Artificial Intelligence (AI) Tech Digest - March 2017

OpenAI agents make a language

OpenAI has announced findings that its AI has developed its own language. The OpenAI team taught AI agents to develop their own language by placing them in a set of simple 2D worlds, and giving them tasks to achieve that could only be accomplished by communicating with each other. If they achieved a goal – such as looking at or moving to another location – they were rewarded. Before the agent takes an action it first assesses communications from other agents from the previous time step, as well as information about relative locations of objects and agents in the world. It then stores the information in a private recurrent neural network, giving it a memory. 
The language showed properties of being grounded (use of non-abstract, concrete nouns) and compositional (used multiword sentences to convey an idea such as location data). Problems encountered with the language were excessive length due to agents creating a single utterance and littering it with spaces to make meaning, and agents trying to use a single word to encode a sentence-long chunk of information. 

When a raven is a writing desk

Scientists at the École Polytechnique Fédérale de Lausanne (EPFL) have been revealing the limitations of machine learning-based image recognition systems by introducing small flaws in an input picture that causes an incorrect identification. The flaws were slight changes in the image’s pixels, that would be invisible to the naked eye but caused the AI system to categorise a joystick as a chihuahua, and a coffee-maker as a cobra. The implications of this kind of error become serious when deep learning methods are being used in zero-error-rate applications such as medical imaging. The code the researchers used to research the perturbation is available to the public to find a solution or carry out further study.

CycleGAN changing Monet into photos

University of California Berkeley researchers have created an AI program called CycleGAN that can convert paintings into near-photorealistic images. The system was trained with paintings by Monet and landscape photographs from an internet site. The system could then be used to convert pictures from Monet paintings into ‘photograph’-like images. They then applied the same process to other impressionist painters’ works. CycleGAN can also convert elements of a picture, such as turning input of two zebras in a field into two horses. 

People tracking

Hitachi, a Japanese electronics manufacturer, has demonstrated an image analysis system using AI to conduct real time people-tracking and detection via public security cameras. Hitachi claims its system will be able to track and find a person based on eye-witness descriptions, and track the individual through multiple cameras using just an image of the individual’s back. The company claims that its system uses over 100 characteristics related to 12 types of appearance such as clothes, age, sex, hair style and gait. Hitachi says its system can process the required information 40 times quicker than other systems. 
Another Japanese company, NEC, has also developed face recognition technology that uses feature point extraction technology and that can identify individuals in a crowd, or under varying environmental conditions. The technology has won its fourth consecutive Face in Video Evaluation test carried out by the US National Institute of Standards and Technology. 

Remembering in deep neural networks

Google’s DeepMind has published research that demonstrates the ability of deep neural networks (DNN) to remember old tasks after learning a new one. Catastrophic forgetting is a big issue, as deep neural networks overwrite previous knowledge when presented with a new task. To provide them with clues the people at DeepMind looked at the human brain and how it is able to integrate and layer new with previous information. One method cognitive scientists posit as a memory formation strategy is synaptic consolidation – connections between neurons are less likely to be overwritten if they have been important in previous tasks. DeepMind termed this weighted process Elastic Weight Consolidation (EWC). The company gave an AI agent Atari games to play sequentially. Without using the EWC algorithm the agent quickly forgets each game after starting a new one. Using EWC the agent is able to retain a large amount of information about each game long after finishing. 

Add this: