IBM says it has broken new ground and achieved a new deep learning record by discovering a way to accelerate data processing, which is crucial to the technology’s functionality.
This breakthrough regards the time it takes to train deep learning platforms to recognise images, an ability at the core of training technology to process information like humans.
The time required by IBM to train a ResNet-101 neural network was a record breaking seven hours, trumping Microsoft’s ten day process taken to train the same system.
In recent testing the tech giant achieved a record recognition accuracy rate of 33.8% when processing 7.5 million images, beating the previous 29.8% record held by Microsoft. This statistic marks a 4% stride forwards, which is significant compared to previous advances of 1% or less.
According to IBM these results were achieved by harnessing the power of 64 servers and hundreds of GPUs, an innovative advancement on the process of using a single server that was employed until recently.
One server was used previously because it is crucial that the massive flow of data that involved remains in sync throughout the process, but for the recent deep learning record IBM gained synchronicity across its numerous servers.
To achieve these results IBM used a clustering technology that acts as an overseer of the many processors that are involved. This means that all of the processors could function efficiently, each managing a set amount of data for the optimum outcome.
IBM is involved throughout the spectrum of technologies that are currently top trends within the tech industry and beyond. Blockchain for example has also become a prime focus area for the giant, working with the likes of Maersk to bring the cutting edge technology to shipping and logistics.
The company is also partnered with the London Stock Exchange on a distributed registry project to provide a blockchain solution to small and medium sized (SME) businesses.
This article is from the CBROnline archive: some formatting and images may not be present.