On the TechTO stage ×1
First seen on the TechTO stage in 2017. Every talk is searchable — ask the archive about Shiva ↗
In their words
we thought this is great, autism is really difficult to diagnose, we had a new model — and then we tried our model on other data sets and we got conflicting reports almost everywhere
we think actually that the pre-processing and feature selection play a far bigger role
Quick answers
What is Shiva Amiri known for?
Leading AI and data work at the life-sciences investor Pivotal Life Sciences, and before that running BioSymetrics, which applied machine learning to combined clinical, imaging and genetics data.
What was her TechTO talk about?
Fools Rush In, in July 2017 — an argument against treating deep learning as the whole job. She used a failed autism-diagnosis model of her own as the cautionary example.
What went wrong with the autism model?
A model built on 1,000 children's brain-imaging data looked convincing, then produced conflicting results on every other data set they tried. The team concluded the pre-processing and feature selection mattered more than the algorithm.