Ashmanov Neural Networks Chooses SONM to Train Its Image Recognition System

The company Ashmanov Neural Networks develops machine learning systems, for data analysis. With the help of the SONM decentralized fog computing platform, the company was able to acquire the resources it needed for training deep neural networks.

Ashmanov Neural Networks has become one of SONM’s first customers. Pilot testing showed that a SONM customer can rent computing power five times cheaper in comparison to the cost of the services of cloud providers.

The challenge for Ashmanov Neural Networks

The organization of “forward” (image and object recognition) and “backward” computing (neural network training) requires a flexible approach in the use of computing resources. Requirements for specific configurations and volumes of memory are not the same, and a specific cluster configuration or specific OS may be needed.

Renting capacity from “centralized” cloud platforms, especially if the task involves GPU computing, is not always optimal in terms of either cost or hardware/VPS specifications.

The customer approached SONM with the following task:

Ashmanov Neural Networks needed the functionality of an application for image analysis to be tested. The application would determine whether an uploaded picture or link leading to it (URL) contained content matching 18+ criteria.

Services like this are in demand by regulators (the Russian media watchdog) and Russian providers for automatic blocking of unwanted content. It was decided to carry out testing on hardware provided by the SONM platform.

The test images had resolution of 800х500px.

Solving the problem with SONM

Decentralized fog computing and the possibility of choosing the required resources at the required time for the required price on the SONM marketplace allow computing power for machine learning to be rented far cheaper than from big-name vendors. There is no increase in the time needed to process the task, thus, SONM platform successfully allows using its functionality for machine learning.

The customer began the task independently on the hardware provided. Guided by SONM developers, the customer succeeded to:

  • install a Linux distribution;
  • run the SONM client application on their system;
  • run their application (server) in an unwrapped Docker container.

The whole process caused no difficulties.

Testing of the application with preliminary neural network learning took two days, during which the application processed various images and gave a result from zero to 100% — an indicator of the probability that a given photograph contained 18+ content.

“Although machine learning tasks are not as demanding on resources as, for example, rendering, the efficiency and especially the cost of using miners’ GPUs for ML on the principles of decentralized computing without intermediaries is not only justified, but also convenient,” explains Eugene Manaev, Product Owner at SONM.

In the course of testing, one of the partners connected a mining farm to the SONM platform.

Specifications of the hardware made available to Ashmanov Neural Networks by SONM suppliers

8Gb RAM and GPU NVidia 1070 8Gb x 6 units.

Result

As a result of image analysis the application calculated the probability of 18+ content on a scale from 0 to 100%. It took one second to analyze a single image, and it scales to 600 images per second on one typical worker with 6 NVidia 1070 GPU cards. On a single NVIDIA GPU, meanwhile, 25 images can be processed per second at a resolution of 800х500px.

The possibility of training and using neural networks and computer vision services on SONM hardware is proven, and the potential costs of analyzing images using mining rigs have decreased by at least five times.

The client received the service free as part of active testing of GPU computing on SONM.

Potential for co-operation

Based on the results of pilot testing, Ashmanov Neural Networks has begun a new commercial project based on SONM computing resources: a neural network collects information from online sources about people in photographs.

This project is in demand, for instance, for bank scoring, determining the trustworthiness and creditworthiness of the client, for tracking down criminals, for marketing tasks in advertising and trade.

“The test project was completed in June. Until the end of the year we are planning to implement a new project on a larger scale. Our technical specialists are interacting with developers from Ashmanov Neural Networks for the launch of SONM hardware. It will be a full case: training a neural network and issuing the results — all on our hardware”, says Alexei Krotov, Key Account Manager at SONM.

The partnership between the SONM platform and Ashmanov Neural Networks allows AMD video cards to be used for neural network learning. This is a unique opportunity for both companies, since none of the world’s neural network libraries currently supports AMD cards — except Ashmanov Neural Networks’ PuzzleLib library. At the same time, AMD video cards are popular among miners who rent out the computing resources of their rigs on the SONM platform.