In November 2018, our Data Science cohort Tensor Thyme visited Mobius Labs. The project manager and two engineers took a lot of time to present the company and answer our questions. Here are our main takeaways:

What does Mobius Labs do?

Mobius has trained “two huge neural networks” on millions of labeled photographs. The first network annotates photographs with a set of over 5000 keywords, so that you can do semantic analysis, similarity search and many more. They also rate the aesthetics of images, allowing you to tell good from bad images. The models are available as a mobile SDK and an on-premise version.

The team consists of experts – mathematicians and computer scientists – with an impressive record: 4 PhDs and 80 publications, many on computer vision.

The network has been extensively benchmarked, and the benchmark is published on GitHub here.

What sounded familiar to us?

We were happy to hear many concepts that we know from our curriculum: Convolutional Neural Networks, the layered architecture of the network, feature vectors and T-SNE to name a few.

On their mobile demo App, Mobius allows users to personalize the model by labeling comparatively few (100+) training images themselves. It’s no secret that this includes some transfer learning.

And of course, Python and TensorFlow are being used at many points in the process.

Why is Mobius far ahead?

To train a model that exceeds the capabilities of models like VGG-19 by far, Mobius has spent 3 years fine-tuning. Model hyperparamters like number, size and configuration of layers is nothing you could determine by wrapping your model in a for loop or running GridSearchCV from scikit-learn.

Coming up with an aesthetics score that works for diverse groups of users and that is double-checked by experts also took time and diligence. And of course, curating millions of images is much more complex than running a Scrapy script.

On top of that, training and running the model requires some infrastructure: storage, GPUs and a software development pipeline are enough to keep data engineers busy.

In our eyes, Mobius has made a powerful entry to “tame the flood of images” that surrounds us. We have learned a lot from the encounter and are curious to test their Android Demo App ourselves.

 

Nov 8th, 2018 Dr. Kristian Rother