ETHICS - AI Dynamics

Unbiased Algorithms in Differential Privacy Research

By

–

01 February 2023 14h55

BONUS TWEET Be sure to also check out this other paper posted at the same time, by @thesasho and Haohua Tang, also focused on unbiased algorithms in differential privacy. Despite similarities in the titles, the settings are mostly different. https://
arxiv.org/abs/2301.13850 9/8

→ View original post on X — @thegautamkamath

1 February 2023

Differential Privacy Limitations for Gaussian Distribution Estimation

By

@thegautamkamath

–

01 February 2023 14h55

But we also show that pure (epsilon, 0)-DP is hopeless, even for really simple classes like Gaussians, and that the delta in approx DP is needed to perform unbiased estimation. 7/n

→ View original post on X — @thegautamkamath

1 February 2023

Privacy-Bias-Variance Trilemma in Mean Estimators

By

@thegautamkamath

–

01 February 2023 14h55

Our main result: no. There is a *trilemma* between privacy, bias, and variance of a mean estimator: essentially, one can not simultaneously have strong privacy, low bias, and low variance. This shows the clip-and-noise algorithm is optimal. 5/n

→ View original post on X — @thegautamkamath

1 February 2023

Bias-Variance-Privacy Trilemma in Statistical Estimation

By

@thegautamkamath

–

01 February 2023 14h54

"A Bias-Variance-Privacy Trilemma for Statistical Estimation," with @argymouz
, Matthew Regehr, @vkerdos, @shortstein
, and @thejonullman
. https://
arxiv.org/abs/2301.13334 Private estimators MUST be biased! 1/n

→ View original post on X — @thegautamkamath

1 February 2023

Scientific Rigor in AI Research Communication and Paper Interpretation

By

@dotcsv

–

01 February 2023 11h46

Hoy en mezclando churras con merinas… Hay gente que quiere ver en mi divulgación cosas que no hay. Si dedico tiempo a explicaros el porqué un paper sobre X tema se está entendiendo mal, no es por un posicionamiento personal, sino porque me interesa esto se trate con rigor.

→ View original post on X — @dotcsv

1 February 2023

Stable Diffusion’s Limited Copying Capabilities Analysis

By

@dotcsv

–

01 February 2023 11h27

Si te parece una probabilidad representativa de algo, pues úsala para construir tus conclusiones. Pero por conectarlo con el contexto del primer tweet, ya que has querido unir ambas cosas. Lo que se muestra es que Stable Diffusion, copiar copiar… copia poco.

→ View original post on X — @dotcsv

1 February 2023

Training objectives: data reconstruction versus distribution learning

By

@dotcsv

–

01 February 2023 11h21

Como comenté, y de donde sale ese tweet, el objetivo de entrenar para "reconstruir una copia de los datos de entrenamientos" es apuntar a plagiar esos datos. Y no es lo mismo que el objetivo al que se apunta con estos modelos generativos: "aprender la distribución de los datos".

→ View original post on X — @dotcsv

1 February 2023

ChatGPT Demonstrates AI Power and Job Market Implications

By

@bernardmarr

–

01 February 2023 8h46

#ChatGPT is a tools that is enabling anyone to see the power of #artificial #intelligence and #machine #learning. It is also forcing people to think about the implications of such powerful #technology on #jobs.

→ View original post on X — @bernardmarr

1 February 2023

Co-Chief Scientist discusses content creation and source attribution in LLMs

By

@ai21labs

–

01 February 2023 8h44

Interview with our Co-Chief Scientist, @YoavLevine
: "When someone creates content, they want to be comfortable to put their name on it. And in certain use cases, the ability to connect the content to sources really facilitates this” https://
bdtechtalks.com/2023/01/30/ai2
1labs-llms-ralm/
… via @bdtechtalks

→ View original post on X — @ai21labs

1 February 2023