Machine Learning - From Fundamental Capabilities to Features to Solutions

Machine Learning use cases in Energy

Machine Learning use cases in Industry

A few points to consider:

1. The Big Data analytics space is noisy. Artificial Neural Networks, Machine Learning, Deep Learning, Artificial Intelligence, etc., terms are being used loosely/interchangeably. Research is headed towards building a system with human brain like capabilities but most businesses do not need such capabilities right away to extract value out of their data.
2. You only need 2 data points to describe a line. You can store a million data points about the line and use excessive compute power to find the slope because memory and compute are cheap. But you only need 2 data points. Machine Learning is a means to an end goal of solving business problems. Enterprise customers should spend time on asking the right questions and prioritize those that have business value. Enterprises should avoid wasting time and effort on combining data sets and implementing a big data project merely for the sake of it.
3. There are 3 main stages in any Data Science project. Data Preparation (includes Collection, Integration, & Preparation), Model Building (includes Visualization and other data exploration and analysis tools), and Model Deployment.
4. Data is the most valuable asset and is the only sustainable competitive advantage. Models are not useful if not tested on real enterprise data. The second most valuable asset is the talent to use this data. 

Another question frequently encountered by enterprises thinking about using ML technologies: Should we buy horizontal technology platforms or should we buy a product/application that solves pain points in my industry (and related industries)?

To answer this question we should look at ML/AI market maturity and how sophisticated the buyer is (do they have a solid data science team, have they executed big data projects previously and so on....)

To simplify the challenge, let's take a look at Solar module value chain:

1.An end customer – Residential home or commercial building wants to buy a solar module or solar system

2.What will those customers do with polysilicon or wafer or a single cell?

3.If the buyer is a system integrator and is interested in controlling the value chain, he/she can think about backward integration to compete well. BUT If the market is nascent and there is no customer buying a solar module, then it is useless to think about polysilicon or wafer in the short term.

4.In a nascent industry such as AI, end customers are first looking for new capabilities or solving a business problem i.e. can they do a better job of predicting X - reduce churn, detect fraud, prevent failures etc,. So Applications come first to demonstrate the value proposition of a technology.

5.Therefore it is advisable to first think of Applications in the short term and then underlying technology enablers in the value chain in the long run. This recommendation doesn't apply universally to all types of organizations. Some are more advanced than others.

(image source: The Big Data Market: A Data-Driven Analysis of Companies Using Hadoop, Spark, and Data Science by Aman Naimat)

Based on the buying hierarchy concept first outlined by Windermere Associates, most customers follow a four phase buying pattern: functionality, reliability, convenience, and price (source: Clayton M. Christensen's Innovator's Dilemma)

1. Initially, when no available product satisfies the functionality requirements of the market, the basis of competition, or the criteria by which product choice is made, tends to be product functionality. Once two or more products credibly satisfy the market’s demand for functionality, however, customers can no longer base their choice of products on functionality, but tend to choose a product and vendor based on reliability.
2. As long as market demand for reliability exceeds what vendors are able to provide, customers choose products on this basis—and the most reliable vendors of the most reliable products earn a premium for it.
3. But when two or more vendors improve to the point that they more than satisfy the reliability demanded by the market, the basis of competition shifts to convenience. Customers will prefer those products that are the most convenient to use and those vendors that are most convenient to deal with. Again, as long as the market demand for convenience exceeds what vendors are able to provide, customers choose products on this basis and reward vendors with premium prices for the convenience they offer.
4. Finally, when multiple vendors offer a package of convenient products and services that fully satisfies market demand, the basis of competition shifts to price. The factor driving the transition from one phase of the buying hierarchy to the next is performance oversupply.”

• Translating that to the ML/AI field (just one scenario of how market could evolve), customers are first interested in a machine intelligence capability that will allow them to solve a business problem. So the first phase is Functionality. Functionality could comprise any of these attributes: ability to handle large data sets, time to value/quick prediction, reduction in manual intervention, etc.
• Once the customer is satisfied with Functionality and validated the value proposition in a single department, he/she then wants Enterprise Wide Scalability and Convenience. Customers want the tool to be usable by all of their engineers and not just data scientists. So the second phase is Scalability and Convenience.
• Intermediate Phases……..Customers are interested in advanced features like better data ingestion, faster image recognition etc…….
• In the Final Phase when market is mature, customers are looking for best solutions at lower prices and the basis of competition shifts to price.

Obviously it is not practical to predict market adoption in such generalized fashion. (remember all prediction are made up). However customers should fully understand how their data science project is going to benefit the organization in quantitative terms: total cost of project, time to fully implement, payback period, ROI, prioritizing higher NPV projects............and the criteria goes on.