Machine Learning Up to Date #10
Here's ML UTD #10 from the LifeWithData blog! We help you separate the signal from the noise in today's hectic front lines of software engineering and machine learning.
LifeWithData strives to deliver curated machine learning & software engineering updates that point the reader to key developments without superfluous details. This enables frequent, concise updates across the industry without information overload.
Application
- Google Colab Tips & Tricks
- Realtime Object Tracking at the Edge
- AWS Copilot as Your CLI Helper for ECS
Theory
Google Colab Tips & Tricks β
Amit Chaudhary summarized some of the more useful aspects of Google's Colab notebook environment on his personal blog. Jupyter is great! There is also much more than it, though π. Between the Jupyter fundamentals and the GPU/TPU access provided, Colab is a fantastic way to democratize data science to the masses.
... keep reading
The Rundown
Realtime Object Tracking at the Edge β
Stephen Cass from [IEEE Spectrum](https://spectrum.ieee.org/) wrote up his experience using Nvidia's $100 [Jetson Nano](https://developer.nvidia.com/embedded/jetson-nano-developer-kit) development kit. The kit was re-released this March with several upgrades, and is competition to Google's [Coral Dev AI](https://spectrum.ieee.org/geek-life/hands-on/the-coral-dev-board-takes-googles-ai-to-the-edge) kit. Long-time embedded tinkerers, relax: both support Raspberry Pi camera modules π€. Have you used either of these? How was your experience? [Let me know](https://lifewithdata.org/contact).
... keep reading
The Rundown
AWS Copilot as Your CLI Helper for ECS β
Amazon's Elastic Container Service (ECS) has long been overshadowed by the almighty Kubernetes (and EKS, it's hosted version) for container orchestration in its cloud platform. However, ECS has found a solid supporter base for smaller-scale applications that have only one or a few containers. With the new copilot CLI, that process got a lot easier. I'm working up a demo project of deploying an application on ECS with FastAPI, more to come soon!
... keep reading
The Rundown
A Survey of Machine Learning Tooling in Industry β
Chip Huyen gathered a list of [over 200 tools](https://docs.google.com/spreadsheets/d/1OV0cMh2lmXMU9bK8qv1Kk0oWdc_Odmu2K5sOULS9hHQ/edit?usp=sharing) used in the machine learning tooling space and analyzed how this space has been developing over the past 12 years. Analyses like these are exactly what is needed to provide some more coordinated effort in the nascent space. Once we can consolidate onto a few tools and platforms that accomplish our typical workflows, we can iterate much faster.
... keep reading
The Rundown
On Data Science vs Machine Learning Engineering β
The distinction between data science and machine learning engineering is subtle, but important when it comes to software platform development. Platforms, as opposed to experiments and one-offs, must provide scale and flexibility to a wide range of tasks. Those that don't are...poor platforms. Caleb Kaiser reflects on how [Cortex](https://github.com/cortexlabs/cortex) aimed directly for machine learning engineering as it built its open-source ML platform, which has caught on very nicely in the cloud ML space. After reading his well-written article, check out some of their example application deploys to see how smooth it is!
... keep reading
The Rundown
Using Graph Neural Networks to Write Equations β
Graph Neural Networks (GNN) have been increasing in adoption the past few years, especially in applications involving complex node interaction (read: physics). Miles Cranmer from Princeton published some riveting research which combines symbolic regression and deep learning to learn equation relationships among particles in a simulator.
... keep reading
Shortcut Learning in Deep Neural Networks β
Dr. JΓΆrn-Henrik Jacobsen from the University of Toronto published a paper and accompanying article detailing the class of phenomena known as "shortcut learning" in which neural networks learn to generalize in ways that are unintended by the human designer. This is the cause of the public outcries for fairness in machine learning model predictions as well as the confusion when otherwise super-human models are embarrassingly fooled by nothing more than Gaussian perturbation. There are several different types of ways in which this occurs, and this sub-field of research will continue to identify and address them. Time to plug those gaps!
... keep reading