long tail
In plain wordsThe huge number of rare things — uncommon words, unusual road situations, obscure facts — that each appear seldom but together make up a large share of what an AI system meets.
The many low-frequency categories or events of a skewed distribution, which individually occur rarely but collectively account for a substantial share of occurrences. Language is long-tailed — word frequencies roughly follow Zipf's lawZipf's law, n.: The pattern that word use is lopsided: in a big text the top word is roughly twice as common as the runner-up and ten times as common as the tenth, so a few words dominate and most are rare., so most distinct words are rare — and so are object categories in images, user requests to assistants and situations met by autonomous vehiclesautonomous vehicle, n.: A vehicle, usually a car, truck or shuttle, that can drive itself, sensing the road, deciding what to do and steering, braking and accelerating with no human driving.. Learning systems master the frequent 'head' and handle the tail poorly because each rare case supplies few examples: language models answer questions about facts mentioned in many pretraining documents far more accurately than about facts mentioned in few, and much of the residual risk of deployed systems lies in rare edge cases. Responses include targeted data collection, reweighting or resampling rare classes, retrieval-augmented generationretrieval-augmented generation, n.: Giving an AI an open-book exam: before answering, it looks up relevant passages in a set of documents and bases its reply on them, often citing where each point came from. for rare knowledge, synthetic datasynthetic data, n.: Data created by computer programs or AI models rather than collected from the real world — such as practice math problems written by a model, or simulated street scenes for self-driving cars. and simulation for rare events, and evaluation sets designed around the tail.
The model handles common requests well; the long tail of odd ones is where it fails.
Origin
Long-tailed distributions are a long-standing topic in statistics; the phrase entered general use through Chris Anderson's October 2004 Wired article 'The Long Tail' and his 2006 book, about online retailers profiting from niche products. In machine learning, 'long-tail recognition' and 'long-tail knowledge' became research labels in the 2010s and 2020s.
Usage note
In business usage the long tail is an opportunity (many niche sales); in AI it is usually a problem (many rare failures). model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy. erodes the tail first.
Related
- Zipf's lawZipf's law, n.: The pattern that word use is lopsided: in a big text the top word is roughly twice as common as the runner-up and ten times as common as the tenth, so a few words dominate and most are rare.
- class imbalanceclass imbalance, n.: When one category is far rarer than the others in the data — say, one fraud case per thousand transactions — so a model can look accurate while missing nearly every rare case that matters.
- out-of-distributionout-of-distribution, adj.: Describes inputs that differ from anything an AI system saw in training — like a model trained on daytime photos facing a night scene — where its answers become less reliable.
- power lawpower law, n.: A rule linking two quantities in which scaling one up by a given factor always changes the other by the same multiple, wherever you start. Plotted on logarithmic axes, it is a straight line.
- model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy.
- sampling biassampling bias, n.: A distortion that arises when the data collected over-represents some kinds of people or situations and under-represents others, so conclusions drawn from it don't hold for everyone.
Sources
- Kandpal, N., Deng, H., Roberts, A., Wallace, E. & Raffel, C. (2023). Large Language Models Struggle to Learn Long-Tail Knowledge. ICML 2023. arXiv:2211.08411. arxiv.org/abs/2211.08411
- Anderson, C. (2004). The Long Tail. Wired 12(10).
Mentioned in
class imbalanceclass imbalance, n.: When one category is far rarer than the others in the data — say, one fraud case per thousand transactions — so a model can look accurate while missing nearly every rare case that matters., model collapsemodel collapse, n.: The gradual decline of AI models trained on the output of earlier AI models: rare details vanish first, then quality and variety erode, like a photocopy of a photocopy., power lawpower law, n.: A rule linking two quantities in which scaling one up by a given factor always changes the other by the same multiple, wherever you start. Plotted on logarithmic axes, it is a straight line., sampling biassampling bias, n.: A distortion that arises when the data collected over-represents some kinds of people or situations and under-represents others, so conclusions drawn from it don't hold for everyone., Zipf's lawZipf's law, n.: The pattern that word use is lopsided: in a big text the top word is roughly twice as common as the runner-up and ten times as common as the tenth, so a few words dominate and most are rare.