Thank you for Subscribing to Apac CIO Outlook Weekly Brief
The Woes of Machine Learning Fragmentation and the Solution to it
Machine learning has made a place in people’s daily life through the use of smart devices they interact with as well as its increased adoption in commercial and industrial sectors. Although machine learning capabilities are being applied to social media platforms, IoT devices, cameras, and automobiles among others, the speed of innovation is leading to fragmentation, which may result in stalling.
Check This Out: Top Banking Companies
Fragmentation is a problem that plagues many industries that either lack the standards or have to comply with various competing standards, if not judged accordingly; it can stifle innovation crippling the emergence of new technologies.
In machine learning, format fragmentation increases the time needed to create, train, deploy neural networks, thus, affecting developers, data scientists and researchers involved.
The benefits from all of the features offered by the various frameworks and formats that a neural network is trained from may be monumental, but it makes transfers between frameworks difficult causing fragmentation.
Format fragmentation can be solved by creating a single transfer standard that is compatible across various platforms. A standardized file format across the industry could be a “PDF for neural networks” and would allow researchers, data scientists, and developers to transfer their neural networks between various frameworks and to numerous inference engines without spending extra time.
Seeing that a standard is undoubtedly needed to solve the fragmentation problem, Facebook and Microsoft will create an open-source format called ONNX (open neural network exchange) for AI frameworks, and also, The Khronos Group, a global consortium, is focused on developing a standard called NNEF (neural network exchange format). It is unclear whether a single standard format can address all the demands of the machine learning industry, but it is clear that there is an imminent requirement for neural network standards, and the two groups are committed to making sure the AI industry make its next leap.