We are well and truly living in the age of AI and machine learning (ML) integration, with new innovations routinely revolutionizing both the way businesses operate and how we live our personal lives. Technology is endlessly improving the way we shop, the way we work and the way we interact with each other.
And, as AI and ML integrations become more commonplace, progress in a lesser talked about branch of AI, deep learning, is beginning to gain more attention. Similar to how ML made the leap from research labs into the mainstream, scientists continue to find new ways around our current limitations through deep learning networks capable of solving complex problems.
Deep learning, a branch of the ML family, processes data in a similar way to traditional ML methods, relying on algorithms and having the capability to self-modify. Deep learning models do not, however, require inputted data to be structured, instead relying on layers of algorithms, with each layer approaching the data with a different interpretation. This is why deep learning is sometimes referred to as hierarchical learning, as each algorithm layer ranks and categorizes elements of the unstructured data, similar to how a human brain – the inspiration behind networks – processes information.
The ability to process unstructured data is one of the many reasons deep learning is being used to develop the 'brains' of self-driving cars, as they need to be able to take in and prioritize massive amounts of data continuously to react in real time. The burgeoning self-driving car industry is currently one of the sectors investing most heavily in deep learning; practically every major automotive manufacturer has teamed up with a tech firm in an effort to be first-to-market with an adequate autonomous vehicle or mobility service. For example, earlier this year, BMW and Daimler combined their current efforts in the mobility services field to try and achieve "sustainable urban mobility for customers" faster than their competitors.
Some tech companies, such as Google's parent Alphabet, are even aiming to enter the fledgling market without the backing of massive vehicle brands. Alphabet's subsidiary Waymo is working on one of the most promising self-driving car projects and have decided to partner up with Canadian automotive parts supplier, Magna International, to help it achieve its goal of bringing the first car manufacturing factory which is 100% dedicated to the creation of level 4self-driving vehicles to fruition.
Research and Markets predicted the autonomous vehicle market is set to reach $30bn by 2023 due to the overwhelming demand for the safety and cost savings (for industry at least) these vehicles may be able to provide. So expect this field to remain one of the core hubs for deep learning innovation over the coming years.
Deep learning networks are also making waves in the healthcare industry, especially in image analyses. Branches of deep learning such as convolutional neural networks (CNNs) which were inspired by the patterns of connectivity between neurons in the visual cortex, have become extremely adept at analyzing visual imagery.
How adept, you ask? In January 2019, scientists at the National Cancer Institute's Division of Cancer Epidemiology and Genetics announced they had developed an incredibly cheap yet highly effective method of detecting cervical cancer. A new screening technique called automated visual evaluation (AVE) which was initially developed from Faster R-CNN, was able to detect precancerous cells at an accuracy rate of 91%, beating both pap smears and human expert accuracy levels.
Another simple, yet incredibly useful way deep learning is impacting healthcare is through the categorization of electronic health records (EHRs). Deep learning is already used in many natural language processing (NLP) programs, but EHRs present a uniquely complex problem, even for deep learning networks. Free text notes are often completed in a rush meaning they are messy, full of medical jargon, sometimes incomplete or filled out by multiple people, rendering them inconsistent. However, deep learning still represents the best method we currently have to analyze these records. A deep learning model developed by Google in 2018 was capable of predicting clinical outcomes, such as mortality and unexpected readmissions, better than traditional models after it had analyzed 216,000 patient EHRs across two hospitals.
Media manipulation and identification
In the age of "fake news", one use case of deep learning which has been all over the tabloids has been related to the progress that has been made in the video manipulation department.
The term "deepfake" first entered the world as the pseudonym of an anonymous user on Reddit after a number of faked pornographic videos were posted to the social media platform in 2017. Users utilized a form of deep learning, a generative adversarial network (GNN), to convincingly supplant the faces of some Hollywood actresses onto those of a more X-rated variety.
These were soon debunked as fakes and the UK government even took measures to allow for the prosecution of deepfake video producers. However, before that happened, the method had been refined by other users with one ultimately packaging the script into a program called FakeApp which users could download to make their own deepfake videos.
This technology has since been further refined and is fast unmaking the saying "seeing is believing". It has been responsible for a number of public hoaxes, such as when Argentina President Mauricio Macri's face was replaced with the face of Adolf Hitler. It also made news in 2018 when Get Out director Jordan Peele teamed up with Buzzfeed News to produce a realistic public service announcement video of former US President Barack Obama telling the public not to trust the internet.
The University of Washington also developed a neural network capable of creating the same effect in 2017:
In this case, researchers converted audio clips into realistic lip-synched videos by utilizing a neural network that can convert the sound from an audio clip into mouth shapes, even when the clips are older and not as a clear.
"There are millions of hours of video that already exist from interviews, video chats, movies, television programs and other sources," says Supasorn Suwajanakorn, the lead author of the paper Synthesizing Obama: Learning Lip Sync from Audio. "And these deep learning algorithms are very data hungry, so it's a good match to do it this way."
On the less dystopic side of media manipulation, deep learning applications are also helping us give history a much-needed touch up. For example, DeOldify uses GNNs to restore old, faded images to their former glory, as the image below demonstrates.
Its creator, software engineer Jason Antic, says he did it by "just training the same model to reconstruct images that augmented with ridiculous contrast/brightness adjustments, as a simulation of fading photos and photos taken with old/bad equipment".
Other teams, such as those behind "let there be color!", have managed to use CNNs to colorize black and white videos automatically:
And more recently, deep learning networks have even been used to amplify the crudest of artistic abilities. In March 2019, Nvidia Research unveiled an app called GauGAN which utilized GNNs to allow users to create photorealistic images from very basic sketches:
Computer vision is another area being upgraded by deep learning. Much has been written about the capabilities deep learning-enabled surveillance systems have given the Chinese government in recent years. AI chips and software created by Chinese firms such as Horizon Robotics and DeepGlint aid not only in facial recognition, but can also track the behavior of those it surveils, able to identify micro-expressions that precursor an act of violence. It has become a big component of the social credit system the region has launched, but the tech may begin to spread due to its usefulness anywhere there are large crowds and the potential for violence.
These are but a fraction of the ways deep learning networks are being utilized right now. Deep learning enabled AIs are powering surgical robots and helping surgeons make fewer mistakes in operating theaters, while others are helping astronomers generate photos of galaxies. Google Translate is capable of translating images of text within larger images into your language of choice in real time. Some deep learning networks are even helping geologists predict earthquakes sooner. The fourth industrial revolution is looming and suffice it to say that deep learning networks will form a pivotal part of it.