Data, Deployment, And Modeling

I just came across this year-old tweet by my former Udacity colleague, Mat Leonard.

By far the hardest part of machine learning is the dataset. Collecting, labeling, and cleaning the data. Unless you're chasing state-of-the-art results, training a model is a few lines of code and a bit of time.

In order of effort, IMO:
1. Data
2. Deploying
3. Model
— Mat Leonard (@MatDrinksTea) February 8, 2020

When I read that, I thought, that’s ironic, because so much of how we taught machine learning at Udacity focused on modeling.

And sure, enough, the next tweet in the thread:

Online courses, including my work at Udacity, focus on modeling. It's easiest to teach and the most interesting for students too. There were times I taught bits on data preprocessing, some students complained about it because they just wanted to learn modeling.
— Mat Leonard (@MatDrinksTea) February 8, 2020

I think a lot of labeling and cleaning is outsourced, either to specialty companies, or to specialty teams within a larger ML organization. Perhaps there’s an opportunity for ML engineer to learn more about data labeling and cleaning.

Mat, by the way, now leads the education team at OpenMined.

2 thoughts on “Data, Deployment, And Modeling”

I think it’s safe to say that the emphasis on models (in the ML “lore” writ large) contributes to a number of hiring managers’ inflated expectations and lack of appreciation for the extremely outsized role that data plays in the process. I speak from experience.

LikeLike

Pingback: Data, Deployment, And Models – Fifth Level

Share this:

Like this:

Related

2 thoughts on “Data, Deployment, And Modeling”

Leave a Reply Cancel reply