Why am I writing about Cloud AI?
With hardware and architecture of cloud instances within their control, cloud companies have got complete freedom to build cutting edge platforms.
Grammarly, Google Assistant, DeepRacer, and DeepComposer are all AI models built on cloud. (Do read the application section at the bottom)
What is AI?
We will talk about AI, ML, and Deep learning in this article. So let’s get the basics right! AI means getting a computer to mimic human behavior.
Is there AI without ML?
Yes! A calculator is a basic example of it, there is no training data in it while the machine does mimic the human behavior of calculating.
AI without ML is called good old-fashioned AI (GOFAI). It was extensively used before ML was possible. In GOFAI, the machine is given basic rules and logic instead of training data (which happens in ML). 2+4=6.
Whenever we give instructions to the machine on how to make decisions or how to respond to a sequence, we are using GOFAI. In nutshell, GOFAI is used when ML is not required, or when we need to give a human/business logic layer to ML-based systems to make it a useful model.
Back to cloud AI,
Considering there are millions of AI platforms on the Cloud, I am following Gartner’s approach- mentioning only the Cloud AI platforms that serve more than one purpose.
Cloud AI Platforms
These AI platforms allow you to prepare data, build ML models, and deploy them. The models can be used for comprehending text or images or responding to a sequence of events/logic.
Let’s dive into an AI platform pipeline
While all the other stages are self-explanatory, the third step is called feature engineering. Where you add/segment/modify data or data extraction to align it with business context.
Companies have built separate sub-products for each stage. And most of these stages are inter-operable with other companies or open-source platforms. For example, you can use Google’s Tensorflow to build models in Amazon’s cloud or H2O.AI models on Google Cloud.
Let’s look at Google’s and Amazon’s Pipelines,
The first step is collating training data. These platforms provide a library of labeled data and ML models for starters.
Additionally, both the companies give manual data labeling as a service, using which you can outsource your data labeling to them. The more varied the data trained by them, the larger their library, hence they provide the service.
Once you have the data models ready, the platform allows you to build models (within the platform or export from other platforms — TensorFlow, PyTorch, and Apache MXNet) and deploy the solution wherever you want using Jupyter Notebook.
AI Platforms give you tools to make the implementation of AI, ML, Deep learning easier, faster and less skill dependent (less code). You can do everything that AI platform offers by coding it yourself or by using libraries & other tools, but that’s too much effort, rather these platforms are the one stop shop.
Use and components of AI Platforms
IMP: Using the AI platforms you can deliver ML Models for specific business/human tasks (including image & language related tasks). However, over the years companies have built strong ready to use language and image (vision) specific models and hence they are sold as separate products.
That makes it three products on the platform — 1. Custom ML Models, 2. Vision Products and 3. Language Products.
Custom ML model building feature on Cloud AI platforms is sold under AutoML category because of the added features making ML Modeling easier.
Charles Babbage, “the father of the computer”, conceived a general all-purpose device using punch cards in 1843 and a century later Alan Turing wrote his paper the “Universal Machine”.
In the 20th Century, inspiration was taken from the human neural system to build logical systems. IBM made the first self-learning checkers game in 1952, NETTalk self-taught pronunciation in 1985, and in 1997 IBM’s Deep Blue beat chess grandmaster Garry Kasparov.
The early 2000s saw a reboot of neural network as deep learning and 2014 saw a Eugene pass the Turing test.
While researchers keep pushing the envelope to make machine human like, ML has become mainstream with free courses from Andrew Ng and platforms like Tensorflow.
AI platforms have further simplified end to end deployment of ML models.
2. Language Services
The above mentioned evolution of ML and AI allowed researchers to use the models for understanding language.
The first neural language model, a feed-forward neural network was proposed in 2001 by Bengio et al. It was tasked to predict the next word/phrase in a sentence after going through thousands of articles.
Then came “Word Embeddings” in this case the model was doing more than a lookup to predict the next word, it was mathematically building relations between words. Good can be positive, while bad can be negative.
Over the years many neural networks were made which led to various models to predict or understand words.
- LSTM(Long Short Term Memory)
- GRU(Gated Recurrent Unit)
- BILSTM (Bidirectional Long Short Term Memory)
Many of these pre-trained language models are available in most AI platforms as drag and drop features. The application of Language Models is found in,
Natural Language Understanding
Speech to Text and Text to Speech
3. Vision Services
In 1959, Russell Kirsch and team figured a way to transform images into grids of numbers — the binary language machines could understand. In 1963, 2D and 3D images could be created on paper.
In 2001 the first face recognition model was developed and in 2012 the biggest breakthrough came when Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton designed a new object recognition algorithm, ensuring an 85% level of accuracy. In 2015, the accuracy of facial recognition surpassed 95%.
Now there are different ways to recognize objects and most of these are available in libraries which can be pulled into the Cloud AI platforms. These models are used for
Image Analysis and Services
Market & Companies
The Cloud AI Market was valued at USD 5.2 billion in 2019 and is expected to reach USD 13.1 billion by 2025, registering a CAGR of 20.3%, during the period of 2020–2025- MarketIntelligenceData
By 2025, 50% of data scientist activities will be automated by AI, easing the acute talent shortage- Gartner.
The increase in cloud adoption, skilled labor, trained models, and use-cases of AI will augment this growth.
The market consists of companies that offer specialized AI Platforms for AutoML, Language, and Vision services in combinations or stand alone. Companies such as Google, AWS, Microsoft provide all these services together and have managed to build an ecosystem around it.
Top Cloud AI Platforms: Google, Amazon, Microsoft, IBM, Intel, H2O.ai, Peltarion, Aible, and Prevision.
While Google, Microsoft, IBM, and Amazon market their product/capabilities, the smaller companies market use-cases.
Also, the big four have channel partners across the globe to sell and implement their products, while smaller ones have only a few partners or sell directly to customers, holding back their growth.
TOP AMAZING APPLICATIONS
DeepRacer — Amazon allows you to code model for driverless cars. You can deploy the model on AWS to participate in virtual races or in the Physical DeepRacer Evo Car. If you win online often you will be called to their annual race event. Amazon has managed to build an online-offline community around a software run race!
NFL NextGenStats- NFL takes RFID of each player and gives real-time insights on what's expected out of a run of a player. They take in a runner’s speed, distance from others, past record to give real-time predictions.
DeepComposer- A keyboard that completes your melody. You can make a tune with a few presses, and after that model will take over. It not only completes the tune but also adds drum, guitar, and other instruments.
Favorite Business Applications
- Google E-Mail text prediction/sentence completion
- Article summarization
- Detection of issues in rails, windmill blades, and CT/MRI/X-ray scans of humans.
- Facebook / Google Photo element recognition
- Fraud detection
- Realtime competitive pricing models
Most companies have usage-based pricing and they provide calculators for quick estimations. Following are the links to a few of the pricing calculators,
Amazon: They haven’t made or I wasn’t able to find a calculation page. But they do mention pricing for each service separately like done here for vision.
While smaller companies charge per user per month as seen here.
Img & Content Sources:
R. Olson et. al. (2016) “Evaluation of a Tree-based Pipeline Optimization Tool for Automating Data Science.”