Artificial Intelligence (AI) Tech Digest - February 2017

Machine learning helps credit-scoring

ZestFinance, a machine learning and big data company, has developed a new credit scoring program, called ZAML. ZAML uses machine learning to judge credit worthiness and can be used by lenders and consulting services. The platform has been tested in collaboration with search engine company Baidu. Using the data Baidu has about people’s search habits and internet behaviour Zest discerned patterns which indicated whether or not someone would be suitable for a small loan. One of the factors Zest used was self-reported income matched against modelled income (a calculation of actual wage based on a user’s behaviour). The crucial information here is the discrepancy between the two income amounts. The less discrepancy the more loan worthy the individual. 

Google rates comment toxicity 

Google’s Perspective project is an API (application programming interface) designed to flag up ‘toxic’ comments online. The system uses machine learning models to score the likely impact of a comment on the conversation. ‘Toxic’ comments are comments that are likely to insult, denigrate or otherwise make someone uncomfortable. The toxicity of a comment is rated by percentage. The project is offering API access to developers. Partnering projects are being carried out by Wikipedia (to detect abuse aimed at editors), and by The New York Times, The Economist, and The Guardian – to moderate comments. 

GAN makes galaxies

Anh Nguyen and his team at Cornell University have taught deep-learning neural network (NN) software to generate high resolution, photo-realistic images of volcanos and galaxies. The work uses a technique known as a generative adversarial net (GAN) – pitting an image generating network against an image recognition network that helps both improve. The image generator attempts to ‘fool’ the image discriminator into believing that the image it has created is a real image. After the image, has been judged, the discriminator network is told which is real and which fake and uses that information to refine its future discriminations, as well as inform the generator – which never sees the real image –  how to improve its images’ realism. The eventual aim is that the discriminator will not be able to accurately judge which image is real and which fake. Ian Goodfellow, a scientist at OpenAI, says that GAN networks are more efficient learners than other NNs because they require less input data to create an accurate output. 

Facebook use AI to check for suicide

Facebook will be using pattern recognition technology to enhance its existing program to identify suicide risks and offer support and help. The AI system will monitor feeds from people that have previously been flagged for worrying content. The pattern recognition system will also identify posts that it believes are indicative of suicidal thoughts, which will then be assessed by Facebook’s community support team. 

Dyson to invest GBP2.5 billion in new UK research centre

It has been widely reported that Dyson, a British electronics company, has plans to open a new GBP2.5 billion research campus for robotics, AI, batteries, vision systems and other advanced systems. One of the first areas for research could be how to incorporate pattern recognition and decision making into the 360 Eye robot vacuum cleaner.

Baidu acquires Raven Tech

Baidu has recently acquired Raven Tech, a Beijing based AI startup that has created a smart home control centre, H-1. The price of the acquisition is not known. Raven Tech’s Cheng Lu will join the Baidu Duer (Baidu’s digital voice assistant) team to assist with product development.  

British Government promises a digital future

A recent paper published by the British Government emphasises the need to invest and nurture the digital economy of the UK. The paper outlines several areas of focus for investment including AI and robotics. The paper says the Government will award GBP17.3 million in grants to support the development of robotics and AI in UK universities. This will include GBP 6.5 million investment into the UK Robotics and Autonomous Systems Network to help translate robotics and AI research into practical applications.

System that builds code by taking it from elsewhere

DeepCoder, a machine learning system created by researchers at Microsoft and the University of Cambridge, is starting to solve  coding problems like those posed in coding competitions. DeepCoder codes by combining lines of code taken from other software in a process called program synthesis. Given a list of inputs and outputs for each code line, the system can learn which codes to combine to achieve the desired result. Currently DeepCoder can only solve code problems involving around five lines of code. 

Seeing a computer think

A series of beautiful pictures has recently been released by Graphcore, a machine intelligence company, that show an AI system thinking. The orderly structures in the pictures that resemble a coral reef seen from above, or a petri dish of various bacterial colonies, show the visual representations of the layers of calculations which machine learning systems go through to complete a process. As the computations race between data sets they create what appears in the picture to be luminescent lines between data centres.  The original aim of the work was to help the company visualise its Graphcore Intelligent Processing Unit’s (IPU’s) machine learning computations. 

AI takes on quantum systems

Guiseppe Carleo of ETH Zurich has devised an artificial neural network that could help scientists describe quantum systems. The network that he constructed with Mattias Troyer of Microsoft, was designed to reconstruct wave functions of a multibody quantum system – a set of probabilities describing how each state could be arranged. The proof of concept experiment tested the system on problems where the answer is known, and it was claimed to have performed better than any other multibody tool. 

Bayesian AI

Currently deep neural network AI systems seem to be in vogue. However, there are other methodologies approaching AI from different angles. One such methodology is called the Bayesian approach and is being pursued by start-up and DARPA-backed Gamalon, based in Cambridge, MA. It believes in creating AI through the scientific method of starting with a hypothesis and through experimentation refining it – a process called probabilistic programming. This process allows programmers to build AI in a way similar to how coders would build software, inputting specific code and editing as they go. It could also help AI researchers better understand AI decision-making and alter those decisions if they don’t match expectations. Gamalon’s AI system is able to translate languages and is currently being schooled to analyse raw data. 

Add this: