Clarifai Mobile SDK - Machine Learning On The Phone
Our portfolio company Clarifai introduced something exciting yesterday.
It is a mobile SDK that any developer can put into their mobile app and it will allow for machine learning on the device:
Machine learning (the process by which computers can get smarter through data examples instead of explicit programming) requires massive computational power, the kind usually found in clusters of computer servers in massive datacenters (ooooh, the cloud). This means that machine learning technology is usually only available to those who can connect to the cloud.
Not anymore! Clarifai’s Mobile SDK gives users the power to train and use AI in the palms of their hands by installing machine learning capability directly on their devices, bypassing the traditional requirement of internet connectivity and massive computing power. After all, these days we have tiny supercomputers in our pockets – our mobile phones. Starting with an iOS SDK, Clarifai is on a mission to make user experiences uniquely personalized on any device from your cellphone to your toaster, anywhere in the world.
Here’s a slideshow that explains how this works:
Congrats to Clarifai team. My fun and games with Tensorflow image recognition.https://uploads.disquscdn.c…
DARPA has a great video on the three waves of AI and their comments on what Deep Learning can’t do are really informative.https://youtu.be/-O01G3tSYp…
Thanks.Good to hear what DARPA thinks.Their first wave was on expert systems. I worked in that, wrote some core software, did some applications, published some papers, and then worked up some applied math that was quite different, applied math that totally knocked the socks off our expert systems work. What the video says about expert systems is not really correct.In the video, for the second wave, okay, can do that with neural networks. But for their example of the big rotating sphere (sounds like data from the first three principle components) could attack via the L. Breiman classification and regression trees (CART), and no doubt the more recent work is son of CART.He, we, should emphasize that what they are trying to do is just classification.For the third wave, he gives a poor, unclear description of what, even if as successful as he hopes, will be disappointing.Calling any of this work in any meaningful sense intelligent is from hype to scam to fraud.We’ve had an AI Winter. We just left an AI Spring of Hype. We are now in the Summer of Scam. Soon we will be in the Fall of Failure. Then get your warm hats, coats, and gloves for the coming AI Winter — it promises to be darned cold.DARPA can pursue this stuff. And sometimes DARPA has some very good results.But, for business, there are much better opportunities pursuing other parts of and directions in what is essentially some applied math.That is, for the second wave, I just didn’t see that I had sitting around classification problems with terabytes of training data and needed some applied math and GPUs to process it!
That was a fantastic video. He did such a great job of explaining it. Any other recommendations 🙂
statistically impressive but individually unreliable.
The image recognition model will change from CNNs, though. That’s too compute-intensive.This is from Vicarious. They only needed 1,406 images compared with the typical 8 million for DL systems trained on ImageNet. https://uploads.disquscdn.c…
Good luck with this. I just flipped through the slide show (makes no sense without someone giving the talking points)…and then I clicked over the details you linked to (lots of hype around ‘machine learning on the device’ but no real details).This is an SDK, they can skip most of the marketing hype (the developers who will want to use it will know *why* they want to use it) and get right into the details and examples. After 10 minutes of clicking and reading, I still don’t know what this thing is specifically helping me as a developer *actually* do.Examples, beyond just name dropping one client, with specific problem/solution details would go a long way here (it took a little rabbit hole searching to figure out that the one client they mention hired Carlifai to custom design a ‘visual recognition model especially for ear pictures and video’)…Presumably, this SDK means you can do something like a “image recognition” out of the box…just supply some initial pictures (example of how many?), tweak a few options (example of one or two), and wha-la you’ve got a trained model that can start to classify any new pictures your mobile app takes in…if this actually is the case, why don’t they just explain it like that? (BTW, it’s actually pretty cool if that’s the case).Hot Dog vs. Not Hot Dog is a great, and timely, example they could be running with here…
Hey! I won $100 gift vouchers in Amazon’s Alexa Challenge with my Da Vinci skill.
Awesome! If it’s an alexa skill in production, you should also be eligible for the $100/month AWS credits too ( https://developer.amazon.co… ) 🙂
Thanks, applying for this!
Can you elaborate? I loved your video.
It’s a very simple app which just gives random Da Vinci quotes:* https://www.amazon.com/UR-D…
This. And assuming “image recognition” captures what they do – why would I use that and not, let’s say, the solid rekognition framework from aws.EDIT: wrt ‘not hotdog’, the ios implementation is using embedded TensorFlow.
Hey @falicon:disqus – the docs are available here: https://developer.clarifai….This should give you more information about the offline SDK. Would love to hear your thoughts/feedback on it! 🙂
Thanks – took a real quick look. Def. a lot more clear than the articles linked above (so thanks for that).
What’s the url for your startup website? Would love to check it out.
Thanks.Well, the startup is not live yet. Why not? Because of nonsense, simple, blind, stupid, brain-dead, unpredictable doo-dah bad luck! That is, independent, exogenous interruptions!As I type this, my startup software won’t run. Why not, it did run apparently just fine? Well, there was a hardware problem. It looks like I fixed it. But it looks like during the hardware problem, and a lot of sudden shutdown and reboot episodes, some of the data on my boot drive got corrupted. So, as my software starts and tries to pull in standard Windows .NET, ASP.NET, ADO.NET code, poof, the corruption means that my code fails.So, reinstall everything on my boot drive? Nope: I have multiple instances of good boot drive backups from Microsoft’s somewhat amazing (get a good backup of a boot drive while it is running) old NTBACKUP. So, I need to restore from that. For that I’m being careful.Meanwhile, I wore out my mouse. Fixed some of it, and it’s working now.Meanwhile I wore out my keyboard. I have another one that works, but I wore it out also and need to use some epoxy to fix a tiny plastic part of the torsion bar part of how the space bar works.The nonsense goes on this way. I was doing fine, but then the unpredictable, exogenous interruptions started. Screaming likely wouldn’t help. So, just beat back the interruptions. Order a new mouse from Amazon. Fix the plastic on the keyboard I like or spring for a new keyboard I might not like. Etc.As soon as I get my software running again, I want to do some more testing. Some of the core code for some of the applied math is a little tricky (tricky indexing), and I want to be sure a third time that that code is just right.Then I want to add some more data to the database.Contact some ad networks and write the code to collect the data they want and code to stuff their ad URLs into my HTML code — my Web pages show ads nicely now, but the ads are coming just from my hard disk instead of an ad network — that is, my code handles ad URLs now.Then I want to do some alpha testing with others. Then I want to do some beta testing.Then do the rest to go live, the routine stuff, static IP address, domain name, business name and its registration, tax ID, etc.I do have some changes in mind: For the server log file, I’ve been using what Microsoft supplies. That worked well enough to trace bugs during my software development. But for production, I want a better solution. But that’s easy: Just take the code I wrote for the Web session state server (I didn’t like Microsoft’s ideas for a session state server, either, so wrote my own code using just TCP/IP sockets, object de/serialization, and two instances of a standard collection class), rip out the code for the collection class instances, put in a few lines to write to a file, and done. Getting all my other code to use the new code for the site log file is easy: All the current writes to the log call a subroutine I wrote; so I just change that one subroutine to use the new log file and, presto, all the log file writing will use the new log file server.There is one place I want to slightly tweak the UI — just a few lines of code in just one Web page. Otherwise the code is 24,000 programming language statements in 100,000 lines of typing which all seem fine.All the work that is mine has been fast, fun, and easy. Fixing hardware problems, boot drive data corruption problems (e.g., the corruption means that my installation of Excel won’t work), mouse problems, keyboard problems, an async port problem, a B/W laser printer problem, an Ethernet port problem, need each three months or so to take apart the processor cooling and add three drops of 3-in-1 oil to the shaft of the processor fan problem, my running shoes coming apart problem, the leaking tub boot seal on my Maytag washer problem, my kitty cat’s bugs in her fur problem, etc. — what a mess!
AI everywhere . Cool. It’s like the super fuel for Apps?
This is super cool. Been a fan of Clarafai for a while (Andrew Pile brought it up while I was working with Vimeo in 2014/15). I’ll share this with the guys on our team at Citizen.
Making a separate post here as well … for all interested, the docs are available here: https://developer.clarifai….We are currently reviewing all applicants ( clarifai.com/get-sdk ), and will be responding back with installation/download details end of this week/beginning of next week!
Despite all the hyperbole, I think they have room to the upside.