Machine Learning Platforms and the Cloud Computing Era

There is a battle looking for the next platform. Specifically the next platform that attracts developers. Right now, the race is on to create machine learning platforms that attract customers to use the latest generation of tools and commit to a cloud platform with machine learning advantages over the other cloud platform. From an open model, the leaders here are Amazon, Microsoft, and Google.

What makes this next battle interesting, is it will be the first time the platform truly moves from the native device to the cloud. When we have spoken of platforms before, we have mostly talked about them running on a local device, where the main brain (CPU) lives. This conversation is now shifting to cloud platforms where millions of brains (CPU,GPU) live. Machine learning will set at the center of this new cloud platform battle and we are just getting started.

The big three attacking this right now in Amazon, Microsoft, and Google are racing to build network training models, proprietary software, backend services, and in some cases even proprietary hardware in order to differentiate their offering and attract companies to build solutions on their cloud platform. While the cloud platform battle has been going on for some time as we think about Amazon AWS, and Azure, what’s new to the angle is the machine learning technology which will be the most important differentiating factor to anyones cloud computing platform going forward.

Without diving too deep on the common learning techniques, it is important to note two hurdles I have been watching as the leaders try to jump over. The first being labeled data. Most data, particuarly on the visual side for things like photos and videos, require some form of tagging so a computer knows it’s looking at a picture of a dog, cat, tree (what kind of tree), street sign, pedestrian crossing the street, parked car, etc. To train networks today, massive amounts of labeled data is necessary. When it comes to things that are already text based it gets a little easier. Hence things like predictive text on your keyboard, or search query, are a bit easier to acquire data and train a network. The holy grail is to develop a solution that allows a computer to be trained with unlabeled data, and even learn from that unlabeled data to be more accurate from larger samples to learn from.

The second is communal data. This one is tricky because of issues around privacy in many pure consumer use cases where machine learning and AI will be mainstream someday. The best way to learn from things like autonomous cars, or the data we generate on our notebooks, tablets, smartphones, smartwatches, etc., is to gather all the user data together and train the network from the community using the service. Obviously, people are sensitive to the idea their car is being tracked and data is collected, or what they type on their keyboard is being tracked and sent to servers. However, the reality is the best way to train a network will be from its community of users so dealing with this prvately will be essential.

All three companies I mentioned take the needed steps to ensure privacy, and some take more precautions than others. But the techniques they employ are crucial for the end result to still be useful data to build better network models.

A recent blog post from Google, highlights a few things the are doing that struck me as particularly interested in both areas I mentioned above. Google is using an approach of federated learning to make communal training faster so users can enjoy the benefits quicker. This paragraph provides some brief detail:

It works like this: your device downloads the current model, improves it by learning from data on your phone, and then summarizes the changes as a small focused update. Only this update to the model is sent to the cloud, using encrypted communication, where it is immediately averaged with other user updates to improve the shared model. All the training data remains on your device, and no individual updates are stored in the cloud.

Google mentions a couple of interesting roadblocks today they are looking to overcome and both have to do with speed of training. The primary one they addressing is bandwidth latency. This process they are employing is a cross between on device training, then taking that training up to the cloud to better train the computer. In order to do that with the slow and unpredictable upload speeds we are accostomed to, they compress the file, encrypt it, then uncompress on the server side to allow the data to train the cloud network. The main goal is to quickly, and in near real-time, train the network using data collected on the device in use. This sentence nails what they are after and the hump they are trying to get over:

But in the Federated Learning setting, the data is distributed across millions of devices in a highly uneven fashion. In addition, these devices have significantly higher-latency, lower-throughput connections and are only intermittently available for training.

It really is an interesting approach, and I’ll be curious how Amazon or Microsoft respond when they are not as privey to as much end user data as Google, particularly on device smartphone data. Regardless, both those companies are working to create cloud platforms, similar to Google, with the goal being for businesses to run their services on which consumers will use on their smart devices.

Where’s Apple in This?
No doubt many of you are asking where Apple fits in all of this. Apple seems to be taking a different approach and Apple’s puts them in an interesting position. The first thing we have to realize is Apple does not sell a cloud computing platform to host and run backend cloud services and apps for businesses. So any approach Apple takes with machine learning, will have to someway play nicely with the backend machine learning systems a business uses to run their cloud computing platform on. Whether I develop apps, enterprise software, banking solutions, etc., I’ll be on either Amazon, Google, or Microsoft’s cloud. Except for China where their are different players.

So Apple is in a unique position that they own a great deal of the customer experience from their customer base, want to provide the value of machine learning and AI to their customers which will show up mostly at an OS layer, and have to work with the machine learning computing platforms where their software ecosystem uses on the cloud backend.

Here is where I think Apple may have to move iOS to a bit more of a hybrid native and cloud OS. They will also have to make sure parts of Siri play nice with whatever backend machine learning tools their software and services ecosystem is using. Let’s use Netflix as an example. Netflix has a machine learning backend. They have chosen a specific machine learning tool set to use but want to make sure that gets implemented into their iOS app. Similalrly Apple will want to make sure Siri can tie into that data in order to let Siri help me find and control Netflix to get what I want to watch. Just saying Netflix can make a Siri app is not enough when they are using something like Amazon Lex on the backend or Google Tensorflow since both of those are specific languages. Apple will have to somehow let those developers who host their solutions on the cloud computing platform of someone else work with their more localized platform. I’m not sure how they navigate this yet but it will be interesting to watch.

My main takeaway from all the discussions I have had with companies leading this space, is we are certainly on the cusp of the cloud truly becoming the centralized computer platform. I expect we will see more of this computing done on the backend, with machine learning the realy value driver, and operating systems continue to become more tightly integrated with these cloud computing backends. How long I’m not sure, but advancements come more than once a year as companies race to own the cloud computing platform with machine learning and eventually AI the central figures.

Published by

Ben Bajarin

Ben Bajarin is a Principal Analyst and the head of primary research at Creative Strategies, Inc - An industry analysis, market intelligence and research firm located in Silicon Valley. His primary focus is consumer technology and market trend research and he is responsible for studying over 30 countries. Full Bio

2 thoughts on “Machine Learning Platforms and the Cloud Computing Era”

Leave a Reply

Your email address will not be published. Required fields are marked *