MLOps Career Advice

"Models were the heroes whereas data was the necessary evil"

How did a guy who was born in prison, raised by his grandparents, went to one of the worst schools in New Jersey end up studying computer science at MIT and doing his Ph.D. in ML at Stanford?

Cliches we have heard. Seems like every week there is a new one coming out in this space. Data is the new oil, 80% of all ML projects make it into production, data scientists spend all their time cleaning data. Now that Andrew Ng has talked about the data-centric approach to doing ML, it's becoming cliché.

Luckily for us, we have Cody Coleman on the pod to actually talk about Data Centric AI. No Clichés. Actual value.

Check out the episode we recorded with him below and don't forget to like it for the algorithm's data feed!

Geospatial Data

What makes Geospatial Data different? What are some unique challenges when working with this type of data?

Anne Cocos, director of data science at Iggy talked to us about how the company was launched by a former Airbnb data scientist who was frustrated by not getting internal resources to be able to solve key geographical questions about website listings. Questions such as 'how far is this house from the beach?' turned out to be extremely complicated to understand when only analyzing the text from the listing descriptions.

Anne explained why what Iggy is doing isn't exactly a feature store or data as a service. If it can be called anything maybe, features as a service? Their goal is to empower a data scientist to instantly have clean geospatial features and data without the headache of having to do the ugly parts that go along with it.

State of AI Report is out now! 'Nuff said, go download it!

Community member, Mark Craddock has created an "awareness exam" and "foundational exam" around Wardley Maps. He was kind enough to create a promo code for the community to give the first 50 to sign up for the foundational exam 50% off. Use the code 'mlops2021'.

Community members André Godinho and Carlos Leyson resurected the bi-weekly reading group sessions! Last week we read the paper "On the Opportunities and Risks of Foundational Models". André had these insights to share with the rest of us:

How to test foundation models? Ribeiro et al. 2020 proposed an approach for this by taking inspiration from BDD. Basically the idea is to create several tests for language models like BERT. For example, imagine you are doing sentiment analysis with BERT, you can generate sample tests (i.e. sentences) where you change some words of a sentence (that should not change its sentiment) and see if the output changed.
What kind of solutions are there regarding explainability/interpretability for foundation models? @Kilian Kluge gave two interesting references about this issue:

https://github.com/pytorch/captum - I haven't seen this one in further detail, but from what I've understood it seems that you can use this tool from pytorch to understand which features are contributing to a model’s output according by chocking the model with input changed with noise -> compute gradients with respect to the predicted class
https://facctconference.org/ - a conference that deals with this sort of issues (fairness, XAI, etc)

Btw, the notion page is already updated with the questions that were discussed during the reading group :)

Vegan Options Available

Open source feature store time. I've been trying to do more hands on coding and presentations in the meetups these days in case you haven't noticed.

For this meetup we will discuss the key concepts underlying Feast, the open-source feature store. We will then do a short coding workshop to explore these concepts and explain the value of using a feature store.

Our guest on honor this week is Felix Wang. Felix is a software engineer at Tecton working on Feast. He recently graduated from Stanford, where he studied math and computer science.

If you haven't heard already, we have a public cal you can subscribe to. Otherwise, see you next week by clicking on the link below.