The Data Driven Podcast was joined by Tristan Rouillard, a co-Founder of Hasty.ai. Hasty.ai supports vision AI practitioners and their evolving needs by developing best-in-class vision AI tools that are supported by a community of machine learning engineers, data scientists, and software developers.
Some of the topics that Tristan and I discussed:
-
The genesis of the idea for Hasty came from the founders’ own experiences
-
Manual image labeling and just how time-consuming that is
-
The idea was NOT to change the technology but to change the process
-
The necessity to bring the neural networks into the process earlier on
-
The importance of having a clean data set
-
With Hasty, one is not looking for mistakes in the dataset, but fixing the ones the neural network has found
-
There is no substitute for a ground-truth data set
-
When you are building a product, you are looking for ‘love’ and ‘wow’
-
Hasty is changing a process and that is the hardest thing to change
-
Bringing agile methodology to machine learning
-
How sometimes the biggest competition is the status quo
-
Taking domain-specific assets and leveraging them in a good way
-
Data ownership and one of Hasty’s USPs
-
It is likely that the training data is the differentiator, not the neural network
-
The nuance and complexity of image recognition is yet to be understood
-
The need for data documentation
-
Hasty is teaching machines how to identify and see the world the same way we do
-
Everybody that joins Hasty has to label a data asset
The quote from Alan Nichol, a co-Founder and the CTO of Rasa, came from this LinkedIn post.
(A shout-out to the incredible Head of Operations at Hasty.ai, Lisa Wantig, for helping to make this possible.)