TheSequence Edge covers:
+ML concept you should learn
+Review of an impactful research paper
+New ML framework or platform and how you can use it thesequence.substack.com/p/-edge10-feat…
6/6
• • •
Missing some Tweet in this thread? You can try to
force a refresh
The centralized nature of AI makes it difficult for startups to compete with the large tech incumbents that have access to:
+massive datasets
+virtually unlimited computing resources
+world-class research talent
Decentralized AI is the key
Thread⬇️
The research in decentralized ML is nothing new and can be traced back to the late 1970s
But the space has caught new momentum w/ blockchains and distributed ledger technologies
2/⬇️
However, blockchains are not the only technology trend influencing decentralized ML
Decentralized ML has benefited from:
+Blockchains
+Federated Learning
+Private ML
3/⬇️
TensorFlow Serving = the first mainstream model serving architecture in ML frameworks
+It serves ML models inside Google
+It is available in the cloud and via open-source
How it was created and how it works?
Thread⬇️
Deep dive into "TensorFlow-Serving: Flexible, High-Performance ML Serving" by @JeremiahHarmsen, @FangweiLi, @sukritiramesh, Christopher Olston, Noah Fiedel, Kiril Gorovoy, Li Lao, Vinu Rajashekhar, Jordan Soyke
2/⬇️
Paper outlined the architecture of a serving pipeline for @TensorFlow models
Capabilities of TensorFlow serving:
+model lifecycle management;
+experiments with multiple algorithms;
+efficient use of GPU resources
3/⬇️
Model serving goes a bit beyond deployment, given the unique nature of the lifecycle of ML programs.
ML models operate in a circular lifecycle, where phases such as training and optimization are continuously repeated.
2/⬇️
Some of the most important aspects of any model serving pipeline:
+API interface
+real-time vs. batch execution
+versioning
+A/B testing
+scalability
3/⬇️