Artificial Intelligence (AI) Tech Digest - May 2017

Curiosity makes for better AI agents

Researchers at the University of California, Berkley, have developed an intrinsic curiosity model that can increase the ability of AI algorithms to work without a strong reward, overcoming a problem that some other AI systems face. In a demonstration, the model encourages the AI agent to explore its environment and aspects of the environment that affect it. In two game demonstrations, AI agents equipped with the curiosity model were able to learn to navigate their environment more quickly than those not equipped. In Mario Bros. the AI learnt not to get killed as it wished to continue to explore its environment. 

AI for crime protection

AIS (AI Surveillance) is a security platform created by Deepscience that promises to deter crime and protect shop-owners. The company claims that its system can use video footage to spot weapons, people concealing their identity, and other potential risks that shop-owners could face. When any threat is detected a warning is sent through to a centre where a human monitor can contact the emergency services if the situation demands it. The service is commercially available on a per camera basis with more than two cameras costing USD54.99/month per camera. 

Deep Voice mimics voices

Baidu’s Silicon Valley AI Lab has published a paper on Deep Voice 2, an advance on its previous Deep Voice 1 system that can mimic a single human voice in real-time. Deep Voice 2 is able to mimic hundreds of voices using a training set of half an hour of voice data from each speaker. The system generates speech by finding shared qualities between different voices and discovers what makes them distinguishable allowing it to imitate the individual speaker’s cadence, pitch, accent and pronunciation almost perfectly. 

Code can write code based on screenshot

Pix2code is an AI system capable of writing software code based on just screenshots. Created by Danish start-up Ulzard Technologies IVS, pix2code can create program code by the user feeding it an image of an existing program which it will then emulate. The codes are created by a neural network and machine learning techniques. Codes of up to 77% accuracy can currently be created. 

Lighting up the house with smart monitor

Lighthouse, an American start up, has announced pre-orders for its eponymous house surveillance equipment.  Lighthouse is an Amazon Echo-like device, designed to be placed in the house, with a camera that watches the environment around it for activity. The device links to an app that uses natural language processing to allow the user to ask it questions such as ‘did anyone walk the dog?’; it then goes through the activity captured that day and shows the user related footage. The user can also set up Lighthouse to send push notifications when certain events occur, e.g. when the user’s child returns home from school. Lighthouse says that the device is able to recognise who should and who should not be in the house and if it detects a stranger it will send a security alert to the user.

Human resources AI

Interviewed is human resources (HR) software that helps remove unsuitable candidates from the selection process. An employer looking for new staff chooses a set of tests and trials that assess the skills needed for particular jobs. The system then makes the applicant undergo these tests, helping the hirers see real-world job aptitude. The tests include basic skill tests in language, maths, customer service, sales etc. It also measures verbal skills by analysing speech answers to questions over the phone or on the computer. The software can also be used to simulate on-job tasks such as building spreadsheets, interacting with AI powered customers and so on. 

Neural networks moving to devices

Google has developed a deep neural network, called TensorFlow Lite, that runs locally on the user’s phone without needing to upload or download data to the cloud. The announcement at the company’s I/O conference came with a demonstration of a copy-and-paste function that is built into the Android O operating system – a new OS also announced at I/O – that automatically highlights phrases that the user may want to copy such as place names or addresses so that the user can simply double-tap the highlighted words to copy them. TensorFlow Lite is expected to be available opensource as part of TensorFlow project in late 2017. 

Allo turns selfies into cartoons

Google has released a new feature for its Allo chat app that allows the user to turn a selfie into a cartoonish sticker. The user can take a selfie and neural network software embedded in the app will convert the image into a cartoon alter-ego. To create an algorithm capable of this the Google team ran experiments using computer vision networks similar to DeepDream and discovered that some of the neurons in the network were good at focusing on things they were not trained to focus on. Artists were employed to draw common identifying features of human faces on which the neural networks were trained, matching the illustration against a selfie that matched that feature.  The feature is available on Allo for Android and is expected to be released for the iOS app too. 

Cisco to acquire MindMeld

Cisco has announced plans to acquire MindMeld, a company that’s developed AI natural language interfaces for applications and devices. Cisco says it has an interest in machine learning (ML) that MindMeld can provide for use in its network and cloud systems. It is expected that the collaboration will provide Cisco with new functional interfaces that can improve users’ interaction with existing products through natural language command inputs and others. 


Add this: