• Zev Isert

5 Innovative Applications of Deep Learning

What is deep learning?

Deep learning is a subset of machine learning that hopes to imitate the way we as humans learn and process information. Deep learning requires vast amounts of data, with the ultimate goal of processing and recognizing patterns within the data.

An artificial neural network consists of layers of nodes - each representing decision variables. These nodes process and apply mathematical formulas to the inputted data, and if the information passes a certain threshold, the nodes “fire” and pass the information on to the next layer. This is similar to how human brains process information. The technology is extremely innovative, and can be applied to a variety of industries.

Unlike some other types of machine learning, deep learning works well with unstructured data; it is able to process unstructured and unlabeled data, meaning it can be trained using an unsupervised approach. Current common implementations include facial recognition tagging in photos, google translate and self driving cars.

Innovative applications of deep learning

1. Fake News Detection

With fake news currently creating unrest and political issues globally, many scientists have wondered whether deep learning could be applied to correctly identify and classify fake news stories. This could greatly help reduce misinformation.

An MIT paper detailed how it used deep learning’s pattern recognition abilities to find common characteristics, language patterns and phrases within fake news articles. Social media networks are already beginning to use this technology to identify and remove misinformation posted on their sites.

2. Solar Savings Potential

Google has recently created a project called Project Sunroof, using the image recognition capabilities of a deep learning network to create a model of a roof using satellite imagery and mapping data. The model accounts for trees, shadows and other things that might obstruct the sun to reach solar panels, and models the amount of sunlight a roof will receive.

For more information, Google has released a paper describing their methodology. After creating a model of a roof, Project Sunroof will estimate the amount of usable sunlight per year, the amount of room available for solar panels, possible power bill savings, pricing and financing options.

Project Sunroof is currently only available in the US, but the program shows how deep learning can be applied to help stop climate change and create efficient energy solutions. It allows people to easily get the correct information they need to make decisions.

3. Photo Enhancement

a. Pixel Restoration

Researchers at Google Brain trained a deep learning network to reconstruct images based on an 8x8 pixelated, low resolution photo. They call this process “super resolution”. The deep learning network was fed low resolution images, and asked to predict what the image would look like. Here you can see the results of its prediction.

This technique could be very useful in enhancing low quality images for scientific or personal purposes, and was not possible beforehand.

b. Colorization

Deep learning networks are also frequently used to recolor black and white photos. Currently, most grayscale photo restoration is done in a manual, painstaking process in Photoshop. Researchers and hobbyists have created projects such as DeOldify, created by Jason Antic which automatically restores and recolors photos, and Let there be color!, a deep learning network that recolors photos, created at Wasana University in Japan. The engines have even started to be able to work on black and white videos!

Here are some examples of colorization that the deep learning networks have created.

"Migrant Mother" by Dorothea Lange (1936) - Source

Zitkála-Šá (Lakota: Red Bird), also known as Gertrude Simmons Bonnin (1898) - Source

4. Creating Music

Open AI has created a deep learning network, called Jukebox, to create songs and works of music. While still not able to create fully polished music, the model is able to

create songs in a variety of different styles and genres, with some parts feeling eerily familiar.

Deep learning music generation without lyrics seems to currently be more realistic. Here is an example of Ragtime Music created by Tyler Doll in a project called DeepWave.

5. Fraud Detection

With the availability of personal information now available online, paired with online transactions, there are many opportunities for online fraud. Companies like Sift are capitalizing on this, and creating products to detect and eliminate fraud.

Deep learning networks can analyze massive amounts of transaction data to understand patterns within the data and find anomalies. These anomalies may pass through traditional rules based validations, but may actually be fraud. The process of analyzing all this data would take humans months or years to complete, and a deep learning network can complete the process much faster, and many times can even offer real time fraud detection. Companies such as Amazon and Paypal already use such networks for transaction fraud detection. Additionally, JP Morgan is also using deep learning for fraud detection, in a program called Deep X.