The big tech companies keep trying this (commoditizing ML models exposed via APIs), but I am extremely skeptical that they will succeed with this approach.
The reason is simple. All of the work lies with validating that a model solves your particular problem well enough to be cost-effective for your customers or stakeholders to be satisfied. Machine learning is not a commodity the same way cloud infrastructure is, because the model development and validation aspects are inherently tailored to your extremely specific, one-off data generating processes and performance characteristic requirements.
Even if you outsource the model itself, this still requires developing some notion of acceptance testing for the model's performance in your use case, on the distribution of data that matter to you. And such acceptance testing still requires high literacy in statistics and model evaluation (e.g., you can't skip on hiring that expensive machine learning engineer even if you outsource because then who, internally, is going to be able to tell you if the outsource solution is a bunch of junk, and why, and what to do about it.)
Short of turning the big tech ML offerings into flat out consulting arrangements, where you trade-off having your own in-house and application-specific and data-specific machine learning staff in order to rent time from experts at the big co's, this model couldn't work.
Don't get me wrong, just as IBM somehow still finds ways to leach from enterprise clients after all these years of not actually helping them, I am sure these lines of business will generate enterprise consulting money.
I just think lots of people will be disappointed that it doesn't somehow mean you're getting Google ML engineer caliber attention or results spent on your company's bespoke needs, and that performance of generic and naive transfer learning from big cookie cutter models often ends up being way worse than you thought.
I feel bad for people who end up suffering the amplified version of vendor lock-in this could create.
Core ML is bring-your-own-model (or use the built-in classifiers, which have generic labels). This is an easy way to train custom models (for example, on images from your own product catalog).
Because I lack imagination and or don't understand ML enough, could someone give a couple of examples of what this could be used for? Particularly at these scales with presumably(?) a smaller initial data set.
More complex applications: apps that do clever audio processing, smart image filters, making some sort of intelligent predictions or recommendations based on a users behavior, communications, preferences, etc.
These are not revolutionary or even helpful ideas, but they can be convenient or fun. For example:
Some medical applications (just for example, you would need quite a robust model):
* Categorize moles on your skin as cancerous or not.
* Categorize cuts on your skin as infected or not.
* Have the user input certain characteristics (images, temperature data, etc) and give some preliminary diagnosis.
Less interesting/helpful:
* Detect photos of certain foods in your app (e.g. Twitter, Instagram, Yelp) and recommend relevant emoji or hashtags.
* For a note-taking or to-do app, classify notes into categories automatically for the user.
* Suggest actions in your app based on what the user types or does.
* Recommend solutions or help articles from the feedback form in your app. Tag the feedback with a certain sentiment to determine how quickly you should follow up on it.
* Detect your products in images/videos and tag them to make them searchable, and recommend relevant things to the user (e.g. in a beer tracking app [2], detect a certain kind of beer, suggest similar types).
If any of this sounds like it's been done before, it probably has, but it's important to note this is done entirely on device (private!) and with custom labels. The current alternative is to use TensorFlow Lite [1], which is a bit more involved. I'm sure as the field develops we will see more creative (and useful/helpful) applications.
This looks like a really easy way to integrate your own custom ML models into your app. I imagine this would appeal to a lot of hobbyist developers, especially given how simple it is.
The docs [1] also seem to imply they're using transfer learning from more robust models for their image classifier: "Use at least 10 images per label for the training set, but more is always better." and "Create ML leverages the machine learning infrastructure built in to Apple products like Photos and Siri."
Could not seem to find any mention of where the model is being trained...I would assume locally, but again, can't seem to find a mention of recommended hardware for this.
It's a known problem in ML community. Unfortunately saying "hey we do pattern matching in our product" dosen't sound so good to investors as "We are building thinking machines using artificial intelligence".
Things like this convince me that we will never get to AGI. I have a, perhaps untestable hypothesis, that an AGI could only be built by an AGI, or that if we create it, we won't entirely know how we created it in the first place. (Much as we don't completely understand how to create life force, even though we do create it, through the process of reproduction). We may be able to build components of it (like growing organs in a lab) but I am quite skeptical we will be able to create "life" or in this case "intelligence".
In my uninformed opinion, ML is more akin to the subconscious parts of our brain, like our ability to recognize peoples faces. We don't learn how to do this, it's there when we are born. It is a hardwired part of our brain that was formed through evolution instead of learning and does not involve higher order thinking. It's obviously useful, and could/would form part of a AGI system but unlikely to be enough.
27 comments
[ 5.0 ms ] story [ 85.4 ms ] threadThe reason is simple. All of the work lies with validating that a model solves your particular problem well enough to be cost-effective for your customers or stakeholders to be satisfied. Machine learning is not a commodity the same way cloud infrastructure is, because the model development and validation aspects are inherently tailored to your extremely specific, one-off data generating processes and performance characteristic requirements.
Even if you outsource the model itself, this still requires developing some notion of acceptance testing for the model's performance in your use case, on the distribution of data that matter to you. And such acceptance testing still requires high literacy in statistics and model evaluation (e.g., you can't skip on hiring that expensive machine learning engineer even if you outsource because then who, internally, is going to be able to tell you if the outsource solution is a bunch of junk, and why, and what to do about it.)
Short of turning the big tech ML offerings into flat out consulting arrangements, where you trade-off having your own in-house and application-specific and data-specific machine learning staff in order to rent time from experts at the big co's, this model couldn't work.
Don't get me wrong, just as IBM somehow still finds ways to leach from enterprise clients after all these years of not actually helping them, I am sure these lines of business will generate enterprise consulting money.
I just think lots of people will be disappointed that it doesn't somehow mean you're getting Google ML engineer caliber attention or results spent on your company's bespoke needs, and that performance of generic and naive transfer learning from big cookie cutter models often ends up being way worse than you thought.
I feel bad for people who end up suffering the amplified version of vendor lock-in this could create.
- https://aws.amazon.com/sagemaker/
- https://developer.ibm.com/ai/
More complex applications: apps that do clever audio processing, smart image filters, making some sort of intelligent predictions or recommendations based on a users behavior, communications, preferences, etc.
Some medical applications (just for example, you would need quite a robust model):
* Categorize moles on your skin as cancerous or not.
* Categorize cuts on your skin as infected or not.
* Have the user input certain characteristics (images, temperature data, etc) and give some preliminary diagnosis.
Less interesting/helpful:
* Detect photos of certain foods in your app (e.g. Twitter, Instagram, Yelp) and recommend relevant emoji or hashtags.
* For a note-taking or to-do app, classify notes into categories automatically for the user.
* Suggest actions in your app based on what the user types or does.
* Recommend solutions or help articles from the feedback form in your app. Tag the feedback with a certain sentiment to determine how quickly you should follow up on it.
* Detect your products in images/videos and tag them to make them searchable, and recommend relevant things to the user (e.g. in a beer tracking app [2], detect a certain kind of beer, suggest similar types).
If any of this sounds like it's been done before, it probably has, but it's important to note this is done entirely on device (private!) and with custom labels. The current alternative is to use TensorFlow Lite [1], which is a bit more involved. I'm sure as the field develops we will see more creative (and useful/helpful) applications.
[1] https://www.tensorflow.org/mobile/tflite/
[2] https://untappd.com/
The docs [1] also seem to imply they're using transfer learning from more robust models for their image classifier: "Use at least 10 images per label for the training set, but more is always better." and "Create ML leverages the machine learning infrastructure built in to Apple products like Photos and Siri."
[1] https://developer.apple.com/documentation/create_ml/creating...
> The docs [1] also seem to imply they're using transfer learning from more robust models for their image classifier
I read that as a recommendation for the minimum number of training samples per label (e.g. 10 photos w/leopard tagged).
Apple's Turi Create [1] also also been getting some love these last few months to add support for exporting to Core ML.
[1] https://github.com/apple/turicreate
[1] https://github.com/DeepRNN/image_captioning [∞]