Edge AI for Beginners

18 hours ago (github.com)

It seems this is focused on on-device computation - as distinct from, say, Cloudflare's definition of the "edge" as a smart CDN with an ability to run arbitrary code and AI models in geographically distributed data centers (https://workers.cloudflare.com/).

Per Microsoft's definition in https://github.com/microsoft/edgeai-for-beginners/blob/main/...:

> EdgeAI represents a paradigm shift in artificial intelligence deployment, bringing AI capabilities directly to edge devices rather than relying solely on cloud-based processing. This approach enables AI models to run locally on devices with limited computational resources, providing real-time inference capabilities without requiring constant internet connectivity.

(This isn't necessarily just Microsoft's definition - https://www.redhat.com/en/topics/edge-computing/what-is-edge... from 2023 defines edge computing as on-device as well, and is cited in https://en.wikipedia.org/wiki/Edge_computing#cite_note-35)

I suppose that the definition "edge is anything except a central data center" is consistent between these two approaches, and there's overlap in needing reliable ways to deploy code to less-trusted/less-centrally-controlled environments... but it certainly muddies the techniques involved.

At this rate of term overloading, the next thing you know we'll be using the word "edgy" to describe teenagers or something...

  • I work at an industrial plant, we use "edge" to refer to something inside the production network.

    As an example the control system network is air-gapped so to use ML for instrument control or similar the model needs to run on some type of "edge" compute device inside the production network all of the inferencing would need to happen locally (i.e. not in the cloud).

  • Yeah, Cloudflare is in the minority with their definition of "edge."

  • In GPU compute land, "edge" means on the consumer device. The latency of delivery is negligible in comparison to the wall clock compute demands, so it doesn't make much sense to park your GPUs near the consumer.

    IoT is "edge".

    The only place I've seen "edge" used otherwise is in delivery of large files, e.g. ISP-colocated video delivery.

  • maybe a decent definition could be compute as close to the user latency-wise as practically possible while having full access to the necessary data.

    For certain things this will be able to go as far as the device if you're only ever operating on data the user fully owns, other things will need data centers still but just decentralised and closer to the user via fancier architectures ala the Cloudflare model.

This is far from what I expected. There is not much related to quantization, pruning, common architectures, precision or benchmarking. For those interested in this topic, I would recommend content from MIT HAN Lab.

I remember when we bought and installed, among the first in the world, the AWS Outpost, sold as an "edge" (of in between cloud and on prem) infrastructure product. Same term has been previously (ab)used also in the security space, at - again - the confluence between cloud and on-prem. And then - yet one more time - the "edge" was a closer data center for localized delivered cloud services.

Isn’t edge AI just a way to deploy AI to meet product requirements? What is special about this course? Is Microsoft trying to sell this as a service? If so what is the revenue model and hardware used?

One of the most common uses for edge AI not listed in this course is computer vision. You similarly want real-time inference for processing video. Another open source project that makes it easy to use SOTA vision models on the edge is inference: https://github.com/roboflow/inference

I would say this is a poor beginners guide for quantization/compression, it's mostly an API guide for tf/keras quantization APIs it doesn't tell the beginner why or when or which layers (and why) they should apply it to.

But the modules that compare the different model families are quite good. As are the remaining modules that are "How to deploy to $platform 101", including microsoft's, of course ;)

Not that I have a better resource at hand for quantization/compression _for beginners_, and I am probably a bad judge for how beginner friendly Song Han's TinyML course was...

What are the best Small Language Models (SLMs) these days?

  • Best is very subjective depends what you want it to do and if you want to fine tune and how big you consider small

    • Let me ask the same with: - runs on a laptop CPU - decide if a long article is relevant to a specified topic. Maybe even a summary of the article or picking the interesting part as specified in prompt instructions. - no fine tuning please.

      Thank you for any response!

It's funny that they used AI to translate into other languages, because the Arabic cover image is just gibberish.

They are really embracing ai! I can feel them all around even. Above me. Below me.

  • given how bad their software has been historically

    imagine how much worse it will be soon, given everything they seem to be outputting now is entirely generated slop

The very first sentence:

> Welcome to EdgeAI for Beginners – your comprehensive...

Em dash and the word "comprehensive", nearly 100% proof the document was written by AI.

I use AI daily for my job, so I am not against its use, but recently if I detect some prose is written by AI it's hard for me to finish it. The written word is supposed to be a window into someone's thoughts, and it feels almost like a broken social contract to substitute an AI's "thoughts" here instead.

AI generated prose should be labeled as such, it's the decent thing to do.

  • Or just by somebody that knows how to use English punctuation properly.

    Is it so hard to believe that there are some people in the world capable of hitting option + “-“ on their keyboard (or simply let their editor do it for them)?

    • I said em dash _and_ the word comprehensive. If you work with LLM generated text enough it gets very easy to see the telltale signs. The emojis at the start of each row in the table are also a dead giveaway.

      I am guessing you are one of those people who used em dashes before LLMs came out and are now bitter they are an indicator of LLMs. If that's the case, I am sorry for the situation you find yourself in.

      7 replies →

  • You forget that MS Word loves to substitute things like em dashes in where you don’t want them. The “auto correct” to those directional quotation marks that every compiler barfs on used to be a real peeve with I was forced to use MS junk.

  • > AI generated prose should be labeled as such, it's the decent thing to do.

    The decent thing to do is to prefix the slop with the prompt, so humans don't waste their time reading it.

  • I don’t really care if it was.

    It’s also documentation for an AI product, so I’d kinda expect them to be eating their own dogfood here.