Disco - Decentralized Collaborative Machine Learning
Disco enables collaborative and privacy-preserving training of machine learning models. Disco offers both decentralized and federated learning. Disco is easy-to-use mobile & web code. The latest version of Disco is always running on the following link, directly in your browser, for mobile and desktop:
๐ https://epfml.github.io/disco/ ๐
Key Question: Can we keep control over our own data, while still benefitting from joint collaborative training with other participants? - or - Can we train an ML model which is equally good as if all data were in one place, but respect privacy?
Federated learning: The key insight is to share weights instead of data, each user trains on his own machine and periodically shares his learned weights with a central server. The server will agreggate all these weights and send them back. We support all modern deep learning architectures running on device (currently via TF.js).
Decentralized learning: makes this possible, following the same principles as in federated learning, but going one step further by removing any central coordinator. Disco only uses peer2peer communication, while keeping your data local at all times. It puts users in control of the entire collaborative training process, without a central point of failure. We support all modern deep learning architectures running on device (currently via TF.js).
Applications: We're investigating many applications which could be enabled by Disco, including from the medical domain. Currently we offer a predefined list of training tasks, but will facilitate creating new tasks for everyone soon. For now if you have a new application in mind, just send us a pull request.
Join us: We follow an open development process - you're more than welcome to join the conversation on our slack space, as well as on the issues pages here.
Science behind Disco: In this project we aim to build and improve decentralized versions of current machine learning algorithms, which are at the same time (i) efficient (R1,R2), (ii) privacy-preserving (R3,R4), (iii) fault-tolerant and dynamic over time (R5), (iv) robust to malicious actors (R6,R7), and (v) support fair incentives and transparency on the resulting utility of trained ML models. We currently follow a public model, private data approach.
How to use the platform
Tasks
The platform already contains several popular tasks such as Titanic, MNIST or CIFAR-10.
New tasks can be created the following form. To do so, practical information related to the task (e.g. description, features, learning rate, etc.) must be provided. Furthermore, two extra TensorFlow.js
files need to be uploaded:
- A model file in
JSON
format. Please refer to the following official documentation pages to create and save your model. - A weight file in
.bin
format. These are the initial weights that will be provided to new users upon joining the training of your task. You can either provide a pre-trained model or use a simple random initialisation scheme.
Note: for the moment, only
CSV
andImage
data types are supported but stay tuned: it will soon be possible to add new types directly on the platform ๐ฃ
Settings
Under the sidebar on the left, you will find a settings button. We offer multiple personalisation options for the user. In particular:
- Platform type: train
decentralised
orfederated
according to your needs ๐ Be aware that changing the platform type during training will reset the state of the platform. In other words, your training progress will be lostโ ๏ธ - Model library: storage options can be enabled so that your models are safely saved in your browser database ๐พ
- Themes and colors: customise the look of the platform as you please ๐
If you have any questions related to Disco, feel free to raise an issue or join our slack workspace โ