Featured Post

Hello Deep Learning

Oct 26 2015 · by
image


Deep learning is a hot and fascinating research area, particularly when applied to classifying images. While researching the Fast Forward Labs Deep Learning: Image Analysis report, we played with a lot of very cool technology. In this blog post, we offer a guide to getting started with deep learning by using APIs from some of the most interesting deep-learning-as-a-service startups.

These APIs accept images and/or video, and quickly classify objects, ideas, and items shown in the images and video. We used this capability to build pictograph.us, an app that allows you to visualize your Instagram photos by the contents of the photos.

We have many ideas for other applications, and wanted share some tips and code to help you get started working in this field as well.

If you want to get started with deep learning for both images and videos, check out our repo hello_deep_learning.  There’s a nice jupyter notebook to get you started as quickly as possible.  We chose to focus on MetaMind, Dextro, CloudSight, and Clarifai because their mature and robust APIs handle large numbers of images while maintaining high quality image predictions.

We strongly recommend starting with an API, because it’s still very challenging to install and configure the open source libraries. That said, if you want to dig a bit deeper, the model zoo has many pre-trained models you can load up into caffe.  

Interested in image recognition? Check out the GoogLeNet.  

Want to play with automatic video descriptions? Check out Translating Videos to Natural Language.  

Want to play with face detection? Check out Face CNN descriptor.  

All these models have been pre-trained and simply need to be loaded up into caffe, then you can start throwing data at them!

Finally, if you want to dig even deeper, we recommend keras. It’s a very simple python library to start building your own neural models. But don’t mistake simplicity for lack of power: we’re using it to train new and interesting language models, and the simplicity enables us to focus on the algorithms and techniques.

With these services and libraries, you’ll be able to start identifying objects in images or videos in no time. Have a cool idea to build something using this new capability? Tell us at @fastforwardlabs!

-Micha

More from the Blog