Chapter 4. Challenges to Delivering Value from Custom AI Development and Engineering Countermeasures
Deep learning can have a profound impact on the enterprise, but developing and implementing it is not easy, and organizations will face unique challenges that are unlike those that accompany the adoption of other technologies.
Despite amazing breakthroughs in artificial intelligence (AI) software and hardware, organizations must confront the poor interoperability of open source software components and the need to optimize highly specialized hardware, not to mention the challenge of first accessing and then harnessing both high-value and high-velocity data, working across multiple cloud environments, and doing all of this at scale. Further, deep learning methods are a radical departure from traditional statistical and machine learning techniques. As such, they can challenge even advanced data-driven organizations.
In terms of operationalizing, most organizations struggle in the transition from insight to action because of analytical systems that are incapable of reliably serving millions of decisions at the speed of real-time business. Many will also underestimate or discount the governance and risk management aspects of developing AI solutions, elements that must be considered for a successful strategy.
Figure 4-1 summarizes the barriers to AI adoption discussed in this chapter. We will also address some approaches that we can take to address them.
Figure 4-1. Five pillars of enterprise considerations for AI initiatives
Strategy
AI capabilities are not just about data and models. As a cross-functional technology, we will need to implement and socialize AI across the organization, an effort that will require coordination with many areas like legal, line-of-business, security, and compliance and regulation. Though developing an AI solution will be a significant technological challenge, its successful operationalization will ultimately be a human one requiring skillful change management and a deep understanding of your organization’s operational processes.
One of the most crucial steps of a successful AI initiative is achieving executive buy-in and support. Here, organizations can encounter a variety of difficulties. The great deal of hype surrounding AI can lead to inflated expectations of what can be achieved, sabotaging a program’s success. Even when it isn’t overhyped, the value that AI can bring to an enterprise can be poorly understood, leading to overt skepticism and inaction. In addition, you might need to confront indifference, open resistance to change, or the tendency for the organization to focus on short-term emergencies instead of long-term change. Furthermore, even though it will be necessary to express the AI solution in terms of its business value, it can be difficult to communicate its potential and path to production accurately across different kinds of stakeholders in your organization. The resulting lack of a shared AI vision can make it difficult to execute anything beyond an AI science experiment or point solution.
Executive education—at an appropriate level of complexity—will be key to set those expectations and determine the right solution to build. A shared AI vision comes from gaining a mutual understanding across key stakeholders of a series of achievable AI applications that are aligned to the strategic goals of the enterprise. Peer benchmarking and collaboration with experienced partners are particularly helpful in uncovering applications that are currently achievable.
AI is advancing rapidly. As such, your strategy should not only contemplate business outcomes, but also ensure that the decisions that are made lead to a robust and future-proof platform. A good starting place is to conduct an independent review of your current capabilities, identify the most important gaps, and develop a high-level roadmap to address them.
To spark and sustain strategic momentum, it is highly recommended to demonstrate proof of value with AI using an Agile, iterative process to find new, actionable insights in your data. This incremental approach will give stakeholders the chance to experience the AI solution’s power and become its advocates as the technology proves its value.
Technology
Working with deep learning means working with open source code, and although there’s no doubt that AI owes much of its advancement to open source software, there are some drawbacks.
Much of the cutting-edge research and code is academic grade and must be retooled before being used in the enterprise. The field is still relatively immature, and the code lacks many of the features that enterprises demand (e.g., security and usability).
The fragmentation and lack of interoperability between components can be difficult to navigate, especially because of significant gaps in support. Many vendors have taken a lock-in model, offering support only to those running code in their respective clouds. This is at odds with the hybrid model of cloud and on-premises computing that most enterprises prefer.
Deep learning also requires specialized hardware, and any data science team working with it must know how to manage and share a cluster of graphics processing units (GPUs). Working within these GPU compute frameworks can be challenging because their architectures are much different from those that are CPU-only. Significant engineering is required to optimize the software to ensure efficient parallelism, manageability, reliability, and portability. And because not all learning will occur solely in GPU architectures, they must also integrate with the rest of the analytical ecosystem.
Operations
Whereas many companies are comfortable with analytics in batch processing, deploying neural nets at scale is a completely different type of data analysis. Many companies lack the infrastructure required to use their data in a way that can service fully scaled and productionized deep learning models. In fact, this was cited as one of the top barriers to AI in the enterprise according to a 2017 survey of 260 IT and business decision-makers at a VP level or higher from organizations with a global revenue of more than $50 million per year. Almost all respondents (91%) expect to see barriers to AI realization when trying to implement it across their business. Figure 4-2 summarizes these anticipated barriers with lack of IT infrastructure and access to talent leading the challenges.
Building modern capabilities on top of or alongside legacy systems can be difficult, expensive, and tedious. It can present challenges like maintaining service-level agreements (SLAs) while moving data in and out of a mainframe and through a GPU or integrating data from customer relationship management (CRM), enterprise resource planning (ERP), or other enterprise software. Depending on the industry, there could be additional challenges such as lack of interoperability between diverse sets of equipment, each with its own proprietary control system and data interchange standard.
Figure 4-2. Barriers to enterprise AI realization
After you complete the tricky work of scaling the models over multiple servers, datacenters, and GPUs, they must also be managed, monitored, and automatically retrained.
AnalyticOps
Training models that work well in the lab environment can—in some ways—bear little resemblance to a production environment where they are relied on to make decisions. To maintain, manage, and monitor them, you will need to develop a culture and capability of AnalyticOps.
AnalyticOps is a system of best practices that has risen as a response to the demands of supporting automated decision-making. It applies the principles of DevOps and IT operations to data science so that companies can better manage their analytics resources and projects.
AnalyticOps provides a fixed framework for operationalizing models, helping automate your processes to ensure that you don’t have more work in progress than you need, that you’re monitoring the work you do have, and that people are adhering to your processes.
It helps organizations to be smart about their data science resources, putting applied analytics in the hands of IT operations so that data scientists need to be called in only when something unusual happens. Though creating AnalyticOps might seem like overkill at the start of an AI project, it will be essential for scaling it.
Model Transparency
Model interpretability is another key challenge because deep learning techniques make it more difficult to explain how the model arrived at a conclusion—the information they use to make decisions is hidden away, quite literally in hidden layers. Whereas the machines’ output might occasionally seem self-explanatory, such as when algorithms correctly identify images, in other cases it might be completely opaque.
The ability to explain these models is often an imperative, however, especially in cases for which laws prevent decisions based on data that can be considered discriminatory (e.g., approving or denying a loan) or cases that involve significant exposure to litigation (e.g., medical diagnosis).
For example, financial institutions in Europe must remain in compliance with the EU’s recently enacted General Data Protection Regulation (GDPR), which levies heavy financial penalties on companies that cannot explain how a customer’s data was used. In this case, it is not possible—nor is it legal—to tell a customer that their financial transaction was declined simply because the model said so. Even aside from regulation, leaders often need to know the factors that went into a model’s decision in order to trust it.
Though the problem of model interoperability is far from solved, enterprises are using several approaches to address it. One approach uses a method called Local Interpretable Model-Agnostic Explanations (LIME), an open source body of research produced at the University of Washington. LIME sheds light on the specific variables that triggered the algorithm at the point of its decision and produces that information in a human-readable way. In the case of fraud, knowing this information can provide security from a regulatory standpoint as well as help the business understand how and why fraud is happening.
The Move to Autonomous Decisions
After the models are built, making AI decisions operational at scale is one of the most significant challenges that companies will face as they adopt deep learning. AI is not like plugging in another app, and it is unlike current forecasting models, in which a report is delivered to a human who then makes a decision. Even for cases in which decisions are currently algorithmically driven, it is a daunting leap between trusting static models and trusting those that learn. This is automated decision-making, and it is different. It can disrupt your enterprise’s processes and mean that you need to entirely rethink ones that were designed pre-AI. New types of cognitive platforms driven by voice, perception, natural language, and so on will especially require reimagining how humans interact with machines.
Many organizations will struggle here because of analytical systems that are disconnected from the front line where second-to-second decisions will be made, or because they rely on manual processes that can’t keep up with the rapid pace of deep learning capabilities.
Using a model’s decision not only touches on explainability and transparency, but also consent, policy, ethics, and security; elements that can make or break the success of an AI solution in an enterprise setting. And, ultimately, the limiting factor of applying these advanced techniques will not be technology considerations—it will be managing the risk associated with pushing a high level of complexity into an operational context.
Just as enterprise risk management currently touches every part of the business, AI risk management and governance will need to incorporate perspectives from across the entire organization to support its adoption. Ultimately, it will be the coordination of these efforts that will build consensus and enable the transformation to a data-driven, analytics-based culture.
Operationalizing an AI solution is not just a question of building a machine that makes the right decision most of the time. It’s a question of how to build it with a level of sophistication that enables it to produce a decision based on a blend of considerations, describe the reasons and evidence for its decisions, remain compliant with all relevant policies and regulations, and do this all fast enough to be useful in real time and at an extremely high degree of accuracy.
Because they are crucial to the project’s ultimate success, operationalization concerns must be kept at the forefront of the project from the beginning. Leaving them for later risks seeing the project wither.
Executive buy-in and a commitment to redesigning core processes will be crucial to see success and work through some of the larger organizational changes. At the same time, building transparency and trust into your models from the start helps facilitate their adoption, enabling humans to move forward and act on model decisions.
Data
Deep learning requires the ability to fuse data from multiple sources of transactions, interactions, and rich media. Deep learning models are particularly adept at finding patterns and correlations across multiple heterogeneous subject areas. As an example, predicting failure of a complex industrial asset has the greatest chance of being accurate when diverse subject areas such as maintenance records, operational usage data, master data, and sensor readings are all factored in. As deep learning algorithms become mission critical to ongoing business operations, the data foundation must be resilient and hardened, facilitating efficient analytic operations.
However, enterprises are at different levels of maturity in being able to connect the dots across their internal and external data sources. Even for the most advanced organizations, deep learning often requires revisiting data management principles, particularly with respect to labeling, in order to properly train new supervised and unsupervised deep learning models.
In fact, 80% to 90% of the total work of a machine intelligence project will be in the data preprocessing—getting access to and cleaning the datasets and setting up data pipelines. This work should not be underestimated and can come with a truly staggering number of potential problems of different varieties.
Some organizations will struggle with even getting access to the high-value structured and unstructured data on which the deep learning project depends. The data itself might be mislabeled or unlabeled—a problem that is especially tricky for unstructured environments. You might also have data that is missing. Or, you might find that your data is imbalanced or that you can’t get different features available for analysis in real time.
This is just a sample of the nefarious, subtle, and downright maddening issues that you can confront in data preprocessing. Data snags can persist after processing, as well—one of the most overlooked problems is training your model on data that does not look like the data that will be used in production, so you end up with the AI equivalent of learning to swim in the ocean by practicing in a goldfish bowl.
The antidote to these issues is to implement proven design patterns for both tightly and loosely coupled data management techniques in order to form an Agile data foundation. This will be crucial for success in model training and discovery because deep learning projects must embrace experimentation and rapid iteration to determine the highest-performing network architectures and what techniques can be applied.
Talent
So much of deep learning’s success depends on the people behind it and their capabilities for guiding the program through each step. Many problems with deep learning are unique to the domain, such as model selection and the requisite data preprocessing, and require an experienced eye.
Unfortunately, there is a severe shortage of people who have production-level experience with deep learning in the industry, making it another one of the biggest challenges for enterprises that want to develop deep learning capabilities.
Though more and more people are acquiring deep learning knowledge through avenues like massive open online courses (MOOCs), it is rare to find people who can do it in practice, and the difference is important. The classroom is certainly a step in the right direction, but it does not replace real-life experience.
Companies have several choices when looking to acquire the talent that they need for their AI programs. Hiring or developing it is an option, though you should expect competition from digital giants and startups, as these have proven themselves to be magnets for both new grads and experienced practitioners in the field.
Acquiring other companies and their IP and talent is another route, or companies could also bring in a partner to gain external experience. To get the most value here, you should work with a partner that both has access to state-of-the-art information and will engage in knowledge transfer.
Conclusion
The challenges of working with deep learning are both technical and organizational. They begin at aligninng key stakeholders on where deep learning can be applied to address business priorities. Then, to execute a deep learning initiative, you will need to assess and address current capabilities with respect to data management principles, the analytic platform, and talent, and make plans to socialize and operationalize the solution across different business functions.
Over the next two chapters, we examine several custom AI solutions built to harness the power of organization-specific data. We also do a deep dive into how Danske Bank used machine and deep learning for fraud detection and its approach to some of the technical challenges discussed.