Rivals have been left trailing in the wake of Google’s TensorFlow, whose framework is clearly positioned as the dominant open-source AI software system. This market monopoly has driven Microsoft and Facebook to form an alliance of their AI software systems, acknowledging that as individuals, their challenges could end in futility.
AI software is powering the rise of machine learning across the business world in applications from product recommendations to chat bots. In a bid to take on TensorFlow, as well as in tacit deference to the slow progress of their in-house Cognitive toolkit, Microsoft is to put less emphasis on its Cognitive ToolkitOpens a new window AI software and instead will help develop Facebook’s more widely received PyTorch offering.
Where previously Cognitive Toolkit competed against PyTorch, Microsoft’s Chief Technology Officer, Kevin Scott, says that while the company has changed its strategy and will help develop and code Facebook’s system, Cognitive Toolkit will still be available as a standalone product and will receive updates.
Unmatched Machine Learning
TensorFlow has dominated AI software since its launch as an open-source project by Google in 2015. The software has been widely used in Google projects including speech recognition, Google photos and search. It’s also the framework behind Gmail’s auto-compete system, Smart Compose, which suggests sentences when writing emails.
TensorFlow has brought deep learning (described as “machine learning on steroidsâ€) to the mass market, allowing companies across the economy to use advanced machine learning to power their offers, for example facilitating product recommendations for business, as does Amazon.
TensorFlow is also strong in speech recognition, enabling the development of customer service chat bots. Rival systems, of which there are plenty, have struggled to match its depth and ease of use, although no market share figures are available.
Microsoft launched Cognitive Toolkit in 2016 as open-source software on GitHub, but is understood to have made limited headway, and Facebook launched PyTorch later that year. While Microsoft’s Cognitive Toolkit is known for its powerful speech recognition capability, PyTorch has had greater uptake by developers and excels in other areas of AI.
Microsoft and Facebook have already workedOpens a new window together on AI. In September 2017, they introduced the ONNX system, allowing developers to export models trained on one AI framework to another. This was designed to reduce complexity and allow interaction between the many AI frameworks in use, such as exporting a model in Cognitive Toolkit to PyTorch.
Facebook says this is useful as AI projects often begin using one framework but later need the functionality only available on another. ONNX simplifies the work involved in moving models between systems.
A Whole New Business World
These increasingly complex systems are expediting advances in AI and machine learning to businesses, allowing them to create their own data science and developer teams and build in-house AI applications.
The democratization of AI is gradually reshaping the business world as functions from customer service to sales recommendations are automated using one of the AI software systems.
AI worksOpens a new window by analyzing huge amounts of “training†data to identify patterns. These patterns are then tested against data the AI system hasn’t previously analyzed. Once the patterns are validated, the AI is strengthened by being fed more data.
An AI may be fed thousands of images, for example, and shown that some of them are cats. The AI will look for the patterns that distinguish cats from other animals. The system will then be tested with other pictures which are marked up as cats or not cats to measure accuracy. Then more cat images will be fed in the system to improve its recognition.
The system for speech recognition works similarly. Companies can analyze hundreds of thousands of call-center telephone chats to uncover which keywords and approaches are the most successful.
One of the challenges for small and mid-sized businesses is obtaining datasets large enough to identify valid patterns. For this reason, AI tends to favor the biggest retailers, manufacturers, media companies and financial services companies.
Businesses increasingly see themselves as data-centric with the collection of data as a core function, and aim eventually to put their data to use with pattern recognition AI.
TensorFlow is increasing its grip on the market. As more developers use it, more will be trained in the software, increasing the likelihood of its dominance as a platform.
Other systems will need to specialize in different areas if they are to secure their position among the AI frameworks, and Microsoft appears to be one of many to have ceded ground in this area of AI development.