Subscribe
Sign in
DiscreetAI — Build better AI by using on-device data
DiscreetAI allows you to train machine learning models on untapped sensitive datasets while maintaining privacy, security & compliance for your customers, in just a few lines of code. Try it now for free!
35
Replies
Mujtaba Hassanpur
Looks great, congrats on the launch!
Share
Really cool, congrats guys! Some annotated tips here too on conversion optimisation on the above the fold section of your landing page, hope its helpful - https://app.usebubbles.com/c2639...
Hey Product Hunt!
We're excited to launch DiscreetAI for mobile, web, and datacenters, an open-source developer tool that will help you build better AI by giving you access to untapped on-device data.
If you want to build an app powered by a cool machine learning model, like an app that can tell you cocktail recipes just from a picture of the liquors in your pantry, how can you collect the data to build that model? You can find some data online, scrape data, but the best source of data for you will be users who are taking pictures of their drinks with your app! However, your users aren’t just going to upload pictures of their drinking habits to share with you. With DiscreetAI that data becomes accessible to you, without you having to collect it at all!
Here’s how it works; your model updates on the user’s smartphone when they take a picture. We fuse those updates into a model that has learned about all the varieties of liquor your users have ever photographed, without ever collecting a single photo centrally. Now, anyone can train high quality models on consumer data while preserving the privacy of sensitive user data and remaining legally compliant with the latest regulations in consumer data privacy.
You can try DiscreetAI out at https://beta.discreetai.com/signup and check out the codebase at https://github.com/DiscreetAI/de....
To get started developing iOS apps that can generate insights from user data in a secure way, download our demo app on Testflight for an example https://testflight.apple.com/joi... and install DiscreetAI via Cocoapods https://cocoapods.org/pods/Discr....
To use DiscreetAI in your web application, check out a chatbot powered by DiscreetAI https://github.com/DiscreetAI/di... and install DiscreetAI via NPM https://www.npmjs.com/package/di....
We want to get your feedback on how to improve DiscreetAI. Thanks for your support!
@wellford_chan that's definitely a valid concern! We start off by never moving data off the device; the only information that's communicated is updates to the model. Making sure that these updates to the model don't leak any data is an area of active research, and we have some secret sauce that sets us apart from other folks working on this right now. If your organization is concerned and wants to understand the technique, please reach out via our website!
@chris_lin2 yes, this tech is called "federated learning" and has been deployed in such cutting edge products as Google Keyboard and Google Mail! We're making it accessible for any company with our open-source service.
Absolutely - to run the demo notebook head over to discreetai.com and signup; if you don't want to make an account you can use the Colab Notebook here https://github.com/DiscreetAI/de...
Interesting, how does this compare to traditional data labeling services? Most companies looking to train a model will probably will look to them for high quality training data.
@justin_hao That's true! Companies have to be able to centrally aggregate data in order to use a traditional data labeling service like MTurk or scale.ai, and our product helps companies that want to train on data they can't centrally aggregate; you wouldn't be able to aggregate people's health records and send them to a data labeling service, for example, but you would be able to use DiscreetAI to build an application that could generate useful insights from that health data.
Thanks for the feedback @maltan_maltan! We named it DiscreetAI because as per dictionary.com, discreet means "judicious in one's conduct or speech, especially with regard to respecting privacy" and we wanted to build a product that helped AI developers build the powerful models they needed while still respecting the privacy of their users!
Hey @dia_aj! Let's say you have an app that provides a service to users by generating an insight from a sensitive piece of data, e.g. medical diagnosis, credit card fraud prevention, virtual assistant, smart keyboard, etc. The users are using the app because they're getting something out of it. You also want to improve your app, and your users provide useful data to improve your app while they're using it! For example if your app is a smart keyboard and your model autocorrects what someone is typing, if the autocorrect suggestion is correct then the user would click that suggestion and that's a positive signal for the model, if not it's a negative signal. By using this signal you can improve your model, and DiscreetAI provides a framework for you to do that without needing to aggregate sensitive information such as what the user is typing. I hope that helps!
Hey James we don't have an Android SDK out yet but we're working hard on it! Expect to see an Android release from us later this year :)
This looks great! Quick question on user data/privacy, what ensures that users have control of their own data/results that are generated?
Thanks @mmitesh16! Our code is open source, so anyone confirm for themselves that we aren't collecting any data that's generated by users, and that we're instead just generating results from that data. As far as keeping control over their results, anyone can similarly confirm that we don't share individual user updates from anyone, and only the application developer receives the final trained model which doesn't reveal any user information.
Looks awesome! Very promising work by smart founders in an increasingly important field.
@jimmy_yao1 Thanks! You can check out the tutorials in our repo; https://github.com/DiscreetAI/de...
