Artificial Intelligence (AI) Tech Digest - April 2018

Scientists balk at idea of AI-enabled military systems

Over 50 AI and robotics specialists from around 30 countries published an open letter demanding a boycott of KAIST, the South Korean public university, on the grounds that it was planning to develop autonomous AI-enabled weapons, following the opening of its new Research Center for the Convergence of National Defense and Artificial Intelligence. The boycott call prompted a swift response by the university, that confirmed in ScienceInsider that it will not be developing “lethal autonomous systems and killer robots” but pointed out there are many other applications for AI use in the military. Whilst the boycott then ended as swiftly as it began, we can expect many more public spats over the ethics of AI-enabled weapons in the near future.

SenseTime scoops an extra $600 million in Series C funding

Following on from Series B funding worth $400 million last year, one of China’s leading AI businesses has secured an extra $600 million in funding, including a large chunk of money from Chinese web and ecommerce giant Alibaba. SenseTime is a significant player in the development of AI technologies for facial and image recognition, medical imaging, and autonomous driving. It is also working with Qualcomm to put AI on chipsets for smartphones.

AI doctor can see you now

Babylon Health, a UK-based company that has developed AI-based chatbots that can help people to diagnose their health problems, is working with Chinese IT giant Tencent to deliver personal health assessments, treatment advice and health records for users in China. The AI system uses descriptions provided by users, and their answers to questions posed by the chatbot, to identify illnesses and recommend (or not) a trip to a doctor. Its solutions also offer phone-based access to banks of medical professionals for swift diagnosis.

US approves diagnostic device that detects eye disease

The US Food and Drug Administration has approved IDx-DR, an autonomous tool that analyses photos of the retina for diabetic retinopathy. It delivers 87-90% accuracy. It can use photos uploaded by nurses, checks the photo is good enough then makes the diagnosis. It is not clear what happens in the 13% of cases where the diagnosis misses the disease, but presumably wider use will enable improved accuracy, and its use is expected to improve early detection in the 50% of diabetics that do not regularly visit an optician.

Fixing those damaged photos; or removing that unwanted spot

NVIDIA has created an AI-driven software technique that uses deep learning to reconstruct parts of images – for instance where photos have been damaged, or to realistically replace parts of images that are manually edited out of the photo – using computer generated replacements. They trained a neural network with 25,000 images, and a range of blemish sizes and shapes that enables the software to ‘fix’ areas of varying shapes and sizes.

Using AI to write code

Researchers at the Department of Computer Science at Rice University have developed an AI system – called BAYOU – that can generate Java APIs to assist programmers. It uses neural sketch learning to interpret descriptions of what a programmer wants to do and converts those into code. The AI was trained using a database of publicly available code snippets (accompanied by descriptions of what those snippets do).

The programmer supplies a draft Java method with gaps (missing code) for the AI to fill, and including a query saying what they want to do. The AI calculates a probability distribution of results and returns the most likely responses – hopefully including code that does what the programmer wanted.

Its capability is currently limited to one gap in the program though the Bayou website says this is not “a fundamental restriction”.

AI-generated videos

Alibaba is releasing an AI-based video creation tool that creates product-focused marketing videos using content on its Taibao online marketplace. It combines music with pictures, facts and animation effects to create short videos designed to drive engagement.  The music and the content are all selected by the ‘AI’. The service is initially being released for electronics and clothing products with versions tailored to other product types coming in the future.

AI to tell stories

Researchers at UC Santa Barbara have been working on a neural network using Adversarial REward Learning (AREL) that is capable of telling stories by captioning series of photos it is given. Its efforts were tested using Amazon’s Mechanical Turk (AMT). Although it was no Dickens, the AI system was able to fool the AMT human reviewers into thinking its stories were created by humans 3 out of 5 times.

Toughen up your AI

IBM has released what it calls its Adversarial Robustness Toolbox. AI based on deep neural network technology designed to recognise images is susceptible to malicious attempts to confuse or mislead it through the addition of adversarial noise – deliberate input modifications designed to produce a desired response from the AI system. These can be undetectable to the human eye. IBM’s Adversarial Toolbox is being released as an open source library for researchers and developers so that they can measure the robustness of their AI systems to such approaches, harden their deep neural networks, and to detect deception attempts at runtime.

And you thought DOOM was tough in its original incarnation…

According to interviews and analysis in The Register, researchers have trained AI to develop new levels for the computer games Doom, and Super Mario Brothers, using generative adversarial network (GAN) techniques. Their projects are still at the development stage, but the idea is that ultimately the AI will be able to vary the difficulty of levels making them much harder for humans to complete, and different every time they play.

Add this: