Many of our customers want to know how to choose a technology stack for solving problems with machine learning.
In this article, I summarize my thought process when suggesting a tech stack for ML. 🧵
In this article, I summarize my thought process when suggesting a tech stack for ML. 🧵
https://twitter.com/GoogleCloudTech/status/1411098610230386689
A key decision that you have to make for each ML problem is to decide whether to:
(1) buy a vendor's pre-built solution
(2) build your own
Make this decision based on whether you have access to more data than the vendor.
This is also a handy rule to choose between vendors.
(1) buy a vendor's pre-built solution
(2) build your own
Make this decision based on whether you have access to more data than the vendor.
This is also a handy rule to choose between vendors.
When building your own ML solution, avoid the temptation to build from scratch.
The best return on investment early in your projects is going to come from collecting more data (both more "rows" and more "columns", by breaking down data silos)
Use standard and/or low-code models
The best return on investment early in your projects is going to come from collecting more data (both more "rows" and more "columns", by breaking down data silos)
Use standard and/or low-code models
The tech stack depends on the type of ML application:
1. Predictive analytics
2. Unstructured data
3. Automation
4. Recommendations
I'll quickly summarize my recommendations for each, but read article linked from headline tweet for more details.
1. Predictive analytics
2. Unstructured data
3. Automation
4. Recommendations
I'll quickly summarize my recommendations for each, but read article linked from headline tweet for more details.
For predictive analytics, the key thing is to use a tech stack where you can keep growing your data and train ML models without data movement
Build an EDW.
Train BigQuery ML models
When improvements due to data size plateau, build Tensorflow/Keras models that read off BigQuery
Build an EDW.
Train BigQuery ML models
When improvements due to data size plateau, build Tensorflow/Keras models that read off BigQuery
For unstructured data, the ROI of AutoML is hard to beat for small to medium data sizes.
For large data sizes, use pre-built models that have already been written to use TPUs efficiently. Start with transfer learning, then do fine tuning, and train from scratch.
For large data sizes, use pre-built models that have already been written to use TPUs efficiently. Start with transfer learning, then do fine tuning, and train from scratch.
For automation, you will be training several models and orchestrating them.
Some of them will be pre-built (eg Document AI). Others will be low-code (eg BigQuery ML). Others will be no-code (eg Auto ML Video Intelligence).
You need a unified AI platform. Use Vertex AI pipelines
Some of them will be pre-built (eg Document AI). Others will be low-code (eg BigQuery ML). Others will be no-code (eg Auto ML Video Intelligence).
You need a unified AI platform. Use Vertex AI pipelines
For recommendations, you again need an enterprise data warehouse (EDW).
You need one well integrated with your transactional databases.
Use Datastream to do CDC into BigQuery.
Start with BigQuery ML. Move to Recommendations AI. Once improvements plateau, train from scratch.
You need one well integrated with your transactional databases.
Use Datastream to do CDC into BigQuery.
Start with BigQuery ML. Move to Recommendations AI. Once improvements plateau, train from scratch.
And as always, reach out to your Google Cloud account team if you want to talk through your options and brainstorm of what approach to start with.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
