Latest Advances of AI in Q1 2020

Artificial Intelligence, once merely science fiction and a distant dream for the computing world, is now very much a reality. Artificial Intelligence, or simply AI, is the term used to describe a machine’s ability to simulate human intelligence. Actions like learning, logic, reasoning, perception, creativity, that were once considered unique to humans, are now being replicated by technology and used in every industry.

Over the years, AI has evolved from computer vision’s baby steps towards recognizing grayscale handwritten numbers, it has scaled up to state-of-the-art tech being able to recognize faces, perform object detection and instance segmentation, render augmented reality and so much more. Natural Language Processing tech has similarly broken new grounds, with development of models trained on massive datasets, useful for question answering, sentiment analysis, and so on, emergence of novel speech detection and conversational AI technologies, and a whole lot more. The AI sphere has expanded in several horizons and new dimensions, leading to its adoption into multiple use-cases for a variety of tasks.

Here are some of the top advancements in Artificial Intelligence that have taken place in recent times:

Development of a novel powerful antibiotic using Deep Learning:

A deep-learning model developed by the team at MIT has identified a new antibiotic compound that has been successful at battling some of the world’s most resistant bacteria. The newly developed drug was picked out by a computer model that has the potential to screen more than a hundred million chemical compounds in just a few days. The researchers designed their model to look for chemical features that make molecules effective at killing certain kinds of bacteria. They trained the neural network model on 2,500 molecules, which included 1,700 drugs approved by the US FDA and a set of 800 natural products with diverse structures and varying bioactivities. In laboratory tests against five species of bacteria, the researchers found that eight of the molecules showed antibacterial activity, and two were particularly powerful. The researchers now plan to test these molecules further, and screen it further on the database they used for the process.

Current methods of screening new antibiotics being prohibitively costly and requiring a significant investment of time, this model is expected to be a huge boost to the world of healthcare.

Advanced Product Understanding and New Shopping Experiences:

The AI Team at Facebook has developed a computer vision model named GrokNet that seeks to redefine shopping by acting as an AI lifestyle assistant that learns people’s tastes and greatly simplifies the shopping process.

Being frustrated and indecisive due to there being a huge variety of products on sale while shopping is a common feeling. The model developed by Facebook uses state-of-the-art image recognition models to recommend products to buy based on a representation of personal tastes obtained by analysing items that one already owns. It can also generate a virtual replica of the object to enable visualization of how the item might fit in a room or on a person. You can see yourself wearing the clothes or accessories you’re considering buying to make a much better choice; trial rooms might soon become redundant!

The method uses an object detector to identify boxes in images surrounding likely products, matches each box against a list of known products, and keeps all matches that are within a similarity threshold. It predicts a wide variety of properties for an image, such as its category, attributes, and likely search queries. It also predicts an embedding (like a ‘fingerprint’) that can be used to perform tasks like product recognition, visual search, visually similar product recommendations, ranking, personalization, price suggestions, and so on. The model has been deployed on the Facebook Marketplace and is already reinventing the shopping experience.

AI Ability to Automatically create automatons of Business Processes:

UK based startup ZappyAI has accomplished the once daunting task of understanding and automatically creating automation's for processes done in enterprises. Back in the early days of automation, the technology was unable to connect the dots between what had been done in a process previously, with what was being done now; automation tools would require extensive coding and configuration without any cognitive ability.

The artificial intelligence powered system uses long term memory to identify business processes that can be automated. It can figure out the decisions and logic involved in the business process being carried out and generate analysis of automation opportunities within the organisation. ZappyAI plans to use this AI technology to run the backbone of organisational process analysis, saving companies countless of manual hours taken to understand and digitize business processes.

Tiny AI — AI that can run on consumer devices

Some of the biggest technology companies in the modern era such as Google, Amazon, Apple and IBM have developed technology that enables AI algorithms to run directly from people’s phones and other consumer devices without needing to interact with cloud services, which would traditionally be required in the absence of powerful computing hardware. This tech, dubbed ‘tiny AI’, researchers have shrunk the size of existing AI models via a process known as ‘knowledge distillation’ without losing any of the algorithm’s capabilities or performance speed. The benefits of this technology include zero latency because of no communication between the device and the cloud, as well as fewer issues with privacy.

Reduction in the Carbon Footprint of AI

The amount of computational power required for training deep learning AI models is immense- a certain report by the University of Massachusetts, Amherst estimated the resulting carbon-dioxide emissions to be about 626,000 pounds in weight, on average. This is equivalent to five times the amount of CO2 produced in a lifetime of an average car in the US. This issue gets worse in the deployment stage of a model, with involvement of multiple hardware platforms with different properties and resources.

Researchers at MIT have developed a network to tackle this issue called a ‘once-for-all’ network, which reduces the carbon emissions to about 1/1300 of the total emissions being produced by traditional models. The researchers built the system on a recent AI advance called AutoML (for automatic machine learning), which eliminates manual network design. The single, large ‘once-for-all’ (OFA) network serves as a ‘mother’ network, nesting a massive number of subnetworks that are sparsely activated from the ‘mother’ network.

Learned weights by the OFA network are shared with all subnetworks, which leads to the subnetworks being essentially pretrained, in the process. At inference time, each subnetwork operates independently with its derived weights, with no additional training required.