Skip to content
Last Dictionary
Plane
Interpretability & SteeringTraining & Post-trainingLanguage Models & Transformers

MEMIT

abbreviationAlso: Mass-Editing Memory in a Transformer

In plain wordsA method for rewriting thousands of facts stored in a language model at once, by spreading small, calculated weight changes across several of its layers.

  1. Mass-Editing Memory in a Transformer: an extension of ROMEROME, abbr.: A method for rewriting a single fact stored in a language model — say, moving the Eiffel Tower to Rome — by making one small, precisely calculated change to one layer's weights. that inserts many factual associations into a language model simultaneously. Rather than updating one MLP layer per fact, MEMIT spreads the change for a batch of edits across a range of middle-layer MLP modules that causal tracingcausal tracing, n.: A test that blurs the words of a factual question with noise and then restores one internal part of the model at a time, to find where in the model the fact gets recalled. identifies as mediating factual recall, solving a least-squares problem at each layer so that the new key–value pairs are stored while the outputs for a large sample of existing keys are preserved. Applied to GPT-J (6 billion parameters) and GPT-NeoX (20 billion), it scaled to thousands of simultaneous edits with better generalization and specificity than earlier editing methods at that scale. Like other locate-then-edit methods it can damage unrelated capabilities when used very heavily or sequentially, and edited facts often fail to carry over to multi-step inferences.

    We used MEMIT to update ten thousand outdated company facts in one pass.

Origin

Introduced by Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov and David Bau in 'Mass-Editing Memory in a Transformer' (October 2022; ICLR 2023).

Related

See also
  • ROMEROME, abbr.: A method for rewriting a single fact stored in a language model — say, moving the Eiffel Tower to Rome — by making one small, precisely calculated change to one layer's weights.
  • model editingmodel editing, n.: Changing one specific fact or behavior inside a trained AI model — for example, who leads a country — by directly adjusting a small part of it, without retraining the whole thing.
  • causal tracingcausal tracing, n.: A test that blurs the words of a factual question with noise and then restores one internal part of the model at a time, to find where in the model the fact gets recalled.
  • parametric knowledgeparametric knowledge, n.: What an AI model 'knows' from training, stored in the numbers of its network rather than looked up in documents — like facts you have memorized rather than ones you check in a book.
  • machine unlearningmachine unlearning, n.: Erasing specific things a trained AI model has learned, such as dangerous lab techniques, a copyrighted novel or one person's records, far more cheaply than training the model all over again.

Sources

  1. Meng, K., Sen Sharma, A., Andonian, A., Belinkov, Y. & Bau, D. (2023). Mass-Editing Memory in a Transformer. ICLR 2023. arXiv:2210.07229. arxiv.org/abs/2210.07229

Mentioned in

model editingmodel editing, n.: Changing one specific fact or behavior inside a trained AI model — for example, who leads a country — by directly adjusting a small part of it, without retraining the whole thing., ROMEROME, abbr.: A method for rewriting a single fact stored in a language model — say, moving the Eiffel Tower to Rome — by making one small, precisely calculated change to one layer's weights.