MENU

Synthetic data sets boost AI

Synthetic data sets boost AI

Interviews |
By eeNews Europe



In the booming world of artificial intelligence and deep learning, Neuromation’s CEO Yashar Behzadi sees two large elephants in the room that could tame the adoption rate of AI, one is the shortage of data scientists and the other is the shortage of properly organized datasets to train neural networks for specific tasks.

Neuromation’s CEO Yashar Behzadi.

It is not just the mechanics of creating neural networks that can be challenging, but data scientists must also organize and label training data so that each algorithm will be properly trained on its assigned tasks. Even for companies with enough data scientists on staff, this can make machine learning time-consuming and impractical.

Aiming to solve these two issues, Neuromation’s ambition is to create a community of AI developers, connecting them to several ecosystem partners such as data providers and service providers around a unique blockchain-enabled AI market place, along with shared tools to develop and train algorithms, including the use of synthetic data sets.

The Neuromation platform: an AI market place.

With its Neuromation platform, Behzadi wants to make it ten times cheaper and faster to develop AI algorithms than it is feasible today, making it easier for AI developers to connect with their customers or technical experts.

“There are about six million software developers who want to implement AI in the next 12 months, we want to empower those folks” the CEO told eeNews Europe over a phone interview, noting that a recent study from Tencent revealed that there are just 300,000 AI developers and other AI professionals worldwide, unable to fill the surging demand for new AI applications.


Discussing synthetic data, Behzadi describes it as computer-generated data that mimics real data. A simple example Neuromation puts forward on its website is the creation of huge data sets for object or facial recognition, based on computer-generated images where objects and faces can be rendered very realistically and manipulated to create endless variations in sizes, shapes, colour, light exposure, viewing angles and so forth. All that data comes readily

labelled for robust and accurate object classification, at a fraction of the time and cost needed for a human operator to sort and create specific data sets from acquired photos. Neuromation also sees robotics as a promising sector, where fully simulated environments can be more effective to train industrial robots at real world tasks.

Synthetic data for object recognition is exemplified
by computer-generated images rendering real world
products, here for the retail industry.

According to Behzadi, with synthetic data, companies only need 50% of their original, authentic training data to finish the formal training of their algorithms. The CEO argues that some AI applications such as object recognition could even be trained almost exclusively with synthetic data, cutting out on lengthy and expensive human labour.

“As a natural roadmap, we started with image data which is simple to model. We have industry experts in CGI effects working together with deep learning experts to improve our models. The next level of complexity will be to create dynamic data. Where the physics are known, you can create generalized models”, Behzadi explained eeNews Europe.

“We have done work in the financial space to mimic transaction data. It is not about the realness of the data but more about the model performance, its robustness and reliability to create data that can supplement real data” the CEO continued.


Behzadi is keen to emphasize other benefits of synthetic data, such as a lower bias than manually compiled data sets, which eventually makes the end AI application more robust. Rare things or occurrences can also be simulated, making the end algorithms more robust. One natural example is facial recognition for which Neuromation says it is engaged with several customers, not even a year after it was founded late 2017. The startup has raised $50 million from investors earlier this year.

Talking about blockchain, the company sees it as a necessity to establish a trustworthy chain of custody and proof of work, also enabling micro transactions among the different players in the AI market place Neuromation is fostering. The platform has been 9 months in development and will allow collaborative work as well as retributed AI knowledge sharing.

“Our best customers are the ones that are already in AI, such as facial recognition, they see the added value and they are the easy ones to convert. But there are other tremendous domains of expertise with a lack of AI backing. For example, you have treasure troves of unexplored data in the medical field” commented Behzadi. AI algorithms could be put to good use in digital pathology, to identify cancer cells or other health-related patterns. Here as in many other fields, Neuromation could help figure out the most applicable solutions to efficiently leverage AI. The company has rolled out several beta versions of its platform and plans to scale it next year.

Neuromation – www.Neuromation.io

Related articles:

Matlab accelerates deep learning applications on Nvidia chips

AI-powered software boosts facial classification for video law enforcement

Neural ASIC platform promises easy AI integration

Imagination launches flexible neural network IP

Ceva goes non-DSP with neural processor


Share:

Linked Articles
10s