An overall lack of recognition for the invisible, arduous, & taken-for-granted data work in AI leads to poor data practices, resulting in data cascades (negative, downstream events)... “Everyone wants to do the model work, not the data work” 1/
storage.googleapis.com/pub-tools-publ…
storage.googleapis.com/pub-tools-publ…
Paradoxically, data is the most under-valued and
de-glamorised aspect of AI
--Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI by Nithya Sambasivan @shivanikapania Hannah Highfill @NaaShomeh @heuristicity @laroyo 2/
research.google/pubs/pub49953/
de-glamorised aspect of AI
--Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI by Nithya Sambasivan @shivanikapania Hannah Highfill @NaaShomeh @heuristicity @laroyo 2/
research.google/pubs/pub49953/
Data quality issues in AI are addressed with the wrong tools created for, and fitted to other tech problems—they are approached as a database problem, legal compliance issue, or licensing deal. 3/
“In real life, we never see clean data. Courses focus on models & tools but rarely teach about data cleaning & pipeline gaps.” CS curricula don't include training for dealing w domain-specific ‘dirty data’, documenting datasets, designing data collection, training raters,... 4/
ML data collection practices often conflict w/ existing workflows of domain experts. Data creation was added as extraneous work to on-the-ground partners (e.g., nurses, patrollers, farmers) who already had several responsibilities and were not adequately compensated. 5/
Missing metadata led practitioners to make assumptions, ultimately leading to costly discarding of
datasets or re-collecting data. Lack of metadata & collaborators changing schema w/out understanding context led to loss of 4 months of precious medical robotics data collection 6/
datasets or re-collecting data. Lack of metadata & collaborators changing schema w/out understanding context led to loss of 4 months of precious medical robotics data collection 6/
From goodness-of-fit to goodness-of-data:
Goodness-of-fit metrics, such as F1, Accuracy, AUC, do not tell us much about the fidelity and validity aspects of the data. Currently, there are no standardised metrics for characterising the goodness-of-data 7/
Goodness-of-fit metrics, such as F1, Accuracy, AUC, do not tell us much about the fidelity and validity aspects of the data. Currently, there are no standardised metrics for characterising the goodness-of-data 7/
We find drastic differences in data & compute in African countries & India, compared to USA... the Global South is viewed as a site for low-level data annotation work, an emerging market for extraction from ‘bottom billion’ data subjects, or a beneficiary of AI for social good 8/
• • •
Missing some Tweet in this thread? You can try to
force a refresh
