One of the central features of AI development is the various software frameworks needed to make it possible – and while significant rationalization of frameworks is beginning, the likely result will be several interoperable frameworks rather than one de facto framework, says ABI Research.
An AI ‘framework’ refers to a collection of libraries, interfaces, and tools created for generating AI models, such as neutral networks (NNs), which enable deep learning (DL). And there are many of them that serve different use cases.
ABI says it has identified and benchmarked the key software frameworks likely to form the backbone of any AI product development. TensorFlow currently leads the pack, but several other frameworks are emerging as potential contenders likely to shape specific use cases and applications.
These challenger frameworks currently lack the size and scale of developer community interaction but are undertaking intense rationalization around interoperability, in a significant move against the dominance of TensorFlow. And that’s good news, says ABI industry analyst Jack Vernon.
“This is positive for the AI technology ecosystem. No one framework is likely to serve every AI use case and segment of the supply chain well,” he says. “A multiplicity of interoperable frameworks will enable developers to better research and productize AI.”
The current market for AI technology has been so far shaped by proprietary technologies. This has led to a huge fragmentation of the AI applications development and has created confusion for developers and implementers of AI technology. As it has been the case with many industries before, including mobile devices ecosystems, desktop operating systems, and internet browsers, technology rationalization is a key milestone of any technology development. AI will not be an exception.
Frameworks are beginning to see rationalization around two factors, says Vernon.
“First, a few frameworks have either died off in terms of developer and community support (such as Theano) or have been updated to accommodate a greater breadth of deep learning techniques (such as Torch moving to Pytorch and Caffe moving to Caffe2),” he explains. “Second, there has been a number of frameworks governing bodies choosing to cooperate around a series of standards that will enable deep learning models to be exchanged between them. For example, the Open Neural Network Exchange will rebalance the framework ecosystem which is disproportionately dependent on the success of TensorFlow and its commercial backer Google.”
ABI Research has assessed the landscape of AI frameworks and developed a benchmark measuring several KPIs such as developer interest, GitHub commits, scalability, edge accessibility, hardware portability, hardware efficiency, productization, governance, future proofing, and reliability. The benchmarking scores found that TensorFlow followed by Caffe2 are the clear leaders, followed by MXNet. These frameworks have repeatedly displayed a commitment to open standards, and their respective commercial backers have shown a dedication to educating their respective developer communities.