rebuilt from a later paper:

  §3.1 lesson, 2017	what the record now holds
  ”There is considerable ambiguity in how to convert signed values to colors” — and whether absolute value helps is dataset-dependent (MNIST digits are always white, so signed gradients mean something; on ImageNet “taking the absolute value of the gradient produced clearer pictures”)	SOTA-430 step 3, where the signed-versus-absolute choice flips a randomization verdict — three groups, three model families
  multiplying by the input “does tend to produce visually simpler and sharper images, although it can be unclear how much of this can be attributed to sharpness in the original image itself … a black/white edge in the input can lead to an edge-like structure on the final visualization even if the underlying sensitivity map has no edges”	THEORY-113, the input-multiplier account
  cap at the 99th percentile, because a few pixels with far-above-average gradients “have the potential to throw off color scales completely” and “without this post-processing step, maps may end up almost entirely black”	nothing — a display artefact that can masquerade as a finding

• And a structural objection to gradient⊙input, also 2017. “Pixels with values of 0 will never show up on the sensitivity map. For example, if we encode black as 0, the image of a classifier that correctly predicts a black ball on a white background will never highlight the black ball.” The paper still shows results both ways, and gives the case for multiplying: in a linear y = Wx, x_i w_i is what x_i contributes.

Standing in the anthology

Eight held documents named this paper and none held it. It is in SOTA-430’s pass row, in SOTA-435 step 3’s recommendation, in LIT-725 twice (Proposition 2.1’s infidelity optimum is “a smoothing operation reminiscent of SmoothGrad on Integrated Gradients”, and Theorem 4.1’s meta-algorithm “encompasses Smooth-Grad”), in LIT-713 and NOTE-365 as a method that changes under both randomization tests, and in LIT-729, THEORY-113 and THEORY-114 as a gradient method outside both mechanism accounts.

So it is the one method in this cluster that nothing faults — and the one with the least evidence behind it. It passes both randomization tests (LIT-713), it does not carry an input multiplier (THEORY-113), it does not converge to rank 1 (THEORY-114), and it is not doing image recovery (LIT-730 leaves the saliency family alone). Its own paper offers no quantitative result at all.

That is not a contradiction in the record and it is worth being precise about why. SOTA-435 recommends restrained smoothing on LIT-725’s evidence — Theorem 4.1, which is proved, and a measured improvement in infidelity and max-sensitivity across three datasets — not on this paper’s figures. The practice happens to be sourced to the paper that supplied numbers rather than to the paper that supplied the method, which is the right way round and was not deliberate.

Its §3.1 changes the provenance of THEORY-113. That account’s history in the record ran: LIT-713 (2018) suspects the input multiplier and says it does not measure it; LIT-725 (2021) measures it. The edge-injection observation is here in 2017, informally, in a section about choosing colour scales — which is why nobody cites it for mechanism and why the record rebuilt the account from the 2021 measurement. Recorded on THEORY-113 v4.

No practice is filed and no theory. The parameters belong on SOTA-435, which already recommends the smoothing, rather than in a second practice about the same technique — the call LIT-728 got for occlusion sensitivity. And the paper offers no mechanism: it does not say why averaging over noise sharpens a map, beyond the observation that the gradient fluctuates sharply at small scales.

Limitations

• Every claim is qualitative, by the authors’ own statement. The parameters were selected by looking at pictures, so 10–20% and n = 50 are defaults with no measured objective behind them. SOTA-435 v2 carries them with that provenance attached.
• One architecture for the headline figures (Inception on ImageNet), plus MNIST in the appendix. The paper’s own closing question is “whether the de-noising techniques described here generalize to other network architectures and tasks”.
• The discriminativity test is two objects and a colour difference. No ground truth, no metric, and the paper says systematic measurement “could also be valuable”.
• It is the method SOTA-430 would forbid validating the way this paper validates it. That practice’s negative half is “do not accept a method because its maps look like the object”, and this paper accepts one on exactly that basis — reasonably, in 2017, when it says the quantitative problem was unsolved. The record’s own position is that it was solved enough by 2019 (LIT-725) and that is where the recommendation’s support comes from.