Machine Learning (ML) is bringing an irresistible benefit to enterprises- turning inexpensive computing power into skilled (virtual) employees. ML has the massive potential to kickstart an era of self-learning systems that can replace hundreds of manual operations overnight. From predictive analysis and flagging manual errors, the benefits are literally infinite.
Inspired by these possible benefits, IT leaders across several disciplines are betting high on Artificial Intelligence and Machine Learning. This has catapulted Machine Learning as a hotbed of activity across the world. Contrary to common belief, the reach of ML extends beyond tech corporations. Even entertainment giants like Netflix are deploying AI and ML systems to optimize their operations.
According to McKinsey Global Institute’s ‘Artificial intelligence the next digital frontier’ study, AI helped Netflix save $1Billion in annual revenues by providing better results that reduced canceled subscriptions. But, all this did not happen overnight. A good number of trial and errors have been undergone before enterprises were able to taste success with ML. Their trials are all excellent learning materials that would help early adopters join the league of ML systems easily.
This post is a roundup of such learning from early adopters.
Identify Definite Use Cases
Identifying where to deploy ML and for what purpose to deploy is the first thing that IT leaders of successful business leaders got right. Identifying definite use cases for ML adoption within the business operations gives a sense of purpose and unity of direction.
Some possible use cases that are universal to any business model include:
1. Strengthening Security: Using AI/ML systems to raise alerts when security threat patterns are observed. Eg: DDoS attacks
2. Automating Transactions: Automating repetitive and procedure-based transactions for quick results. Eg: Credit appraisal for loan processing
3. Predictive Analysis: Analyzing historical and real-time data for predicting future instances. Eg: Inventory planning in manufacturing.
4. Chat-bot communication: Virtual assistants that provide dynamic feedback to human users based on past preferences. Eg: Chatbots customer support for an online store
Another point worth noting is that all these use cases and their benefits should be measurable. For instance, how much cost-savings can ML bring in by replacing manual resources?
Or, how many man hours (or other metrics) can be improved by an ML system compared to a manual resource. These results should be clearly defined to measure the benefit that ML adoption is bringing to the business.
You Need Clean + Diverse Data
ML systems thrive on data. They feed on data to grow their neural brains that respond to human queries. Such data should be clean; meaning, they should be free of duplicates, errors, empty fields and so on. Clean data helps ML systems to learn and process information smoothly. Also, such data has to be put in a proper format that the ML system can easily grasp.
The data range should also be broad and diverse. A diverse range of data exposes the ML system to a 360-degree view of the process helping it draw patterns and responses accordingly. For example, frequently asked questions by a first-time online healthcare website visitor. Another factor that adopters need to consider is the inclusion of unstructured data. ML systems are well-compatible with structured data. But, the real bang for their buck is in including unstructured data like images, video, audio and other multimedia formats.
Netflix’s recommendation system was able to split its users into 2,000+ taste groups based on machine learning analysis. Deciphering unstructured data played the key to achieving this feat.
The takeaway: Maximum RoI from an ML system can be obtained only by converting unstructured data into structured data models.
The System Needs To Be ‘taught’
The primary benefit of ML is that systems learn on their own. But, even in their self-learning mode, they need to be instructed what kind of data they should learn and what kind of expertise needs to be built. To cite a real-life example, in the healthcare industry, doctors need to weave through several patient records before arriving at a disease symptom. Being a heart specialist also requires the doctor to be good with diabetic and other allied branches. ML systems should be taught to assist the doctor across several disciplines so that quicker diagnosis can be made.
Lynda Chin, director of the Institute for Health Transformation at the University of Texas System, says, “What we need to do to get there is to have a common infrastructure that allows us to break down the silos of different specialties.”
To make ML systems experts at different specialties, it takes time. Depending on the volume and complexity of the data, it could take anywhere from few days or even months. There is no absolute measure of the time required for the data volume and complexity changes according to industry and its subdomains.
Resource Demands Are High
As Theodore Roosevelt once said, “Nothing ever worth having comes easy.” Implementing ML systems demands several resources. You will need to set up an infrastructure that is quickly scalable, flexible and also reliable. The infrastructure should be adept at collecting, hosting and analyzing data to churn out dynamic output.
Having at least one in-house data scientist who can strategize and lead the ML implementation forward is essential. Or still better if you can partner with experts and IT vendors who can provide the resources and technical know-how to implement an ML strategy and road map. Some industrial giants are taking a step further by acquiring AI and ML startups to add more power to their digital operations strategy. IBM has earmarked $3 Billion to make Watson a leader in IoT market. Baidu, Google, Amazon, Tesla, BMW, Facebook, Alibaba are other names that are pouring investments into AI and ML computing services.
Crafting an ML strategy and implementing is a business transformational journey. While the driver seat is taken by IT leaders, other functional departments and their leaders must also participate in the exercise to make data available. Large volumes of data must be available to enable the ML system to draw patterns and synthesize dynamic responses. Such data must also be cleansed and made consistent to avoid discrepancies in responses. Since the task involved is highly technical in nature, the presence of an in-house data scientist or a separate AI-ML arm is required.
Image Source: Forbes.com