In the form of machine learning (ML), artificial intelligence (AI) is making its way into the enterprise, promising to transform many areas of the business including customer experience. Just as humans can learn and act from experience, machines can learn from data, identify patterns and make decisions. Furthermore, this cognitive ability of machines can improve dramatically over time. The higher intelligence demonstrated by machines has led to wide application across industries; virtually every industry, including banking and financial services, and healthcare and life sciences, is embracing it to revitalize core business processes and models.
From our perch, we see ML accelerating across enterprises, impacting fields as diverse as human communications, autonomous cars, fraud detection and disease diagnosis. With traditional computational approaches, algorithms are explicitly programmed to solve particular problems. In ML, however, the system can identify data patterns automatically and improve on experience without being explicitly programmed. Within a few years, ML could potentially become integral to far more effective and widely available applications that perform tasks on their own and create fundamentally new, profitable business models.
ML’s increasing popularity is based on a variety of factors (see Figure 1):
Data availability. The vast and growing amount of structured and unstructured data in various formats — such as plain text, raw images, video files and audio files — is fueling interest in ML as it can learn from all these data sources and autonomously build applications.
Computing power. Processing large volumes of data has become more affordable and cost-effective through cloud services. Efficient parallel computing is enabled by graphics processing unit (GPU) computing, one of the most pervasive, accessible and energy-efficient (i.e., faster and with less infrastructure) platforms for training ML models.
Tools and frameworks. The availability of various open source frameworks, toolkits and libraries makes it easier to build, evolve, implement and scale the ML models using popular programming languages.
Relevant ML use cases at work include:
Banks could apply biometrics for facial recognition and voice authentication to improve customer experience and security.
Insurance companies could automatically recognize and assess vehicle damages, thereby lowering risk.
Hospitals could create a library of fundus images to screen and detect diabetic retinopathy.
ML Process/Development Lifecycle
ML is often an iterative process, and collecting, training and applying large volumes of data to develop suitable ML models is very challenging. The data used by an ML algorithm must contain the information about prior occurrences of a situation or condition, suitable to make predictions. When business conditions change, ML models can simply be retrained on new data, without the need to rewrite the instructions as you would for explicitly programmed systems. The essentials include:
The right questions.
In machine learning, asking the right questions pertaining to the identified business problem will help understand the value to be generated by an ML solution.
The right data.
Having accurate and complete data sets will establish reliable inputs and outputs to train the model for an ML solution.
The ML algorithms define the success criteria, which should be measurable on an ongoing basis. This enables the system to continually learn and adjust its algorithms to help the business meet its objectives.
An efficient ML framework reduces the complexity of machine learning, making it accessible to more developers. An effective ML framework is:
Simple to understand and easy to code.
Developer-friendly for building models.
Easily optimized for high-performance computing.
Capable of automating the computation process.
Figure 2 offers a comparative view of the various ML frameworks.
Emerging ML Prototypes
ML has the potential to redefine how business is conducted across industries. The following use cases are examples of how ML will drive business value to enterprises.
Service Desk Optimization
Service desk or L1 technical support jobs (such as support ticket creation) contain rote and routine tasks that are increasing exponentially across industries. By 2021, more than 50% of enterprise spending will be on smart bots for optimization.
Solution. A user speaks to a smart AI bot through their choice of channel such as voice over Internet protocol (VoIP), a Chrome browser, Alexa or Google Home. The system converts the speech to text using Google ASR, understands the context and manages conversations using Dialogflow and NLP Engine and fetches/validates the information requested by the user from external resources such as knowledge repositories and databases. Once the system receives a response, it rephrases and converts the text into speech using Amazon Polly, a text-to-speech service that uses advanced deep learning technologies to synthesize human-sounding speech, and sends it back to the user. If the smart AI bot is unable to resolve the query, it automatically transfers the request to a human agent for resolution.
Benefit. Cutting down on resolution time produces operational cost savings. AI bots reduce the call length from an average talk time (ATT) of three minutes to two minutes, based on our prototype.
Image/video duplication is a rapidly increasing challenge, as media content volume is already huge and growing rapidly across most organizations. Multiple iterations of images are often used, varying only by format, size, quality, compression and image transformation.
Solution. Key features are extracted from media files, and are indexed. A lightweight scalable search service is developed that identifies duplicates through a similarity score. An AI tool is integrated into the de-duplication service, which is responsible for preventing duplicate media files from being uploaded to cloud storage. The solution comprises a two-engine model — one for querying and the other for indexing. The querying engine is used to check a file for duplication in the existing media corpus and the indexing engine is used to transform the media corpus into an N-dimensional feature space, and index the files for efficient querying.
Benefit. Automating the de-duplication of media files helps enterprises perform efficient querying — with an accuracy rate above 90% from our initial test results in identifying duplicate media files.
Breast Cancer Diagnostics
Breast cancer is a leading cause of mortality for women in the U.S. The Breast Cancer Organization has estimated that 266,120 new cases of invasive breast cancer will be diagnosed in women in the U.S., and 63,960 new cases of non-invasive (in situ) breast cancer. Today, breast cancer diagnosis is still a lengthy process, involving multiple tests and other processes. Hence there is a definite need for ML/AI to partially automate breast cancer diagnosis.
Solution. The AI system uses ML and computer vision technique to analyze images of tissues and predict which lesions are most likely to become cancerous. This approach has huge potential to reduce the specialist’s workload in a typical pathology lab and to improve the incidence of early breast cancer detection.
Benefit. Early detection of cancer greatly increases the chances of successful treatment. The ML system ultimately could provide a more targeted approach and help women make more informed treatment decisions.
How AI Will Inform the Way Forward
ML holds the potential to create huge value for businesses by helping them solve pressing problems. Yet, key challenges remain. They include:
Data security. Enterprises must provide access to their data sets to implement ML correctly and efficiently, but for security reasons the data remains tightly secured and inaccessible.
Infrastructure. This is a must to test the various ML models with different tools. A lack of infrastructure may undermine frequent tests in developing the best possible outcomes.
Talent pool. Creating an ML model requires domain expertise and technologists who can solve the problem — but there are few experts to be found in the industry.
Business model. Implementing ML requires organizations to change their mindset, as well as implement a proper strategy for managing the change. There are few tools, frameworks and best practices to help guide business leaders forward.
If applied right, ML has boundless potential to improve business practices, increase return on investment, and satisfy customers with future-ready platforms and solutions. As a result, we recommend the following:
Start from business decisions. These should be the key strategic decisions that affect the business, and the related metrics that need improvement.
Identify an appropriate business problem. Then solve it with the right ML framework and packages.
Build business-problem-specific data sets. The data set should be labeled (by structured or unstructured data) and aligned with the problem being solved.
Pilot with parallel runs. Conduct and compare ML solution outcomes with those enabled by human decision-makers. Assess and iteratively improve the performance/accuracy of the ML solution.
Scale the ML solution. Build the solution with necessary hardware/software architecture and with associated infrastructure.
Manage change. Institute a broad change management program to shift the internal decision-making mindset.
To learn more, read “Machine Learning: The First Salvo of the AI Business Revolution,” visit the AI & Analytics section of our website, or contact us.