MIT CSAIL News ·
Zheng Dai and David Gifford discuss removing training data to test its effect on model outputs. James Grimmelmann says the paper gives reason to think attribution may fail for what he calls “interesting models.”. Read 3 viewpoints with supporting evidence and source links.
Lines show the reading structure. Select an idea to read its explanation and evidence.
Synthesis
Zheng Dai says that if removing a piece of training data leaves the model’s output unchanged, that piece of data did not affect the output.
Original excerpt
If you take away a piece of data and the output of the model doesn’t change, then that piece of data didn’t affect the output
Zheng Dai · Paragraph 4 — quotation 2
" ,"
Continue exploring
Causal attribution in AI outputs →David Gifford describes the paper’s method as deleting training inputs and all their influence. He says it enables efficient, large-scale deletion and shows that the results do not change.
Original excerpt
This paper introduces the first method that is absolute. You're actually deleting the inputs and deleting all influences of the inputs.
David Gifford · Paragraph 5 — quotation 3
"They really could not absolutely show that deleting individual things did not change the output. This is the first exact method for doing large-scale deletion efficiently and showing that the results don't change."
Continue exploring
Exact deletion method for influence testing →James Grimmelmann says the paper gives reason to think attribution will fail for “interesting models,” and that technologists and courts will need other ways to assess copying.
Original excerpt
this paper provides reason to think that attribution will fail for interesting models
James Grimmelmann · Paragraph 19 — quotation 3
“But . Instead, technologists and courts will need to resort to other methods for assessing copying.”
Continue exploring
Legal implications of attribution failure →Attributed passages with the context to verify them. Open the original text to check the source.
Original excerpt
this paper provides reason to think that attribution will fail for interesting models
“But . Instead, technologists and courts will need to resort to other methods for assessing copying.”
Original excerpt
If you take away a piece of data and the output of the model doesn’t change, then that piece of data didn’t affect the output
" ,"
Original excerpt
This paper introduces the first method that is absolute. You're actually deleting the inputs and deleting all influences of the inputs.
"They really could not absolutely show that deleting individual things did not change the output. This is the first exact method for doing large-scale deletion efficiently and showing that the results don't change."
These viewpoints are linked to their original sources. Paraphrases are labeled and are not verbatim quotes.
Open transcript or source material (opens in a new tab)Report an issueDavid Gifford on Exact deletion method for influence testing. Explore 1 viewpoint by topic, with evidence from 1 source.
James Grimmelmann on Legal implications of attribution failure. Explore 1 viewpoint by topic, with evidence from 1 source.
Zheng Dai on Causal attribution in AI outputs. Explore 1 viewpoint by topic, with evidence from 1 source.