Machine Learning in the Browser (MLitB, http://software-engineering-amsterdam.github.io/MLitB/) is an ambitious software development project whose aim is to bring ML, in all its facets, to an audience that includes both the general public and the research community. MLitB is not just a software library, but is a philosophical approach to achieving our goals of bringing powerful machine learning models to a global audience.
The ubiquity of the browser as a computational engine makes it an ideal platform for the development of massively distributed and collaborative machine learning.
To accomplish this, we have written a lightweight software library that runs in native browsers, which is capable of performing massively distributed machine learning on heterogeneous devices (desktops, laptops, smart phones and tablets), embedding ML models into the devices for both prediction and collaborative learning initiatives, be shared as an encapsulated object containing configurations, code, and parameters.
Implications of this software paradigm include truly collaborative machine learning, open to the public, real privacy preserving computation, field experiments, green computing, and more.
To install, clone the MLitB repository. Then use the node.js package manager npm to install the required packages. MLitB uses two servers, a master Machine Learning server (MLitB) and a data server (that for now only handled zipped image directories).
$ cd MLitB/MLITB
$ npm i
$ cd MLitB/imagezip
$ npm i
The data server uses redis to organize data vectors; download and installation information can be found at http://redis.io/download.
Clients do not require any software installation beyond a browser.
In a terminal, start redis
$ redis-server &
Start the Machine Learning server. From the MLitB/MLITB directory, start the node application:
$ node app.js -h mlitb_host
Start the Data server. From the MLitB/imagezip directory, start the node application:
$ node app.js -h mlitb_host
Now open a tab in your browser to http://mlitb_host:8000 (without specifying the host url the default is http://localhost:8000/). Note that the data server will be running on port 8001.
Click http://localhost:8000/#/new-project. Select a preconfigured Neural Network from the selection on the right. Add a name and select an iteration time. The iteration time is the length in msecs of a synchronized event loop. During an event loop, the ML server and its clients perform MapReduce step, along with other events such as data uploading/downloading, client joining/leaving, etc. Click Add Neural Network and a new NN will be added to the project list.
From the project list, click the new NN. From this view you can upload data, modify hyperparameters, and add slave nodes.
MLitB is a prototype based on image classification. The only data types understood at the moment are images (jpeg/png) that can be uploaded in zipped directory structure. For example, files can be organised like /dataset/cat/cat1.jpg, /dataset/dog/dog1.jpg, etc; these files should be compressed to dataset.zip. Then select dataset.zip when uploading data. MLitB will automatically assign labels cat and dog to the data vectors. Note that you can upload data of any size; the images will be cropped automatically based on the dimensions of the first layer of the NN.
To train, click Add worker, then choose train for their task. Click Restart to begin transferring initial dataset to clients/workers. Click Pause after the transfer has finished and click Restart to start training.
MLitB is a collaborative effort between researchers at the University of Amsterdam with support from Amsterdam Data Science.
- Ted Meeds
tmeeds@gmail.com - Remco Hendriks
remco@topnotchdevelopment.nl - Said al Faraby
said.al.faraby@gmail.com - Magiel Bruntink
magiel.bruntink@gmail.com - Max Welling
welling.max@gmail.com
Please cite MLitB whenever possible:
arXiv bibtex