{"id":258,"identifier":"LOR-00258","url":"https://lifeofresearch.com/preprints/258","title":"A Near Neighbour Is Not a Native Feature: What Nearest-Neighbour Novelty Detection Owes to the Space It Measures In, Why the Case For and Against Pretrained Features Turns on Benchmark and Baseline More Than on Where the Features Came From, and the Comparison That Would Separate Scoring Rule From Representation","abstract":"(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.\nLife of Research\nA novelty detector is trained on normal data and asked to flag whatever is unlike it. Since 2020 two positions have competed over how such a detector should be built. One holds that the decisive ingredient is a feature space borrowed from a network pretrained on a large external corpus, after which a nearest-neighbour distance is enough, and that self-supervised or reconstruction-based detectors trained only on the normal data are outclassed. The other holds that detectors trained in-domain, such as autoencoders whose novelty score combines the reconstructions of an input's nearest latent neighbours with the distance to them, are competitive without any pretrained network and transfer to domains, such as radio-telescope spectrograms, that look nothing like natural photographs. In this analysis the author reviews and synthesises published results to show that the two positions are answering a question neither benchmark design isolates. A published comparison of two detectors almost always changes at least three things at once: the space in which distance is measured, the rule that turns distance into a score, and the kind of novelty the benchmark plants, while a fourth factor, the evaluation (metric, baseline implementation, hyperparameter selection), decides which of the three gets the credit. On one industrial benchmark, published image-level results for detectors that all use ImageNet features and a distance or density score range from 75.0 to 99.1 AUROC, a spread several times larger than the 3.3-point margin by which the nearest-latent-neighbour autoencoder beats its own re-implementation of the pretrained baseline. On one natural-image benchmark, supervised ImageNet features with a simple density score are reported at 80.0, 86.2 and 92.5 AUROC by three groups. Off the pretraining distribution the verdict flips with the dataset: in one seven-dataset medical study a plain autoencoder beats frozen ImageNet features on two datasets and loses by up to 36 points elsewhere, and an autoencoder whose residual is measured in pretrained feature space beats both on most. The paper consolidates nineteen published results in a table and a figure, states a procedure for crediting a novelty-detection gain to a representation's origin rather than to the scoring rule or the evaluation, and names the studies that would settle what remains open.","authors":[{"name":"Pranay Mahendrakar","affiliation":"SONYTECH","orcid":"0009-0003-7224-029X"}],"type":"preprint","license":"CC-BY-4.0","keywords":["novelty detection","anomaly detection","nearest neighbours","autoencoders","pretrained representations","transfer learning","evaluation methodology","model selection"],"publication_date":"2026-11-21","version":1,"views":3,"downloads":3,"created_at":"2026-11-21 11:16:30","doi":"10.5281/zenodo.24408406","source":"Zenodo","source_url":"https://zenodo.org/records/24408406","files":[{"id":282,"name":"a-near-neighbour-is-not-a-native-feature.pdf","size":416288,"mime":"application/pdf","downloads":3,"download_url":"https://lifeofresearch.com/preprints/258/files/282/a-near-neighbour-is-not-a-native-feature.pdf"}]}