Amund Tveit

This blog post describe the Zedge mobile app and primary use case, and then follows on with what we do with Deep Learning.

Our main product is an app - Zedge Ringtones & Wallpapers - that provides wallpapers, ringtones, app icons, game recommendations and notification sounds customized for your mobile device.  Zedge apps have been downloaded more than 200 million times for iOS and Android and is used by millions of people worldwide each month.

People use our apps for self-expression. Setting a wallpaper, ringtone or app icons on your mobile device is in many ways similar to selecting clothes, hairstyle or other fashion statements. In fact people try a wallpaper or ringtone in a similar manner as they would try clothes in a dressing room before making a purchase decision, they try different wallpapers or ringtones before deciding on one they want to keep for a while. 

The decision for selecting a wallpaper is not taken lightly, since people interact and view their mobile device (and background wallpaper) a lot:

The importance and attention to mobile screen and wallpapers gives incentive to brands to publish high quality content on Zedge e.g. several wallpapers packaged into Zedge Stories, as shown for Entertainment One’s movie the Little prince in the figure below.

Zedge explained in a Pokemon Go context

Since the mobile world currently seems to evolve around Pokemon Go, let me try to explain Zedge in a Pokemon Go context compared to other popular mobile apps and services. The columns below each describe features in Pokemon Go and which popular mobile apps and services they are similar to.

1. Journal of Actions   2. Photo Opportunity   3. Customization of Avatars etc.   
Feeds in:
- Facebook
- Google+
- Pinterest
- Twitter
Taking photo in:
- Instagram
- iMessage
- WeChat

Deep Learning at Zedge

One of the moving parts that helps providing our services is Deep Learning, e.g. for ranking and processing of content with Convolutional Neural Networks. We currently use two open source deep learning tools: TensorFlow (from Google) and Caffe (from Berkeley Vision and Learning Center at Berkeley University) for existing services. For new services we are primarily leaning towards TensorFlow for new projects given its strong momentum and rapid progress.

Some of the challenges we work on and either are using, planning to use or considering to use deep learning for are:

  • Finding similar content - since people have a behavior not unlike trying clothing (as described above), it makes sense to provide suggestions to similar content to what they are trying when they are in the try wallpaper or ringtone flow. One challenge we face is the balance between too similar (which might in some cases be considered a duplicate or boring) and provide similar content with a touch of serendipity. Similarity might be along several dimensions, for wallpapers it can be similar in terms of colors and composition of colors used, the degree of abstractness, emotion the wallpaper describe, or the items/people/animals on the wallpaper.
  • Content Recommendations - how can we ensure that the wallpapers or ringtones we present when you open the app are relevant and have high quality? Perhaps we can find outliers 
  • Personalized Recommendations - browsing and searching for ringtones and wallpapers can sometimes be tedious, if we can predict what is relevant based on what content you have previously looked for and preferred this can improve your experience.
  • Search Ranking - people perform millions of search queries in the Zedge app, and 
  • Improve metadata for Search - uploaded content usually doesn't have a rich set of metadata, and deep learning can be used to generate more metadata that describes objects and situations images (wallpapers) and genre, emotion etc. in sounds (ringtones). Image understanding with Convolutional Neural Networks has had a dramatic improvement in the last few years, with Alex Krizhevsky et. al  (University of Toronto) winning the 2012 ImageNet competition with an error rate of 17% to Christian Szegedy et. al (Google) Inception v4 having an error rate of 3.08%, i.e. a more than 5 times improvement in just 4 years! (see my blog post "Why Deep Learning matters" for details)
  • Localizalization of metadata for search - Since people from all over the world is using Zedge we need to make content easily searchable at least for the most used languages. Deep Learning has shown promise in automatic translation, and might be a cost efficient way for us to support more languages better.
  • Creation and Transformation of content - one sub-fields of Deep Learning is the using it for transforming or creating images is also interesting to look into, check out for what is happening in the field. 
  • Monetization - perhaps deep learning can help improve revenue?
  • Operations - making a mobile service that millions of people are using running smoothly, cost-efficiently and with low latency over time is very, VERY hard. A recent example of that is Niantic Labs' Pokemon Go that due to extreme growth has had their fair share of operations issues. Deep Learning has shown great promise in detecting and predicting outliers/anomalies and novel patterns in data, and can potentially be used in operational settings. Deep Learning is applied in operational settings by a few companies, e.g. Baidu uses it to predict which hard disks in Data Centers are going to fail, and Google uses it to reduce energy usage in Data Centers.

(header photo © User: Gengiskanhg / Wikimedia Commons / CC-BY-SA-3.0)

Amund Tveit 
Lead, Data Science Team

btw: If you want to help us build Zedge, we have openings for:

  • Senior Android, iOS and Frontend Developers
  • Senior Backend Developers and Data Scientists
  • Business Development, Content and Product Managers

see Join our Playground for details.

About Zedge

Zedge (NYSE Market: ZDGE) is a content platform, and global leader in smartphone personalization, with more than 200 million app installs and 30 million monthly active users. People use Zedge to make their smartphones more personal; to express their emotions, tastes and interests using wallpapers, icons, widgets, ringtones and more.The Zedge platform enables artists, teams and brands to extend their reach, and gain additional insight, by giving fans and customers a seamless way to personalize their smartphones with content they associate with.