As I’m trying to figure out how and where we might use Machine Learning (ML) in our software businesses, I thought I’d review all the uses I can find beyond the more general cognitive services (like speech to text, image recognition, keyword extraction, etc) that I’ve already talked about and that – by themselves – are incredibly valuable and offer a near-immediate payoff.
I was a little shocked at the diversity of ways ML is being used. According to TechCrunch there has already been over $10 billion in Venture Capital to 1500 AI/ML startups in 70 countries, which is predicted to rise to more than four times that in 2017!
Since I was compiling this list, I thought I’d share it with you, but it’s just a sampling. Even so there are more than 40 applications described here, in addition to the Cognitive Services as stand alone ML tools.
These services don’t require any training. They are available for integration into apps and workflows now. From IBM Watson alone we have:
- Speech to text
- Text to speech
- Natural Language Processing – sentiment, keywords, entities, high level concepts etc.
- Natural Language Classifier – understands the intent of text and returns a corresponding classification.
- Dialog – script branching conversations between a user and an application
- Language Translation
- Personality Insights based on how people write to match them to other individuals, products, opportunities and tailor their experience.
- Retrieve and Rank the most relevant information from a collection of documents.
- Tone Analyzer uses linguistic analysis to detect emotions, social tendencies and writing style, to understand the emotional context of conversations and communications.
- Visual recognition: understand the content of images to tag the image, find human faces, approximate age and gender and find similar images in a collection. Trainable to teach it custom concepts for specific applications. TheTake has launched a site for consumers to buy that thing they saw in that movie
Predicting Emergency Room Wait Times
Health tech companies and healthcare organizations are using ML to predict wait times for patients in emergency department waiting rooms. The models use factors such as staffing levels, patient data, emergency department charts, and even the layout of the emergency room itself to predict wait times.
The Online Privacy Foundation sponsored a competition to see if it’s possible to predict whether someone is a psychopath based on his twitter usage and apparently, you kind of can.
Identifying Heart Failure
IBM researchers have found a way to extract heart failure diagnosis criteria from free-text physician notes. They developed a ML algorithm that combs through physicians free-form text notes (in the electronic health records)and “reads it.” A computer can now simulate a cardiologist reading through another physician’s notes and figuring out whether a patient has heart failure.
Predicting Strokes and Seizures
Singapore-based startup Healint launched an app called JustShakeIt that enables a user to send an emergency alert to emergency contacts and/or caregivers simply by shaking the phone with one hand. The program uses a machine learning algorithm to distinguish between actual emergency shakes and everyday jostling. In addition to the JustShakeIt app, Healint is working on a model that analyzes patients’ cell phone accelerometer data to help identify warning signs for chronic neurological conditions.
Google’s Deep Learning AI has been applied to cancer diagnosis, and results were better than expected “out of the box” but after “tweaking” it has delivered stunning performance. Clinician’s accuracy is about 48% but by the end Google’s ML was scoring 89% accurate diagnosis.
Predicting Hospital Readmissions
Additive Analytics, is working on a machine learning model that identifies which patients are at high risk of readmission. Hospitals can predict emergency room admissions before they happen.
Identify Skin Lesions
Stanford researchers have trained one of Google’s deep neural networks to recognize skin lesions in photographs. By the end, the neural network was competitive with dermatologists when it came to diagnosing cancers using images. While still not perfect, it’s an impressive result.
Managing Diabetes Patients
hHealth management solution provider Medecision used a machine learning platform to gain a better understanding of diabetic patients who are at risk for avoidable hospitalization or emergency room use. It trained the platform on a database of approximately 8 million patients.
Text and Language
This is a task where given words, phrase or sentence in one language, automatically translate it into another language.
Automatic machine translation has been around for a long time, but deep learning is achieving top results in two specific areas:
- Automatic Translation of Text.
- Automatic Translation of Images.
- How Google Translate squeezes deep learning onto a phone
This is a task where given a large number of handwriting examples, generate new handwriting for a given word or phrase.
This is an interesting task, where a corpus of text is learned and from this model new text is generated, word-by-word or character-by-character. The model is capable of learning how to spell, punctuate, form sentiences and even capture the style of the text in the corpus.
- The Unreasonable Effectiveness of Recurrent Neural Networks
- Auto-Generating Clickbait With Recurrent Neural Networks
ML classification (or topic modelling) algorithms are behind how news articles from thousands of sources gets neatly segregated under topics in Google News or any major news aggregating portals. Equally able to be applied to classifying production keywords into a limited taxonomy.
Business & Legal
Find Tax Deductions
H&R Block have trained an IBM Watson learning machine to seek better tax deductions.
Calculations Insurance Payouts
The technology will be able to read tens of thousands of medical certificates and factor in the length of hospital stays, medical histories and any surgical procedures before calculating payouts.
Predict Successful Product Launches
Dunnhumby and hack/reduce are trying to predict in advance whether a product launch will be successful or not.
Predict Trade Prices
Benchmark Solutions is trying to predict the trade price of U.S. corporate bonds.
Understanding Legalese and Contract Law
Legal Robot is translating legal language into plain language andcan determine what’s missing from a contract and whether there are elements in a contract that shouldn’t be there, such as a royalty fee section in a non-disclosure agreement.
Outsmart The Other Litigator
Attorneys have to comb through large volumes of data to build their cases. The faster and more precisely an attorney can separate signal from noise, the more time she and her team can spend on litigation strategy. According to eDiscovery and information intelligence software provider Recommind, “By combining machine learning and search analytics, you can find patterns in language that indicate peoples’ behavior. Sometimes you’re looking for the best example of a particular kind of behavior, or a particular incident.” Humans tend to look for familiar patterns, based on their experiences or other causes of bias. Machine learning can help yield accurate results, speed the process, and reduce related costs.
Handle a Bankruptcy Legal Practice
Law firm Baker & Hostetler has announced that they are employing IBM’s AI Ross to handle their bankruptcy practice, which at the moment consists of nearly 50 lawyers. Ross, “the world’s first artificially intelligent attorney” was designed to read and understand language, postulate hypotheses when asked questions, research, and then generate responses (along with references and citations) to back up its conclusions. Ross also learns from experience, gaining speed and knowledge the more you interact with it.
Prevent Money Laundering
PayPal is using deep learning to prevent fraud and money laundering at granular levels. By combining deep learning with machine learning and other tools, the company can precisely discern between legitimate and fraudulent buyers and sellers.
ML can flag any malpractice in very high volume high frequency data transactions or communications. ML powered systems can now detect a possible insider trading in a stock market, also ML can flag a rogue customer transaction as a fraudulent transaction in high volume business.
Improve Customer Service
Machine learning can improve the efficiency of customer service by understanding customers and their issues at a granular level. According to predictive analytics platform provider Lumidatum, ML can easily discern between the customers that are beginning to use a product versus those that have more experience with the product, which enables efficient customer support. Alternatively, it can recognize and proactively address customer issues as they occur.
Predict the Auction Price of Heavy Equipment.
Fast Iron wants to predict the auction sale price of a piece of heavy equipment: essentially create a Blue Book for bulldozers.
Determine if a Car is a Lemon
Carvana is building a model to determine if a car bought at auction is a lemon.
Image Recognition and Manipulation
Automatically Adding Sounds To Silent Movies
In this task the ML system must synthesize sounds to match a silent video.
Image Caption Generation
Automatic image captioning is the task where given an image the system must generate a caption that describes the contents of the image. In 2014, there were an explosion of deep learning algorithms achieving very impressive results on this problem, leveraging the work from top models for object classification and object detection in photographs. If you have a slow loading Facebook page, you’ll see the automatically generated description of the photo load before the photo.
- A picture is worth a thousand (coherent) words: building a natural description of images
- Rapid Progress in Automatic Image Captioning
YouTube says it has now captioned over one billion videos, in 10 languages.
Colorizing Black and White Images
Deep learning can be used to use the objects and their context within the photograph to color the image, much like a human operator might approach the problem.
Focus attention on objects in images.
- Recurrent Models of Visual Attention [pdf], 2014
Answer questions about objects in a photograph.
Turn sketches into photos.
- Convolutional Sketch Inversion [pdf], 2016
Create stylized images from rough sketches.
Search Images by Content
Facebook’s Lumos computer vision platform is now powering image content search for all users. This means you can now search for images on Facebook with key words that describe the contents of a photo, rather than being limited by tags and captions.
Other Applications of Machine Learning
Jukedeck is one of a growing number of companies using artificial intelligence to compose music. Their computers tap tools like artificial neural networks, modeled on the brain, that allow the machines to learn by doing, rather as a child does.
Employee Access Control
Amazon sponsored a ML contest to determine whether it was possible to automate employee access granting and revocation. According to Amazon, “These auto-access models seek to minimize the human involvement required to grant or revoke employee access.”
Fine-Tune Security Screening
Human screeners often overlook items that machine learning can identify. And, machine learning can easily adapt to seasonal changes affecting bag types and bag contents, or the specific requirements of a particular venue. Operational intelligence solutions provider Qylur are trying to reduce the number of false alarms.
Cornell University is working on an algorithm to identify whales in the ocean based on audio recordings so that ships can avoid hitting them.
Determining Bird Species by Audio Recording
Oregon State University is working on software that will determine which bird species is/are on a given audio recording collected in field conditions.
Stopping Spam and Malware
In 2014, Kaspersky Lab reported it was detecting 325,000 new malicious files every day. Only machine learning and deep learning can deal with that volume, particularly since each new malware only differs about 2%!
Israeli communication services provider Orange (aka Partner) has been using machine learning for the past two years to help protect its business and customer data.
Compete Intelligently (Tour de France)
Software and device maker winningAlgorithms use what people are saying on social media to better understand what’s happening in an event. The algorithm is able to determine the credibility of individuals reporting race details over social media.
Mapillary uses machine learning to stitch together 3D visualizations of photos contributed by its more than 12,000 users. The images are available via an API.
Researchers at Microsoft and the University of Cambridge, have created a system called DeepCoder that solved basic challenges of the kind set by programming competitions.DeepCoder looks across the Internet for snippets of code that it then builds into new applications.
Win at Texas Hold ‘Em
A lengthy tournament of Hold ‘Emended with victory for Libratus, an AI program developed by a professor and PhD candidate at Carnegie Mellon University. Libratus emerged victorious after 120,000 combined hands of poker played against four human online-poker pros. Libratus’ $1.7 million margin of victory, combined with so many hands, clears the “Brains Vs. Artificial Intelligence” tournament’s primary bar: victory with statistical significance.
Jam with the Machine
Google has launched a fun new machine learning experiment: A.I. Duet. This new web-based experiment lets you play melodies on your computer’s keyboard (or a supported MIDI keyboard) and the computer will then try to play a duet with you.
Inspect Cell Phone Towers and Identify Problems
Not linkable, but at the HPA Retreat this last week, Jason Brahms of VideoGorillas – an action sports remote streaming company originally – revealed how they are using ML to analyze incoming video to derive metadata, and he talked about a future project to teach ML how to recognize faults on cell phone towers when inspected by autonomous drones.