Artificial Intelligence with App Inventor

[piero alabrese]

Hi,

I downloaded the tutorial ("WhatIsIt") to create an app that can recognize objects thanks to artificial intelligence thanks to the "Look Extension" extension.

The guide presents this site: "https://teachablemachine.withgoogle.com/", what is the use of this site?

with this site it is possible to create models and export them. can I subsequently insert the model created by replacing it with the "Look Extension" extension in my app?

Currently the app created following the tutorial makes many errors in the recognition of images, how can I correct these image recognition errors? how can i learn the machine?

anyone know how to clarify my ideas?

thanks

[Evan Patton]

Hi Piero,

Google has recently updated the Teachable Machine website, so our documentation is now a bit out of date. Originally, the site did not provide an export feature. We have a separate extension coming out that will support training your own model. The model used in the LookExtensions (part of WhatIsIt) is a fixed model based on MobileNet, a neural network trained and published by Google. At this time we don't have a way of updating it to include more examples. Keep an eye out for the personal image classifier example in the coming weeks that will give you a way of training your own model.

Cheers,

Evan

[BrandnameOra]

Okay..
I'm interested in this discussion and I understand that the export possiblity is coming soon.

I'm the blogger behind this post
https://usinggdpr.wordpress.com/2018/10/08/tradinglive-io-experience/

I've got an idea to make an app that would answer my incoming calls and have an AI character reply to this call.

I really wish that I could make that app with the teachablemachineeithgoogle.com.
Can I do it with it?

[piero alabrese]

thank you very much for the info !! so if I understood correctly the extension I used (suggested in the tutorial) is "static" does not help the machine to learn new things, correct? while in the new extension that will soon be available, it will be possible to teach the machine when the recognition is wrong, correct? I wanted to ask one thing ... after creating a model with the Google platform, it is possible to download code that can be used in the browser, I understand correctly ?? so the idea is this: why not insert this code in the browser used inside the app? in this way I could use a model of image recognition created by me therefore personalized !! is it a feasible idea or are there things I don't know that won't allow me to do all this ??

thanks Piero

[BrandnameOra]

I've now digged deep on the website and I see a lot of possibilities to export your materials.

I'm guessing that I don't quite understand what you guys are talking about above.

Although I had a blast going through the website, I now know that what I can do with it

[Evan Patton]

thank you very much for the info !!  so if I understood correctly the extension I used (suggested in the tutorial) is "static" does not help the machine to learn new things, correct?  

Correct. It's just evaluating a model that has already been trained.  No new training occurs with the Look extension.

 

while in the new extension that will soon be available, it will be possible to teach the machine when the recognition is wrong, correct?  

Yes, but training is done on a laptop/desktop rather than on the phone as training requires more time/memory resources than just running the model. 

 

I wanted to ask one thing ... after creating a model with the Google platform, it is possible to download code that can be used in the browser, I understand correctly ??  so the idea is this: why not insert this code in the browser used inside the app?  

This is essentially what we do as the extensions are wrappers around TensorFlow.js and a WebView that present the underlying JavaScript representations of the outputs as App Inventor compatible types. Of course, unless you make an extension you'd have to implement the rest of your logic in JavaScript as well using the WebViewer component.

 

in this way I could use a model of image recognition created by me therefore personalized !!  is it a feasible idea or are there things I don't know that won't allow me to do all this ??  

Correct. The second extension that we are going to publish allows you to train with your own images so that it is personalized. In fact, the extension name is PersonalImageClassifier.

Regards,

Evan

[yamafu]

Thank you, Evan, for useful information.

I currently do the following to link Teachable Machine2 with App Inventor. That is, after training my model on a PC, I edit the snippet of the JavaScript (or P5.js) that is presented and call it using the WebViewer block. However, the cross-domain problem occurs when providing test images to this trained model, so the trained model should be placed on a local Webserver launched within Android. This is going well, but a little cumbersome. It will be very useful if your "PersonalImageClassifier" is released. I look forward to this release.

Regards,

Fujio Yamamoto