What is AI?
A system or service which can perform tasks that usually require human intelligence.
AI is not something new. The rise of AI started 60 years ago. In 1957 Frank Rosenblatt invented The Perceptron and in 1975 The Backpropagation was published. People talked about AI but it never became mainstream. That is until 2011 when AI started booming.
How does an artificial neural network work?
There is a hidden layer of neurons between input and output. A neuron uses multiple inputs to produce an output. The neurons are organized in layers and the networks are self-organizing. Giving a large set of data as input, the network will be able to learn automatically how to produce the correct answer. There is a catch, it involves a lot of math operations that grow exponentially with the data and also as the layers increase. The curse of dimensionality.
Until recently there was not enough data and calculation power to train the neural networks. By 2010 large neural networks were made possible with algorithms, data, GPUs & Acceleration and Cloud Computing. AWS sits in the middle of them since it allows people to use large computing power, algorithms to train data and GPUs and acceleration.
Innovation is the promise of AI and deep learning by making it possible to provide new features and experiences for products.
AI today at Amazon
- Thousands of robots with AI in warehouses to optimize delivery of packages to customers
- Alexa fueled by natural language and automated speech recognition
- Drones for Prime Air packed with ML and DL, computer vision technology to deliver packages
- Amazon GO, anyone can enter a shop, take what she needs and leave the shop, the experience is monitored by computer vision, AI improving the retail experience of the customers
- Thousands of engineering focused on ML: logistics, search, AWS,…
AI & Deep Learning in the hands of every developer
AWS provides services, framework and platforms for developers. It is a rich ecosystem of services to use to fulfill the requirements. AWS is very popular with a lot of customers interested in using AI or that use AI.
Amazon AI Ecosystem
Text to speech engine. You give it a string of text and Polly returns an audio stream with life-like speech of the input. It uses DL under the hood to transform a text into human language. It includes focus on voice quality and pronunciation. 47 voices in 24 languages.
Use cases: Washington Post uses Polly to turn their webpage into podcast, Y-cam case study uses Polly to provide human sounding voice.
The API is very simple. Go to the AWS console => Amazon-Polly => Text-to-speech => SSML.
Image Analysis. You give it an image or reference stored in S3, it returns information using object and scene detection, facial analysis, face comparison and so on. You can use it for image moderation which social networks use to moderate image for nudity. It will return you a JSON file with the probability and confidence ratio of the label. You can do facial search, compare different faces with confidence index. You can do celebrity recognition which is quite new by sending an image and Rekognition will return the name of the celebrity.
Use cases: Artfinder Case Study to add meta data on images for customers to search for.
Conversational engine. It does speech recognition. It powers Alexa and allows you to create conversational bots. Advanced conversational applications.
Use cases: book a room at a hotel.
How Lex works:
- Intent performs an action in response to NL user input
- Utterance spoken or typed phrases that invoke your intent
- Slots are input data required to fulfill the intent
- Fulfillment mechanism for your intent
Lex extracts the words using automatic speech recognition, it detects words and meaning with natural language understanding, invoke Lambda with that information, respond to the customer using Polly.
- Amazon Machine Learning – Supervised learning algorithms. Fraud detection.
- Amazon EML- Elastic MapReduce. ML Apps on EMR like Spark.
- Spark and SparkXML.
For do-it-yourself use cases. AWS comes with many frameworks like Apache MXNet that are ready to use and experiment with.
Use cases: early detection of diabetic complications at Stanford, autonomous driving systems at TU Simple.
Once you have done training with MXNet, you have your models. You can store your models in S3, then use a Lambda Function with API Gateway to load and query the model in Lambda and do predictions. Alternatively you can run predictions on ECS which is a lot more flexible because you will not run into a timeout of 5 minutes like with Lambda so you have a lot more options.
On the low level of the ecosystem you have Amazon EC2 P2 and G2 GPLUS, Amazon EC2 CPUS, AWS Lambda, enhanced networking and AWS IOT and AWS Greengrass. Leverage the full power.
Use cases: recommendation engines on Netflix.
AI for everyone!
What we see is the democratization of AI. Look at the entire set of AI offering to see what fits best for your use case.
Resource for this article:
- AWS Webinar by Adrian Hornsby, Tech Evangelist