The deep learning technology market is booming with innovation. However, differently from other enterprise software trends, we are seeing as much innovation in deep learning coming from big corporate labs as from startups. At the moment, the deep learning cloud platform market is dominated by the big four cloud incumbents: Google Cloud, AWS, Azure and Bluemix. Recently, I learned about Floyd, a startup that is attempting to change that landscape.
Graduated from the most recent Y Combinator class, Floyd provides a series of native cloud services for the execution and management of deep learning models. The platform uses a cloud runtime for over 10 deep learning frameworks such as Caffe, TensorFlow or Chainer while also providing the corresponding management nad lifecycle automation tools. Additionally, Floyd is planning to build a marketplace of algorithms and curate dataset a la Algorithmia. Certainly an ambitious vision.
While offers such as Floyd are certainly welcomed in a market dominated by incumbents, the big question mark remains to asses whether this type of standalone platform has a long term chance against bigger and also very innovative PaaS competitors. Personally, I am bullish about the opportunity of standalone deep learning cloud platform such as Floyd but entering a young market already dominated by companies such as Microsoft, Google, IBM or Amazon is far from trivial. In the current market, the deep learning PaaS incumbents has some very well known advantages but also some frustrating limitations.
Key Advantages of the Deep Learning PaaS Incumbents
From an enterprise perspective, there are certainly very tangible advantages in the deep learning offerings of platforms such as Watson Developer Cloud, Google Cloud ML-AI APIs or Azure ML-Cognitive Services. For starters, those services are delivered as part of broader cloud platforms that include capabilities in areas such as data integration, storage, messaging, APIs, etc which enables the implementation of sophisticated deep learning solutions. Additionally, those platforms expose really sophisticated deep learning models as simple APIs [ex: Microsoft Cognitive Services, Watson Developer Cloud, AWS Lex…] which can be easily integrated into third party applications. Global availability, established enterprise SLAs and a robust partner ecosystem are also some of the key advantages of the incumbent deep learning cloud services. However, now everything is rosy when comes to these technologies and there are some interesting challenges that could open the door for startups in the space.
Key Limitation of the Deep Learning PaaS Incumbents
Extensibility , on-premise integration and interoperability with open source deep learning frameworks are some of the key weaknesses of the deep learning platforms provided by Google Cloud, AWS, Azure and Bluemix. Excepting, Google Cloud ML (which supports TensorFlow programs) the deep learning services provided by the cloud incumbents are currently unable to execute models authored on popular frameworks such as Caffe, Theano or Torch. That limitation contrasts with Floyd’s ability to support over 10 deep learning stacks. Extensibility is another challenge of the incumbent deep learning services as those services typically limited themselves to allow the execution of custom R or Python scripts. Finally, the big cloud deep learning platforms are currently unable execute on on-premise environments with the exception of some creative work Microsoft is doing to support Azure ML on Azure Stack.
Does Floyd has a Chance?
Absolutely! The unique characteristic of the enterprise deep learning market on which a lot of the innovation is being delivered as open source frameworks (Tensorflow, Torch, Chainer, Microsoft Cognitive Toolkit…) creates an interesting opportunity for cloud services in the space. Additionally, deep learning algorithms and data marketplaces in a nascent segment in the market. How could deep learning startups platforms like Floyd become and stay relevant in a market already dominated by incumbents? We will discuss some ideas about this in the next post…