Why to merge neural networks is more favorable than to compete

Technology of merging several neural networks are open and available.

Students programmers learn to create and train primitive neural networks on the second or third year of education. Nevertheless, neural networks remain something difficult and unimaginable. There are two reasons: computing capacities and data.

Training of each network requires a large number of data. Data have to be relevant, connected, accurate, enough to work with. Relatively recently, collecting, marking and storage of a large number of data became possible. There were already on the Internet free datasets for training. But it is still difficult to find suitable, from the commercial point of view, data. Their cost is also too high.

Training of the neural network process is rather long. At the big complexity of network architecture, a large number of layers and perceptron, training can take days and weeks on a very powerful equipment. Of course, it costs much.

High costs of development and deployment of neural networks force to protect data and the trained networks from competitors. Because of it dozens of teams collect the same data worldwide, train neural networks with similar functions. Besides time and money for development is irrationally spent.

The team of the Israeli startup NeuroSeed has decided that it is better “Make love not war”. They create the platform which will allow to exchange already trained neural networks or to use others developments for the purpose of a further education or merge together. According to the tests were carried out by the team on text, audio, and graphics data, the merge of already existing networks allows to receive network with the necessary accuracy of 45–200 times quicker, than training the new. The economic benefit is obvious.

It is necessary only to wish good luck to developers and to watch their news. I hope that shortly the neural network will become as habitual as the smartphone.