The development of an international team of researchers will increase the speed and quality of recommendation systems, search services, online translators and many other software packages that use machine learning algorithms. In addition, the innovation will allow you to take the load off user devices (smartphones, tablets, computers) that are involved in the process.

An international research team has developed a new technique to enhance machine learning efficiency, with the potential to accelerate AI systems including recommendations engines, translators and more. By optimizing data transmission from end-user devices to servers, the method slashes communication bottlenecks holding back state-of-the-art neural networks.

As models and datasets grow exponentially, distributed training demands can overwhelm networks, slowing iterative improvement. Scientists from the Moscow Institute of Physics and Technology tackled this by synergistically blending three prevailing approaches to minimize expensive server exchanges.

Essentially, the new hybrid strategy transmits only partial dataset chunks to reduce payload volumes. It then exploits data similarities across nodes to further limit transfer frequency. Finally, preliminary local model updates precede synchronization to reduce lag. Experiments revealed this tripled throughput acceleration.

Lead author Alexander Beznosikov explained, "With increasing data and model size, more parallel computing is crucial today. Meanwhile distributed methods have a huge communication bottleneck. Our combination of techniques attacks this from all sides."

By smoothing data flows, the tactic relieves pressure on both server and user hardware during machine learning. This enables advanced neural architectures once impractical due to crushing network demands or unwieldy client computational expenses.

The researchers share the timely innovation as AI permeates standalone devices and services worldwide. The breakthrough may democratize access by making complex intelligence viable on modest smartphones or processors. Quicker iteration could also translate models to new languages or domains faster while optimizing recommendations and translations.

In a machine learning context bogged down by sprawling connectivity strain, the team's shrewd hybrid promises to uplift services through flexibility impossible for conventional networks. Their inventive traffic control for modern data pipelines may put higher-order AI within instant reach across countless emerging applications.

"The essence of our method is that on one of the devices — conditionally, the main one, some kind of server — the data should be in some sense similar to those that are available throughout the network. At the same time, the data on all other devices can be very heterogeneous," the scientist explained.

According to him, the implementation of this method allows you to speed up network communications ten times compared to basic algorithms and about twice as compared to the most advanced of them. In addition, the algorithm is good because most computing operations take place on the server. At the same time, user devices (phones, tablets and computers) remain unloaded and, therefore, can safely perform their direct tasks.

This method correlates with one of the most promising machine learning technologies — Federated learning. This technique implies that the data remains on users' devices, and the model is updated on the server by aggregating trained models from various devices.

Alexander Beznosikov stressed that during the research, the new method was tested on simple experimental tasks. In the future, scientists intend to test it on more complex software systems. Including on language models — artificial intelligence systems that are used to predict the next words and phrases based on the previous ones.