Edge ML integrations
I had a great time speaking at the Ai4 - Artificial Intelligence Conference on August 17, 2021. My talk was titled, "What is ML Scale? The Where and the When of ML Usage." It was a lot of fun to be able to chat with people about machine learning use cases out in the wild. You can watch a video of my talk below. During this talk I focused on the strategy behind both ML scale and MLOps decisions at scale. You can watch the talk via the vimeo link below.
You will quickly find that plenty of scholarly research exists on the idea of edge machine learning.[1] That is not to be confused with the idea of cutting edge machine learning which would be about describing the edge of what is possible vs. the idea of machine learning at the edge. If you consider for a moment how devices are set up and work, then you can sort of draw diagrams of what would be out on the edge of that connectivity. It could be all the internet of things (IoT) devices, your smartphone, or other things that end up getting counted in the ever growing pool of connected devices. Our friends over at IBM even have a page devoted to the joys of machine learning models being deployed at the edge.[2] You are using edge-deployed machine learning models on your Apple or Android based smartphone for photography or voice assistants right now. Somebody trained up a model to do something with photographs or to help drive relevant search and it was deployed on your smartphone without any real fanfare or celebration. This has been done to make things easier at the edge.
Your integrations at the edge are always going to be constrained by two distinct factors related to processing power and access to the internet. You cannot expect that you will have the bandwidth to egress your smartphone and work on a photograph in the cloud at all times. Our wireless network is not that great and most people don’t have endless bandwidth to support that type of endeavor. You would also have another problem related to latency where people just don’t like to wait for things to happen anymore. Long gone are the days where the internet booted up with a series of hums and beeps before things started to happen. A growing number of people using devices right now have never had to wait for a dialup connection to initialize. That reality means that to avoid waiting, latency, and otherwise perceived failure you have to really work to train and deploy machine learning models that achieve real results with the lowest possible processing power requirements and space needs.
Links and thoughts:
Check out this video from Google Cloud Tech, “Transformers, explained: Understand the model behind GPT, BERT, and T5”
I was reading this article called “The Pile: An 800GB Dataset of Diverse Text for Language Modeling” from the arXiv website from Gao et al., 2020. That diverse collection of data includes 825 gigabytes of content which functionally has been cleared of all sources and the authorship removed. This action has removed individuality from the language model in favor of generalization. Future models might end up going the other direction and favoring personality over generalization, but that might end up being more isolated based on what I’m seeing so far in terms of language modeling.
Check out Yannic Kilcher this week: “[ML News] Nvidia renders CEO | Jurassic-1 larger than GPT-3 | Tortured Phrases reveal Plagiarism”
Top 7 Tweets of the week:
Footnotes:
[1] A quick Google Scholar search for “edge machine learning”
https://scholar.google.com/scholar?q=edge+machine+learning&hl=en&as_sdt=0&as_vis=1&oi=scholart
[2] https://www.ibm.com/cloud/blog/models-deployed-at-the-edge
What’s next for The Lindahl Letter?
Week 32: Federating your ML models
Week 33: Where are AI investments coming from?
Week 34: Where are the main AI Labs? Google Brain, DeepMind, OpenAI
Week 35: Explainability in modern ML
Week 36: AIOps/MLOps: Consumption of AI Services vs. operations
I’ll try to keep the what’s next list forward looking with at least five weeks of posts in planning or review. If you enjoyed reading this content, then please take a moment and share it with a friend.