Neuromation is building a critical component of the budding AI eco-system and plan to capitalize on the first-mover advantage: The Neuromation Platform provides an exchange and an ecosystem where participants can either contribute or purchase the components of an AI model.
The Platform will use distributed computing along with blockchain proof of work tokens to revolutionize AI model development. It will combine all the components necessary to build deep learning solutions with synthetic data in one place. Platform service providers, commercial or private, will provide specific resources for the execution and development of synthetic data sets, distributed computing services, and machine learning models, addressing the “three pillars” of AI in the previous slide. Each executed service (data set generation, model computation) or sold data piece (dataset, data generator) is accommodated by a reward, numerated in our currency, Neurotoken. The Neuromation platform will be running an auction model so the customers will negotiate prices directly with service providers.
Imagine a place where you can go and easily address all requests to acquire AI capability. A vendor will create the data generator for you, then a group of Neuromation Nodes will use the generator to quickly create a massive virtual data set. You can then select a set of Deep Learning architectures to train on that data. Then another group of Neuromation Nodes will do the training in record time.
The AI ecosystem is founded upon three pillars:
● Datasets – large datasets of structured (unstructured) data used for deep learning algorithms to learn a specific task.
● Computing power – a standalone server or a farm of interconnected GPUs designed to provide the necessary framework for training AI models with existing datasets to achieve the objective.
● Machine Learning models – a set of algorithms designed to process datasets in the Neuromation Platform environment.
Platform Price Setting:
The price for each service will be determined by the aggregate setting of the Neuromation nodes (price per unit of computation). Each node will have a minimum token price-floor setting. The minimal price floor can also be adjusted dynamically via an algorithm that will maximize the total tokens earned for the node. The Neuromation platform will determine the resources required for each requested task and select the most efficient node pool (minimizing price for the customer).
Between nodes “sniffing” out the market and customers hunting for the most efficient price, the Platform will find equilibrium in supply and demand.
Token Purchase for Clients:
In order to transact on the Neuromation platform, a client will need to buy tokens. To simplify the purchase mechanics, Neuromation will provide a client portal that will make token purchase a one-click process.
To facilitate liquidity and regulate the market, Neuromation will periodically buy Tokens back from the market.
An average retail set has 150,000+ items. A deep NN needs thousands of labeled photo examples for each item. Doing this by hand would take years.
In the Neuromation solution, synthetic data is used for training, and a small validation set is comprised of real, manually labeled data. In essence, this makes using synthetic data for a transfer learning problem: they need to reuse models trained on one kind of dataset (synthetic) and apply them on another kind of dataset (actual images). Their approach, however, has several important advantages that greatly simplify transfer learning in this case.
1. the synthetic training dataset is not given as part of the problem but rather generated by them – they can and do try to make it match real data.
2. even more importantly, the above-mentioned feedback loop between the model and the dataset makes it much simpler to perform transfer learning: they are able to tune not only the model but also the training set, a luxury seldom available to machine learning practitioners
In a current application of this approach, they use state of the art object detection and image segmentation models to recognize objects on the shelves of a grocery store / supermarket.
In further work, they plan improve synthetic data-generation pipelines with machine learning algorithms, leading to a true active learning framework: synthetic data is used to improve a model that is verified on real data, while at the same time the learned model improves the synthetic generation pipeline to further benefit learning.