AI is pervasive in modern society, whether you see it or not.
Most often, AI works in the background, making complicated tasks trivial through automation. Much of the technology we take for granted in our daily lives requires multiple layers of artificial intelligence that work together to deliver informed decisions and outcomes.
Here are a few of the most exciting AI-based deployments and products I’ve come across recently, including one product that makes artificial intelligence and machine learning possible for just about any business or project.
AI in Smartphone Photography
Consider the digital photos we take with our iPhones and Samsung Galaxy smartphones.
These images require processing billions of pixels for color, contrast, saturation and now database management. Little did we know that phones would begin to leverage AI to improve consumer photography — to see, process, and then organize photos by location, recognized faces, subject matter, and even noteworthy events.
AI innovation has enabled exponential leaps in smartphone photography, to the point where AI-assisted photos are often at par with those shot by advanced (and far more expensive) SLR cameras and lenses. An example of this is how some Android phones can now turn dark nighttime shots into what almost appear to be daytime shots, without losing important visual information. Starting with the Galaxy S10 and S10+, AI powers the digital images pulled from three rear camera sensors to extract the maximum lighting information possible. With the built-in Scene Optimiser technology, users will not be able to discern that the phone is using AI to combine up to 17 photos into one composite masterpiece.
When the Galaxy phones are in night mode, they can detect light and objects that should be more visible, and then combine 7 shots over 2-3 seconds to create a final photo. If the camera detects that it’s on a tripod, it will take 17 shots in about 35 seconds to create a “Super Night Shot”. This digital combining effort is orchestrated uniquely to each environment, light exposure and subject(s) in the shot.
As impressive as this system might seem, smartphone photography is but an appetizer when it comes to modern AI.
Imagine what a similar approach can yield for complex use cases that require Trillions of Operations Per Second (TOPS), such as autonomous driving.
AI in Self-Driving Cars
Autonomous or self-driving cars are an emerging technology that is heavily dependent on AI capabilities. This is true, by definition, as the technology aims to allow cars to drive themselves without direct human control.
Elon Musk, CEO of Tesla, just announced that by 2020 all Teslas will not only be fully level 4 self-driving, but will also be ready to compete with Uber by providing robotaxis that will not need a human driver to be present (subject to state laws). Owners of Teslas on the ride-share network will have the opportunity to earn up to $30,000 a year in passive income, simply by allowing their self-driven vehicles to participate in this sharing economy.
Google’s Waymo is also advancing towards preparation for public use, taking the driver out of the equation. However, questions about limitations with the deep learning capacity of its AI currently plague the project.
One specific technical challenge that exists across all projects in this space has to do with computer vision (CV), or the ability to see and understand the environment in real time. As an example, being able to understand that a pedestrian on a sidewalk is about to step into the street (rather than just standing by the curb) is something humans can do intuitively, however this is a significant challenge for computers to do. As another example, being able to discern a large, white semi truck trailer blocking the road seems to be a challenge for current vehicles, as we’ve seen in a few similar accidents with self-driving cars, including this recent example with a Tesla Model 3.
However, thanks to advancements in computer vision and the continual improvements we are seeing in both sensor and computing hardware, we can expect for these sorts of challenges to be overcome in short order. On the whole, self-driving cars are already much safer than the typical human driver (and have been for some time).
AI is now Accessible to Just About any Company
Even if your business isn’t a Tesla or Samsung-like tech giant (and developing your own R&D lab for AI isn’t feasible), there are still sensible ways to implement cutting-edge AI into your product.
Amazon Pre-trained Models
Amazon Web Services (AWS) helps organizations move faster and lower IT costs by offering a broad set of global compute, storage, database, analytics, application, and deployment services. Amazon has created AWS services that support machine learning. This includes plug-and-play API services like:
- Recognition for image and video analysis. Form Capture. You no longer need to manually type out forms and pull data to create new databases. Text in Image Capture. SageMaker allows you to capture text within images.
- Lex for conversational interfaces (chat bots). Another service worth exploring is Amazon Lex. Amazon Lex is capable of breaking down language communication into: intents, utterances, slots and ultimately fulfillment. By breaking down usual activities into intents, Amazon Lex’s machine learning follows the pre-programmed path towards fulfillment. Say for example booking a hotel. Choose the intent under BookHotel and what you want to do (utterances), the AI will ask you where (slots) and you will have to type in your destination. The AI then guides you to fulfillment of this intent– a hotel booking.
- Comprehend for text analysis. Imagine detecting Customer Satisfaction without surveys! No need to troll social media to detect how customers feel about your product. Let SageMaker do that for you. SageMaker streams your social media feed into your machine learning model, collect and make sense of your customer’s sentiments.
- Transcribe for speech-to-text, making transcription much easier.
- Polly for text-to-speech conversions, etc. allows you to use multi-lingual text to speech.
These specialized services provide pre-trained models which can be easily integrated in your applications via API.
A Look at Amazon SageMaker
Amazon’s SageMaker offers the capability to add intelligence to machines that originally are without learning skills, as well as create, deploy and refine original machine learning models. SageMaker models collect data upon deployment and return collected data called “ground truth values” to further refine each subsequent deployment.
This increases the scalability of Amazon SageMaker, as the models improve as the data sets grow bigger and bigger. Amazon makes it easier for users by having prepacked algorithms available to start with, so that models don’t have to be created from scratch. SageMaker offers speech recognition, natural language processing (NLP), text generation, regression and computer vision to name a few capabilities. All of these capabilities can be intregrated into existing apps to provide powerful machine learning capabilities.
- Build highly accurate training datasets
- Fully managed instances running Jupyter notebooks for training data exploration and preprocessing
- High-performance, scalable machine learning algorithms optimized for speed, scale, and accuracy
- Broad framework support
- Amazon SageMaker supports reinforcement learning in addition to traditional supervised and unsupervised learning
- The open source Apache MXNet and Tensorflow Docker containers used in Amazon SageMaker are available on Github
- One-click training
- Automatic model tuning
- Train once, run anywhere
- Easy deployment, one click away
- Automatic A/B testing
- Fully-managed Hosting with Auto Scaling
- Inference pipelines
- Free automatic model tuning
SageMaker simply allows for very easy machine learning creation. The first step is to spin up a “notebook instance” which will host the Jupyter Notebook app. It does not require a connection to that notebook instance (everything is already pre-prepared). Create a new notebook and use it to start the learning process. Jupyter notebook allows freedom for experienced data scientists who are already accustomed to the tool. It is also new user-friendly, with easy procedures on how to use the instances so you don’t have to undergo extensive and time-consuming training to become productive.
SageMaker supports Apache MXNet, TensorFlow and comes with a built-in algorithm which allows you to choose other libraries and frameworks. SageMaker offers many of the most popular built-in ML algorithms. Some of them include K-Means, PCA, Sequence models, Linear Learners and XGBoost. The SageMaker workflow as illustrated below, allows you to prepare data and model, then configure and launch training using SageMaker SDK.
Some instances may cost more, and true to the AWS (pay for what you eat) model it makes sure that clients do not accidentally use more expensive instances by providing limitations.
Check out SageMaker pricing per region HERE.
Companies currently using SageMaker include Haliburton, Plume, Vivid Seats, Buxton, Delta Air Lines, and Huawei Technologies.
AI’s Evolution from Theoretical Science to Thriving Commercial Reality
As the examples I’ve highlighted demonstrate, AI has come a long, long way in recent years. What was once a field explored by a handful of visionary data scientists is now a commercially-viable way for any business to improve their software products and provide game-changing value to users.
If your business might benefit from adding artificial intelligence capabilities to your product, I invite you to contact our team at CitrusBits to explore this opportunity further.