In every application of Artificial Intelligence (AI) you need some form of training to get from the stage of dead code to a more or less intelligent machine. This can be a simple set of rules (think basic chess moves) to petabytes of image data that are annotated in detail. When the application is highly repetitive, e.g. a quality control step in a production line, you do not need many images to identify outliers or failed items. But when the surroundings get more dynamic (e.g. self-driving cars) it becomes exponentially more complex and you need some serious training. Other examples can be found in an earlier article I wrote on DeepBrainChain.

Training

There are many real life examples where AI powered cameras require training. Facial recognition under different conditions; lightning, angle, skin colors are all factors that complicate annotation. Other examples are object detection in satellite imaging or by industrial robots, medical image analysis, precision agriculture, automated inventory management, match analyses, the list goes on.

The big mammoth IT companies not only have unprecendented datasets at their disposal for AI training, but in some cases they also let their millions of users help annotate images or situations by simply tagging their private photocollections. They help training their face recognition or speach interpretation algorithms for free. These companies have both the hardware capacity and endless raw datasets to leave any potential competition behind. Data samples that are labelled correctly (i.e. low noise or error free) are gold dust in AI.

Imagine being a startup in this arena and you have an AI application on the cooker based on some brilliant idea. Only one thing missing at the start: a comprehensive dataset to train your app. This is not a small issue. Building a big enough dataset may be a very tedious and costly work and may drain too much resources from the young company.

Synthetic data

This is where Neuromation comes in. This company provides what is called “synthetic data”, i.e. digitally created data that mimics real-world sensory input but with guaranteed accuracy and perfect labels. It can be created faster, cheaper, less biased and more robust compared to endless series of life images. But furthermore, Neuromation intends to become the world’s marketplace for all kinds of synthetic datasets that can be used and modified for very different applications. Creating newer and more refined datasets along the way, cross feeding from other projects. There will be feedback loops between models and datasets to transfer to deep learning techiques. So not only the model is continuously fine-tuned, but also the datasets will grow in quality over time.

Platform

It all comes together on the Neuromation Platform. Participants can either contribute or purchase synthetic data, receiving or paying the platform’s token NTK. These can be traded directly on the platform exchange against BTC or ETH (and soon even directly for fiat). If you require a new dataset, you can commission it on the platform and accept the best bid (both in price and domain knowledge) from a range of specialist data generators. Once the synthetic dataset becomes available, you can choose the deep learning architectures that would train your model best. Workloads will be run on distributed computers, fuelled by the proof-of-work tokens.

Computation Incentives

Crypto miners can load up Neuromation computational nodes on their own hardware, temporarily stop mining to execute a Neurotoken batch at a price that will surpass Ether mining by a factor of 3-5x. Once the job is done, the normal crypto mining is resumed. Eventually there will be dedicated Neurotoken nodes mining NTK, migrated to their own blockchain.

The Experts Marketplace

Neuromation plans to support all relevant frameworks, deep learning techniques and libraries, applying converters where necessary. When new models or AI architecture is required, requests can be posted on the exchange and domain experts may respond and pitch their work. The platform will make a wealth of datasets available at a fast pace and affordable prices, even for startup companies. Judging from the high numbers of participants in their social channels and the large number of business and academic partners they signed up this past year, Neuromation is going to be the hot house for synthetic data and AI training, attracting specialists from across the globe, capable of implementing solutions in a wide variety of industries.

Like many in this bear market, NTK is trading way below the ICO price. But I think (just my personal opinion) that in the long run, the NTK token may represent a solid value in the AI ecosystem.

Stats & Refs

Total Supply = 60,000,000 NTK (note that CMC again posts a wrong number)
Top 10 wallets have ca 40% NTK
Over 16,000 NTK holders

The white paper makes a good read if you want to know more details, as does the references mentioned on the blog page. It will give you a better understanding of the need of clean, annotated sample data.

https://www.neuromation.io/
https://www.neuromation.io/neuromation_white_paper_eng.pdf
https://www.neuromation.io/news/

10 votes, average: 5.00 out of 510 votes, average: 5.00 out of 510 votes, average: 5.00 out of 510 votes, average: 5.00 out of 510 votes, average: 5.00 out of 5 (10 votes, average: 5.00 out of 5)
You need to be a registered member to rate this.
(1715 total tokens earned)
Loading...

Responses

  1. CryptosDecrypted

    Very interesting write-up Peter. I hadn’t heard of this project – your science background is no doubt very useful when assessing these kinds of projects. Synthetic data & AI training seem an ideal match….now I need to go off and play catch-up in this area!

    (0)