As I previously outlined, an important theme of Google's recent Cloud Next event, was the democratization of AI through the use of cloud services that encapsulate complex algorithms, AI expertise and scalable infrastructure to allow "non-specialists in business to tap into sophisticated models and techniques that were formerly limited to a high-priesthood of ML (machine learning) researchers."
AI is central to Google's strategy due to the company's deep expertise in the field and because of the technology's potential to reshape business processes and decision making. Indeed, at last year's Google Cloud event, chairman Eric Schmidt said he's convinced that the convergence of cloud computing and ML
...will be the basis and fundamentals of every successful huge IPO win in 5 years.
While Google spins a compelling narrative, actions speak louder than words, so I was interested to see how organizations are already putting AI and ML to work. A session on Customer Successes with Machine Learning proved particularly illuminating. Google Product Manager Riku Inoue reviewed seven case studies spanning four business segments from both large multinationals like Airbus and non-profits like Global Fishing Watch.
In manufacturing, Google sees ML used for quality assurance, predictive maintenance, processes optimization and demand forecasting, while in the energy and environmental realm, applications include large-scale monitoring, seismic and energy usage data analysis and carbon emissions monitoring and credit trading. Financial services and commerce present especially lucrative opportunities by using ML and AI for fraud detection, risk analysis, insurance and financial contract pricing, credit scoring, cross- and upselling, product recommendations and inventory planning.
Some organizations already use self-learning security systems to spot unusual file activity that's typical of ransomware. Much of this is not ew per se but it is the rapid expansion of the possible that intrigues analysts like myself and buyers alike.
Turning data into insights
Insurance is an area ripe for algorithmic improvements since companies are gaining access to an increasing volume of data, particularly auto and casualty underwriters, that garner data from cars which are little more than roving, connected computers.
One user of Google ML services, AXA, says only about 7-10% of its insured drivers cause an accident in any given year and just 1% of accidents require payments of $10,000 or more. The trick is predicting which drivers are most likely to be in that risky group.
Even without using automotive IoT data, AXA has identified 70 factors, such as a driver's age, address, other demographics, policy details and vehicle type that are predictive of accident risk. It used these as inputs to a neural network, which AXA trained on historical data to create a risk model.
When used with new data, the results accurately predicted accident incidents and severity 78% of the time. This is significantly better than the company's previous method. AXA uses the new model to improve pricing and investigate new services such as real-time pricing.
Two companies using ML for cyber security are Sparkcognition and SMFG. The former has developed models to do malware detection that it claims is many times more effective at identifying zero-day exploits in real time by using Google Cloud's scale and ML engine to tag and stop new threats. As Sparkcognition describes the technique,
Instead of looking for a hash or a heuristic match, pattern detection and correlation are employed to identify similarities to other known malicious files. Even in the case of handling packed and obfuscated files, machine learning techniques can quickly identify and block any malicious intent present on a client endpoint. Best of all, these techniques learn malicious patterns over time as new file types and threats are discovered.
SMFG, a Japanese financial services company, uses a neural network to analyze credit card transactions and achieve 80-90% accuracy at fraud detection.
Image recognition isn't just for social networks
Image recognition is the quintessential application for neural networks where companies like Facebook and Google use them to categorize photos and automatically tag faces.
Airbus, which has a division providing geographic information and intelligence, does something similar to discriminate between clouds and land features like snow cover, in satellite imagery. Such image filtering has traditionally been a manual process that can't handle more than about ten thousand images per day. Even then, the detection isn't perfect.
Using Google Cloud's ML Engine, Airbus reduced the error rate from 11% to 3%, a 72% improvement. Furthermore, by using recently-released GPU instances, Airbus reduced the training time for its very complex model by two orders of magnitude.
Product inspections as part of the production and QA process is another area ripe for ML-based image recognition. For example, Kewpie, a Japanese food processor and distributor, had a manual process to identify defective ingredients on an assembly line, such as vegetables that don't meet quality, color, size or shape standards.
It's a challenging problem as material rolls down the line that requires intense focus by food inspectors. Kewpie's previous attempts at automation using purpose-built machines were never as accurate as humans; however an ML image recognition system can match the accuracy of trained inspectors and automatically stop the line when defects are found. While the system is much less stressful on workers, it also obviates the need for $100,000 in equipment per production line.
Aucnet, a Japanese used car auction site, takes a page from social networks by using ML to automatically identify and classify car models and features from dealer-submitted photos. The site likes to have 18 exterior and 12 interior images for each entry, however sorting and tagging these photos using its manual process takes about 5 minutes per car. It adds up to a lot of time when you're handling 5 million cars per year. Aucnet turned to technology, building an ML image classifier with 95% accuracy on make and model.
Since image classification is such a common problem, Aucnet bootstrapped the development of its neural network model by using a technique called transfer learning This technique started with a standard image model that was pre-trained on millions of generic images and added neural net layers specific to car makes and components.
Transfer learning significantly reduced the time and computational resources needed for model training. In a demo, the system was able to accurately identify a vehicle and various parts within seconds. The slowest part of the operation was uploading the images from the presenter's laptop.
I agree with Google's head of AI, Fei-Fei Li, that cloud services could someday democratize AI, however we're still nowhere near AI services being usable by the masses.
While products like Google ML Engine or AWS Machine Learning make sophisticated algorithms and infrastructure available to a much broader audience, these remain complicated tools with a steep learning curve: more like Photoshop than Snapseed.
Although AWS, Azure and Google have each introduced pre-built ML services for common uses like image recognition, dictation, translation and natural language processing, building models tailored to specific business problems still requires understanding the features and benefits of various models and programming in an ML framework like TensorFlow.
Unfortunately, the number of business analysts who can tell the difference between a rectifier neural network with activation and a fully connected feedforward network is minuscule. For now, AI experts in business, like their peers in data science toiling with predictive statistics in R, will be unicorns.
Over time, AI services will become further packaged and abstracted from the underlying mathematics, further expanding their usability. For now, organizations without the requisite intellectual horsepower should focus on characterizing the problems they hope to solve with AI and identifying the data sources that can eventually train machines that learn.