A neural network layer that classifies digits by comparison instead of multiplication mostly works - just not nearly as well as the thing it is trying to replace.
Researchers describe a "Soft Dominance Layer" that skips the standard affine transformation - the weighted sum and bias at the heart of nearly every neural network layer since the 1980s. Instead, each output unit compares input coordinates against a learnable reference vector and aggregates the results using a sigmoid relaxation, with a sharpness parameter called alpha controlling how close those comparisons get to a hard yes-or-no threshold. In single-run MNIST tests, the layer hit 90.61 percent accuracy without annealing and 91.73 percent with it, against 98.27 percent for a standard MLP baseline. The paper also notes the learned reference vectors show spatial structure, which the authors read as qualitative evidence the layer is learning something meaningful rather than noise.
The appeal here is not raw accuracy - it is readability. A threshold comparison is something you can inspect directly; a weighted sum of hundreds of inputs is not, which is a big part of why neural networks remain hard to audit. If a comparison-based primitive can approach affine-layer performance, it could open a more interpretable path for parts of a network where that tradeoff matters.
That is a big "if." These are single-run results on one toy dataset, with no repeated seeds and a seven-point accuracy gap still unexplained. Call it an interesting idea worth a second experiment, not a new building block.