When we first are introduced to deep learning, we see it as a better machine learning classifier. Alternatively, we could subscribe to the hype that it is ‘brain-like’ neuro-computing. In the former instance, we grossly underestimate the kinds of applications we can build with this. In the later instance, we grossly overestimate its capabilities and as a consequence overlook the kind of applications that are not general artificial intelligence, but applications that are more realistic and pragmatic.
It is best to look at applications of deep learning from the perspective of improving human computer interaction. This is perhaps the most natural approach. Deep learning systems do appear to have capabilities that approximate the capabilities of biological brains. As such, they can be most effectively used to augmenting tasks that humans or even animals have been employed to perform. It is important to remember that deep learning systems are very different from traditional symbolic computing platforms. Just as humans think very different from how a computer computes, deep learning similarly different.
However, deep learning systems are already intrinsically built from traditional computational technology. So that the tireless mechanistic efficiency inherent in computers are also present with deep learning. Computers are much more capable than humans in performing accurate symbolic computation and inference. Though, deep learning systems are not yet capable of performing complex symbolic computation. They are however by default already linked to this capability.
Applications built using deep learning seems to be straight out of science fiction. Here is a partial sample of some of the incredible applications that have been developed so far:
Photo Captioning for the Blind
Facebook has developed a mobile app that is able to describe a photograph to people who are blind. http://www.wired.com/2015/10/facebook-artificial-intelligence-describes-photo-captions-for-blind-people/
Realtime Speech Translation
Microsoft Skype is able to translate voice into different languages in realtime. Something straight out of the universal translator in Star Trek.http://blogs.skype.com/2014/12/15/skype-translator-how-it-works/
Automated Email Replies
Google Mail is able to automatically respond to email on your behalf. http://www.wired.com/2015/11/google-is-using-ai-to-create-automatic-replies-in-gmail/
Moodstocks (acquired by Google) is able to identify common objects using your mobile phone. http://www.slideshare.net/CdricDeltheil1/moodstocks-mobile-image-recognition-paris-tech-talks-6
Location Identification from Photographs
Google is able to identify the location of where a photograph is taken just my analyzing the scene. https://www.technologyreview.com/s/600889/google-unveils-neural-network-with-superhuman-ability-to-determine-the-location-of-almost/
Organizing Collections of Photographs
Google Photos is able to autmatically organize your photographs into collections with common shared themes. https://www.youtube.com/watch?v=JuFtW1PSYAU
Yelp is able to automatically classify photographs into different business relevant categories. http://engineeringblog.yelp.com/2015/10/how-we-use-deep-learning-to-classify-business-photos-at-yelp.html
A hobbyist is able to teach his car to self-drive in a few hours. http://www.bloomberg.com/features/2015-george-hotz-self-driving-car/
Music can be composed based on different composer styles. http://web.mit.edu/felixsun/www/neural-music.html
Painting based on Artists Styles
Painting can be created based on famous artist painting styles. https://nucl.ai/blog/neural-doodles/
Discovery of New Materials
New materials are discovered with the help of deep learning. http://www.nature.com/articles/srep02810
Playing Video Games
Google DeepMind is able to create video game playing systems that learn how to play well by just watching the game.http://www.wired.co.uk/article/google-deepmind-atari
Playing Championship Level Go
Google DeepMind has created a Go playing system that is able to learn new strategies by playing against itself.http://www.scientificamerican.com/article/how-the-computer-beat-the-go-master/
Face recognition is so common that it is no longer surprising.
Click-bait Headline Generation
A RNN is trained to generate click-bait headlines.
Colorization of Black and White Photographs
A system is trained to convert black and white photographs into color. http://richzhang.github.io/colorization/
http://demos.algorithmia.com/colorize-photos/ is a service that lets you try this out on your own photos!
Realtime Translation of Images of Text
Google has a mobile app that translates the text found in a photo into text that you can understand.
Swiftkey is building keyboards for mobile phones that make it easier and faster for you to type. http://www.slashgear.com/swiftkey-neural-alpha-predicts-what-youll-type-08408912/
Predict the Future
Well, that’s the claim by these folks at MIT: http://news.mit.edu/2016/teaching-machines-to-predict-the-future-0621
3D Object Classification
Learning the meaning of different hand gestures is likely going to be how we interact with devices that don’t have screens.
Converting Photos of People to make them Smile SmileVector is able to take an image of an image of a person and transform it into an image of the person smiling.
Human Like Conversation
Google has create a messaging application that has more natural conversational capabilities.. https://research.googleblog.com/2016/05/chat-smarter-with-allo.html
Augmented Reality – Face Tracking Baidu created a mobile app that is able to track faces using Deep Learning. The app overlays a 3D image over one’s face.
Warehouse Optimization A Deep Learning system is trained to learn an optimal way of pick and placing items in a warehouse. This system is faster than the more traditional operation research optimization approach.
Sketch to Search Sketch an image as a query to a visual search.
EEG-informed attended speaker extraction from recorded speech mixtures with application in neuro-steered hearing prostheses.
Accelerating Fluid Simulation
Leveraging convolution networks to create fast and highly realistic fluid simulations.
Amazon drives its personalization capabilities using Deep Learning. http://blogs.aws.amazon.com/bigdata/post/TxGEL8IJ0CAXTK/Generating-Recommendations-at-Amazon-Scale-with-Apache-Spark-and-Amazon-DSSTNE
Brain Tumor Detection
Results reported on the 2013 BRATS test dataset reveal that the 802,368 parameter network improves over published state-of-the-art and is over 30 times faster.
Reducing your Electric Bill
Google is using technology from the DeepMind artificial intelligence subsidiary for big savings on the power consumed by its data centers.
Amazon sponsored researchers used deep learning to analyze 3D scans of objects that their robot had to pick and replace.
Facebook is using Deep Learning to create more accurate and current maps from satellite imagery.
Identifying people through their voice.
Users may more quickly and accurately comprehend infrared images that have been colorized.
Taking a 3D voxel representation of a shape and a semantic deformation intention (e.g., make more sporty) as input and then generate a deformation flow at the output.
Sketch to Generate Realistically Photos
Convert face sketches to synthesize photorealistic face images.
Predicting Clinical Events
A RNN trained on time stamped EHR data from 260 thousand patients and 14,805 physicians over 8 years. The network is able to make multilabel predictions (one label for each diagnosis or medication category). The system can perform differential diagnosis with up to 79% recall, significantly higher than several baselines.
Skin Evaluation and Recommendation
Using Deep Learning to determine a customer’s “skin age,” identify problem areas and offer a regimen of products meant to address those issues.