Open Innovation: Five artificial intelligence research projects you can access right now.
Over the past few years, AI has been a overwhelming topic with some calling the impact ‘the most important general-purpose technology of our era’ (HBR, 2017). For most people how it actually works is still unknown. For analysts and data scientists creating these new systems, the area is an elegant intersection of science and art that resonates with the free spirit of the hacker and the creative spirit of the artist.
It has become commonplace for data scientists to open source their research, enabling others to build upon their findings to grow our collective understanding of predictive analytics technology and to allow practitioners to apply their research to solve problems in today’s world.
Today’s practitioners range from major technology companies, such as Google, Facebook, Adobe, Nvidia and Uber, who build their own proprietary and open source technology into their products, to small consultancies who use the latest advances in AI research to solve complex problems for businesses small and large alike.
To help decision makers better understand the capabilities of the technology, we’ve put together a list of recent interesting applications of AI.
1. Generating Celebrities’ Faces
By training a system of Generative Adversarial Networks (GANs), models can be trained to generate images from scratch. Simply put, GANs apply two separate artificial neural networks to work with and against one another to optimise a generative output, such as an image. One network is tasked with generating the image from the pixel level and the second network is tasked with recognising that image. These two processes work harmoniously, optimising against one another to create the desired output.
A recent advancement with this technology from a team of researchers at NVIDIA (2017) used a new training methodology to improve the quality, stability and variation of images of celebrities generated from the CelebA dataset. By training the model with increasingly large samples from the training set (4x4 pixels through to 1024x1024 pixels), the GAN is able to produce good quality images.
The full study and links to its repository can be found here.
2. Generating New Design Recommendations
Another recent application of GANs has deployed the method in a system to generate items of fashion aligned with the tastes and preferences of a given user, given their past purchase and interest data. This study by researchers from the University of California, San Diego and from Adobe has used deep neural networks to identify features of existing items of clothing (such as colours, patterns, types and other details) to train a GAN to generate items of clothing that have similar features to the preferences of a user and also match the description of a desired outcome, such as a blue dress.
The full study can be found here.
3. Core ML and Apple’s On-Device Facial Recognition:
With the release of iOS 10, Apple introduced Core ML, a suite of software frameworks that allowed developers to integrate various trained machine learning models into apps, vastly expanding the capacity of iOS applications. With this, Apple also began to implement facial recognition models through its Vision framework to improve things such as photo archiving and camera focusing.
More recently, Apple has made huge advances with this technology to allow these deep convolutional network facial recognition models on their high-end mobile devices in a way that these intensive models can be run quickly and efficiently and not interfere with the numerous other simultaneous functions that a device is likely to be running. Core ML and Vision has made this possible. A full explanation can be found here.
4. Predict Likes From Instagram:
A recent study by students at the The University of Edinburgh investigated the applicaton of convolutional neural networks (CNNs) and Natural Language Processing (NLP) to predict the number of likes on an instagram post. First, a dataset was generated from a large scrape of 972 instagram profiles and 16539 images which took not only the images but key data points such as number of likes, caption text, comments and followers. NLP was used to analyse key features from the caption and comments, and this information was analysed using a CNN alongside other elements, such as features extracted from the images themselves and data points such as number of likes and number of followers.
The combination these different model types yielded mixed results, and does show that there is room for improvement, but it is effective as a proof of concept. The full study can be found here.
5. Fashion-MNIST
MNIST is a large set of tagged images of handwritten digits that have formed the basis of countless studies and models in the Optical Character Recognition (OCR) field. A recent study has developed a drag-and-drop replacement set for MNIST which, rather than being an imageset of hand-drawn digits, is a set of 70,000 grayscale images of fashion items across 10 categories designed for MNIST models to be applied to it for classification processes.
The paper and full set of images can be found here.
About us and other sources
At northraine, we specialise in connecting cutting-edge research like this to organisations of all shapes, sizes and specialties. We are a B Corp with the purpose to ‘recondition the human condition’ and 20% of our work is AI for pro-bono social ventures. Have a chat to us and we can have some cake and walk through some code.
If you are still adventurous one of the best sites to find new code that is linked to new gitxiv (http://www.gitxiv.com/), the ‘hacker news for data’ Datatau (https://www.datatau.com/) and also reading through the Machine Learning subreddit (https://www.reddit.com/r/MachineLearning/)
References:
Harvard Business Review (2017). The Business Of Artificial Intelligence. Retrieved November 23, 2017 from https://hbr.org/cover-story/2017/07/the-business-of-artificial-intelligence
Google Trends (2017). Artificial Intelligence. Retrieved November 21, 2017 from https://trends.google.com/trends/explore?date=all&q=%2Fm%2F0mkz
