I remember so many researchers who confidently pronounced “but words/sentences/paragraphs is not a vector!!” Here we are shoving *the entire internet* into a few vectors. How it started:
DATA
-
Differential Privacy: Correlated Noise in DP Algorithms Explained
By
–
Learning with #DifferentialPrivacy often involves adding noise into the learning process. At 3:20pm today, Krishna Pillutla will be at the #NeurIPS2023 Google booth to explain why it is beneficial in DP algorithms to add noise that is correlated across time. Drop by to hear more!
-
Google Introduces Croissant ML Dataset Format for Interoperability
By
–
Drop by the Google booth at 9:15am today where @laroyo
, Meg Risdal & Omar Benjelloun will introduce Croissant —in collaboration with @MLCommons & others—a format for ML datasets with the goal of making datasets interoperable across repos & platforms. -
Responsible AI: Data Quality and Cultural Diversity in ML
By
–
Don't miss @laroyo give the #NeurIPS2023 opening Keynote talk, “The Many Faces of Responsible AI”, in which she'll discuss the need for culturally-aware, society-centered research on the impact of data quality and diversity for training and evaluation of ML models. Happening now!
-
Overnight Model Training Run Produces NaN Values
By
–
gm Waking up to find out that my overnight model train run produced a bunch of NaNs.
-
Mass Surveillance Through Connected Devices: Privacy Under Threat
By
–
UGH. Please read, share, call. Fighting for privacy can be wearying, feel defeating, but we need to fight this latest attempted mass surveillance ratchet turn, working to make all of our devices, and the infra they rely on, into snitches. This is not a world we can accept.
-
Meta surveillance enables criminalization of healthcare seekers
By
–
And corp (+ gov) surveillance is playing a troubling role in enabling such. To take one example., Meta turned over FB messages bet. Jessica Burgess & her daughter. These were used to criminalize them for seeking banned healthcare. They're now sentenced to 2 years in prison. 4/
-
End-to-End Encryption: Privacy Protection Against Corporate Surveillance
By
–
E2ee provides vital protection. While no corp OR gov should have access to the vast surveillance data typical of tech cos, for now we can celebrate wins like this, which offer the safety of privacy in a world where daily activities and solidarities increasingly put us at risk. 5/
-
Corporate surveillance replaces privacy norm without public consent
By
–
In a short few decades, w/o consent or full public awareness, the norm of privacy has been overwritten by the norm of corp (& gov) surveillance, c/o the normalization of the surveillance business model in tech. And yes, Meta is an exemplar here. 2/
-
Predictive Modeling Challenges: Data Quality and Pundit Predictions
By
–
Making predictive models when you have lots of clean, well organized, high quality data can still be extremely challenging. It can be a very humbling experience. It’s incomparably more difficult when you have no such data. And that’s where most pundits operate.