What’s Different about Enterprise & Industrial AI
and Why Does it Matter?
by Matt Denesuk, Ph.D., Chief Data Science Officer, Noodle.AI
Why does AI matter to industrial firms and traditional enterprises?
Throughout modern business history, industries that have become substantially all about “data + math” have seen both enormous gains in productivity as well as disruption of business models and business leadership. Today, we call this digital transformation. Financial services and banking, which began their transformations at the dawn of the computer age, are among the earliest examples of industries transformed. Since then we’ve seen many more examples in telecom, media and entertainment, retail, and other industries. Firms like Netflix and Amazon have become leaders in both broadcasting and content creation; Amazon is dominating increasing amounts of the retail and distribution sectors; and online video games and social media have overtaken more traditional forms of recreation.
This has dramatic implications for firms currently in these industries who have not fully embraced digital transformation, as they face displacement or gradual marginalization if they are not quick to establish a strategy and rapidly begin executing.
In industries that are lower on the curve in terms of digital transformation, opportunities are present to establish new leadership and drive the very transformations that will reshape these industries in the future. This is especially the case with very physical industries such as transportation, energy, steel, and mining. Early-adopting firms are presented with opportunities to transform the way things are done, resulting in changes in business performance and the firm establishment of new leaders.
What are Enterprise AI & Industrial AI?
We term the core technologies and processes driving these transformations in the modern era “Enterprise AI & Industrial AI.” There is an overwhelming level of noise about “AI” right now, and it can be very difficult for business leaders and practitioners to keep up, to understand what it all means to their business, and how to get started in applying it.
Mature AI
Understand first that most of what you hear from the media today (including the tech media) refers to a class of problems that may be termed “Mature AI.” Such problems generally have the following characteristics:
1. The problems are clearly identified and articulated. It is taken as self-evident that the solution is intrinsically dependent on AI.
2. The data sets used are massive, well-curated, and often purpose-generated. For example, the sensors and data capture policies for self-driving cars were designed and placed from the beginning to generate the very data that AI models would need in order to be trained effectively. These cars are then driven billions of miles in the real world–and even more in a virtual sense via simulation¹, which can be used to generate enormous volumes of additional data and training. The data in Mature AI areas are also typically well-labeled; for example, images labeled with their contents, faces linked to correct individuals, or text in one language linked to correctly translated text in another language.
3. The business value of the work is generally taken as self-evident, and there is a shared expectation of slow, steady progress with substantial sustained investment presumed. It’s taken for granted that a 0.5% improvement in the accuracy of a face recognition algorithm is worthwhile, or even a slight reduction in risk for a self-driving car is worth achieving and paying for.
4. Mature AI areas often have huge global research and engineering communities associated with them. These communities curate and publish journals and hold information exchange conferences dedicated to the vital technical elements in these areas. One can find many professional journals, conferences, and societies dedicated to speech recognition, image and object recognition, and so forth.
5. The performance of the models and applications in Mature AI application areas is generally driven incrementally, in a steady way that can be likened to a “Moore’s-Law-like progress.” The central contributors to this are:
- Continual additions of more of the same kinds of data — more labeled images, more translated docs, more miles driven (real or virtual), etc.
- Riding the curve of increasing processing power and architectures. This includes the steady increases in CPU or GPU speed and parallelism, improvements in storage and retrieval systems and so forth. One just keeps upgrading equipment and systems as they become available.
- Tweaking of core algorithm performance. AI models have many “knobs” that can be tweaked, such as structure and depth of a neural network, the various hyperparameters that finely tune the inner operations of many algorithms, and finer segmentation of key problem cases. After initially finding appropriate settings, continuing to tweak these can result in modest improvements over time.
Emerging AI
It’s easy to make the mistake of thinking about problems in “Emerging AI” areas as if they were in Mature AI areas. This can lead to confusion, misapplication of resources, and failed AI efforts.
In particular, most problems in Enterprise AI & Industrial AI are not mature in the sense described above, and in fact typically fall into either what may be termed “Emerging AI,” or, for those that have moved somewhat toward a more mature path, “Transitional AI.” The strongly differing characteristics of such problems from Mature AI problems requires focusing on different kinds of activities and taking different approaches than those that are suitable for Mature AI problems.
“Emerging AI” problems generally have the following characteristics:
1. The problems are often ill-formed as AI problems, meaning that there is no clear linkage between a nominal business problem being posed and a technical problem whose solution would materially help address that business problem. In these cases, the ultra-critical step of defining what the technical problem is has not yet been completed. This is an opportunity, because it means there is great flexibility in terms of mapping the business problem to a good and tractable technical problem, but it also means there is much work left to do before even starting in earnest on the model development. It also means there is not very much precedent or prior work for one to leverage or build upon. Another complication can result from the fact that the business problem in question is generally currently being addressed by other means (e.g., spreadsheet calculations, rule-based software, dashboards + humans, etc.).
2. The data sets tend to be scarce, heterogeneous, sparse, and of low quality. Enterprise data capture systems and policies were generally put in place for compliance or reporting, not for AI. Industrial operations data capture systems and policies were generally put in place for control, basic monitoring, or post-event trouble-shooting. If data are not essential for one of these things, they are typically not recorded. Even if they are recorded, they are often discarded after some nominal holding period. These restrictions severely hamper the modeling approaches that can be used, as well as the ultimate model accuracy that can be achieved. In addition, each customer tends to have their own unique data models, data capture policies, and data architectures. A developed solution for one environment or customer does not generally apply to another environment or customer, even when the nominal business problem is quite similar.
3. Potential business value is not taken for granted. Many customers require being convinced that AI can be relevant to their business, and that an investment in AI is justified. They typically want to take slow, baby steps and see measurable business impact before justifying the next step of incremental investment. This requires communicating effectively with and winning over diverse sets of stakeholders. It also requires creating a roadmap of delivered capabilities, where each step can demonstrate its own strong and positive ROI as it builds on the prior step.
4. Research and Engineering communities, relevant journal articles, and relevant technical conferences are generally fragmented and thin. There is typically no global community of deep expertise for Emerging AI practitioners to rapidly develop and share technology and ideas. Each team or provider company is essentially its own, segregated community, which slows development of best approaches and practices considerably.
5. The performance of the models and applications in Emerging AI application areas is thus driven by very different factors than for Mature Areas. Model or application performance in Emerging AI is generally driven by creative invention in several specific areas:
- Problem Framing and Setup. “Problem Framing” is the understanding of a business process and important key performance indicators (KPIs), the subsequent articulation of one or more math problems that relate to key constraints in that business process, and the assessment of whether suitable data are available. “Problem Setup” is the process of constructing a particular data and modeling workflow that generates output that can be used to improve the business process. These are both typically adjusted in an iterative manner until a workable solution that moves the KPI sufficiently is obtained. Successful problem framing is currently part art, part science, but is absolutely essential to producing a high-value solution.
- Acquisition and use of new data sources. The introduction of a new data source, whether internal or external to the company, can often have more impact in the short term than model or algorithmic improvements. For example, things like the use of weather data for modeling tire wear, using hotel occupancy rates or pricing for models for airline revenue management; or leveraging current commodity prices as an additional input for optimizing financial performance of a manufacturing operation. These types of external data sources can significantly improve model performance.
- Effective use of domain knowledge and prior experience. Especially because the data are typically scarce and lacking in other ways, incorporating prior knowledge, for example in the form of engineering or theory-based equations, highly engineered model features, and even partially pre-trained model components, can enable an effective solution to be built despite the severe inadequacies of the data². In some cases where accurate simulations are available or can be created, such simulations can be very helpful as well.
- A yet unknown breakthrough. There is so much economic and societal value to be unlocked that many organizations (public, private, for-profit, not-for-profit) in many nations are feverishly working on breakthroughs to address this³.
As you might expect, in terms of basic model performance parameters Emerging AI solutions are typically substantially below those for Mature AI problems, as one would expect in terms of model performance parameters. Thus the performance levers for such Mature AI problems, which tend to create impact in small, incremental amounts, are not as helpful for Emerging AI problems. The performance levers for such Emerging AI problems (outlined above), while requiring more creativity and time, can result in step changes in model and application performance, and thus should be the primary focus in such problem areas.
In Closing
AI modeling thrives with simple problem statements and massive amounts of high-quality data. Data scientists are not usually blessed with such situations when trying to address critical business problems for Enterprise and Industrial firms. The massive value that can be generated by the successful digital transformation of these kinds of organizations is giving rise to major efforts across the globe to make AI relevant for them.
It’s critically important to recognize that the approaches being used successfully to continually improve performance in Mature AI areas will not apply successfully to most Emerging AI problems. We must recognize the unique aspects of these immature problem areas and employ a set of approaches designed for them. This paper has given a broad outline of these approaches⁴. This should result in a maturing of many of these problem areas, gradually putting them on the more incremental and continual improvement trajectories we currently see with Mature AI problems. The resulting global economic impact will be staggering.
The sentiments expected as one moves up the maturity curve, along with the placement of a few selected problem areas, are shown below.
—
1. Mature AI also frequently has access to sufficiently accurate simulators, such as driving simulators, which are very mature and well-simulate the results of nearly any driving action in a virtual environment.
2. See Noodle.ai Whitepaper: “Enterprise & Industrial AI Modeling: How It’s Done”
3. See Noodle.ai Whitepaper: “Coming Breakthroughs in Enterprise & Industrial AI”
4. See future Whitepapers for more detailed elaboration.
—
Photo by Clark Tibbs on Unsplash, Figures by Matt Denesuk
Follow us here for more on the future of AI. You can also find Noodle AI on Twitter, LinkedIn, YouTube, and get our latest insights via our newsletters!