4 ms·

Android to launch TensorFlow Lite for mobile machine learning

119 points by ascorbic 9y ago

pplonski86 9y ago
Interesting, I'm wondering if it will be more popular to learn ML model on mobile or just to deploy model on mobile and run predictions.
- quadrature 9y ago
  if you can learn on mobile you can have the model update it self with user specific data without sending that data back to a server.
- ndr 9y ago
  It's going to be resource intensive, aren't your users going to hate you for draining their battery?
  - tdb7893 9y ago
    If you found a way to have efficient distributed training you could just have each device do a few training runs and the user probably wouldn't even notice.
    - IanCal 9y ago
      You may be interested in federated learning: https://research.googleblog.com/2017/04/federated-learning-collaborative.html https://research.googleblog.com/2017/04/federated-learning-c...
  - danieldisu 9y ago
    they show an example of ML when selecting a text in android and how it predicts how much text should it select depending on what it is (a full address, an email etc).. don't know how they made it but it was almost instant, I'm sure the users wont care.
    - ndr 9y ago
      I think that was just using a pre-trained model, i.e. the training didn't happen on the device but offline.
  - goda90 9y ago
    It'd be interesting to compare the power drain of running the model vs transmitting the data to a server, waiting for the response, and then acting on it. Even if it's still worse, it might be worth the responsiveness and reduction of data usage.
    - Klathmon 9y ago
      Not to mention that training can be done somewhat "out of band". So you can have your phone train on data while charging overnight to get better predictions the next day.
  - mtgx 9y ago
    It will use accelerators, which should help a bit. You shouldn't use this on the phone's GPU.
  - quadrature 9y ago
    Simple models should run performantly, we're all thinking of CNNs and RNNs, but a simple Logistic Regression will probably the most ubiquitous application.
zitterbewegung 9y ago
So I assume it will be similar to this ? It almost sounds like they would keep the interface to this but redesign how it uses the specialty chips. https://www.tensorflow.org/mobile/ https://www.tensorflow.org/mobile/
- tome 9y ago
  I'd like to know how it's different from that at all.
sanxiyn 9y ago
The most important part of the announcement seems to be about hardware acceleration. Maybe it's something like Qualcomm NPE? https://developer.qualcomm.com/software/snapdragon-neural-processing-engine https://developer.qualcomm.com/software/snapdragon-neural-pr...
- nswanberg 9y ago
  It'll be cool to see what Qualcomm releases. Until then, there's already some released Tensorflow code that will run on the Hexagon DSP in a Snapdragon 820: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/hvx https://github.com/tensorflow/tensorflow/tree/master/tensorf... You can run it on a rooted OnePlus 3t (or a OnePlus 3 if you find one used).

119 points by ascorbic 9y ago

Interesting, I'm wondering if it will be more popular to learn ML model on mobile or just to deploy model on mobile and run predictions.

quadrature 9y ago
if you can learn on mobile you can have the model update it self with user specific data without sending that data back to a server.
ndr 9y ago
It's going to be resource intensive, aren't your users going to hate you for draining their battery?
- tdb7893 9y ago
  If you found a way to have efficient distributed training you could just have each device do a few training runs and the user probably wouldn't even notice.
  - IanCal 9y ago
    You may be interested in federated learning: https://research.googleblog.com/2017/04/federated-learning-collaborative.html https://research.googleblog.com/2017/04/federated-learning-c...
- danieldisu 9y ago
  they show an example of ML when selecting a text in android and how it predicts how much text should it select depending on what it is (a full address, an email etc).. don't know how they made it but it was almost instant, I'm sure the users wont care.
  - ndr 9y ago
    I think that was just using a pre-trained model, i.e. the training didn't happen on the device but offline.
- goda90 9y ago
  It'd be interesting to compare the power drain of running the model vs transmitting the data to a server, waiting for the response, and then acting on it. Even if it's still worse, it might be worth the responsiveness and reduction of data usage.
  - Klathmon 9y ago
    Not to mention that training can be done somewhat "out of band". So you can have your phone train on data while charging overnight to get better predictions the next day.
- mtgx 9y ago
  It will use accelerators, which should help a bit. You shouldn't use this on the phone's GPU.
- quadrature 9y ago
  Simple models should run performantly, we're all thinking of CNNs and RNNs, but a simple Logistic Regression will probably the most ubiquitous application.

zitterbewegung 9y ago

So I assume it will be similar to this ? It almost sounds like they would keep the interface to this but redesign how it uses the specialty chips. https://www.tensorflow.org/mobile/ https://www.tensorflow.org/mobile/

tome 9y ago
I'd like to know how it's different from that at all.

sanxiyn 9y ago

The most important part of the announcement seems to be about hardware acceleration. Maybe it's something like Qualcomm NPE? https://developer.qualcomm.com/software/snapdragon-neural-processing-engine https://developer.qualcomm.com/software/snapdragon-neural-pr...

nswanberg 9y ago
It'll be cool to see what Qualcomm releases. Until then, there's already some released Tensorflow code that will run on the Hexagon DSP in a Snapdragon 820: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/hvx https://github.com/tensorflow/tensorflow/tree/master/tensorf... You can run it on a rooted OnePlus 3t (or a OnePlus 3 if you find one used).