(EDIT due to wrongly stating that models run on the Raspberry Pi directly)
The Google Vision Kit will run models on a custom neural processing chip connected to the Raspberry Pi Zero. With the DIY setup from the blog bost, the neural network runs on a "large pc" (potentially with GPU). Depending on the hardware you have at your disposal, you can run more complex (and therefore more powerful) neural networks. At the same time, you'll need wifi set-up and streaming to work. Completely embedded devices are easier to just put in the wild.
In theory, you should be able to use the models from the Vision Kit if you follow their instructions and just put the on a Raspberry Pi directly, and get an additional Movidius compute stick: https://developer.movidius.com/
Inference doesn't run on the RPi Zero. It runs on the VisionBonnet board which has a Movidius VPU tensor co-processor on it. RPi is just for handling the LEDs, buzzers and buttons. For training a model with custom datasets, you are correct - something bigger's needed.
Where can I follow your project? I currently use QOwnNotes on Linux which does exactly that, but would be keen to try some alternatives, in particular if you're working on a mobile version.
> (...) the current hype about AI replacing jobs has much more to do with media clickbait than any real issue we should be worrying about right now.
Hype? For sure. But the current advances in autonomous vehicles will have a very real impact on the working life of the common truck driver (3.5 million in US alone), I think we all agree on that.
I for one don't agree. If the autonomous trucks displace the truck drivers over the span of 50 years, I don't see the problems. Very few current drivers will be in the workforce 50 years from now.
Care to share some details on your technology? On whose handwriting was this trained, did you use any public datasets for this? And of course, how well will this perform on writing styles it hasn't seen before?
We did two things to train it (1) scraped the web for photographs of handwritten notes with known transcription to build our training dataset (2) had our university friends / students write out training examples by hand to get more realistic data on what modern handwriting looks like
Scribble currently only supports English, so it does poorly with other languages, but is pretty robust to poor handwriting in English (such as my own).
It gets about 85% of my handwriting correct (my handwriting is abysmal), so there's definitely room for improvement.
You probably know this already, but if you are just looking for an illustrated demonstration of what RNNs are capable of in the text domain, probably the best brief article is the post by Andrej Karpathy about [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/).
Sidenote: A common practice is to take a pre-trained model (e.g. on the Imagenet dataset) and only learn the last few layers for your usecase. This way you can get a well trained feature extractor if your task data is similar, and then only train the classification, which is a lot faster than full end-to-end training.
You can call it an artificial neural network (ANN) as soon as you have two or more artifical neurons connected to each other. As simple as that.
A small amount of neurons might already solve some problem you're having. The XOR problem can be learned by 4 neurons connected to each other.
When you want raw images or similar as an input and have it be classified into 100 classes (e.g. look up CIFAR-10 or CIFAR-100), you will need an architecture with many more neurons.
After all, ANN are simply a tool. Depending on the task, that tool might need to more elaborate. And when you have all those different possible architectures, you want a common way of naming them. Labels such as Deep Learning are simply nomenclature of talking about certain groups of artificial neural networks.
Have been running the trial for a few hours and am really satisfied so far. I see the same issues with Spectactle, though. Hope this can be addressed in a future version, as I'd love to have both apps running permanently.
Still building up a user base so we can have better results, but here is the top list of our platform: https://podcastprofile.com/top
(the project is in its early stages, basically the result of a hackathon we had a few weeks back)
The Google Vision Kit will run models on a custom neural processing chip connected to the Raspberry Pi Zero. With the DIY setup from the blog bost, the neural network runs on a "large pc" (potentially with GPU). Depending on the hardware you have at your disposal, you can run more complex (and therefore more powerful) neural networks. At the same time, you'll need wifi set-up and streaming to work. Completely embedded devices are easier to just put in the wild.
In theory, you should be able to use the models from the Vision Kit if you follow their instructions and just put the on a Raspberry Pi directly, and get an additional Movidius compute stick: https://developer.movidius.com/