The Modular, Bolt-On View of DS/ML

Spectacular demonstrations of emerging DS/ML capabilities solve isolated and isolatable challenges without recognition of the conditions critical to the success of these demonstration. They encourage a modular, bolt-on view of DS/ML. This view encourages us to see DS/ML as an add-on with high interoperability: conditions do not matter. The view suggests that DS/ML can be added to existing software without reconfiguration; existing processes can be complemented, or augmented, by DS/ML without transformation. The modular, bolt-on view of DS/ML suggests that DS/ML can be deployed as a layer that sits atop or replaces existing products and processes, as neatly as desktop word processors seemed to replace typewriters.

But, even in the transition from typewriter to word processor, problems of translation required halting and stepwise adjustments from one to the other. The dot-matrix printer, TrueType font libraries, and skeuomorphic user interfaces (Laurel 2013) all filled in the gaps between how people had designed their work processes around the typewriter and the new, unique affordances of desktop publishing. The modular bolt-on view of DS/ML draws attention away from the very particular conditions that enable DS/ML successes as well as the expertise and labor that is required to conceive of, develop, and refine these technologies over time and often over many rounds of trial and error; this is at the heart of what we call the modular, bolt-on view of DS/ML. Spectacular demonstrations have broad appeal because of the human tendency to misunderstand what constitutes a computationally difficult problem, to see proof of technological capability as proof of pragmatic capability. Arguably, this human tendency is exploited in the choice of demonstration project to achieve reach and effect. AlphaGo Zero, for example, is a dramatic proof of concept of reinforcement learning (Silver et al. 2017). But, it is not a persuasive proof that reinforcement learning can accomplish tasks we as humans understand to be on the same order of complexity as the game of Go. Indeed, human intuitions for what are easy or difficult problems to solve do not map on to computational difficulty. This problem has long bedeviled AI researchers who have struggled to explain, for example, just how hard comprehension tasks are for computers, when they seem so ‘easy’ to humans. This human tendency to misunderstand what constitutes a computationally difficult problem allows spectacular demonstrations, like AlphaGo Zero, to recast a range of problems that seem ‘easier’ from the perspective of human intelligence as within grasp of being solved by the technologies that are showcased by the demonstration project.

DS/ML as the New Electricity

Because of the prominence of DS/ML as modular and bolt-on, exceptions to this narrative are worth examining. One such narrative that stands in exception to the modular addition of capabilities story is told by Andrew Ng, formerly of Baidu, now Co-Chairman of Coursera and an Adjunct Professor at Stanford University. He describes the challenges of adopting DS/ML as an emergent technology in terms of the challenges that faced industry around the turn of the 20th Century as the emergent technology of electricity began to replace steam power. At that time, electricity was far from the omnipresent and almost invisible commodity that it is today. Few aspects of the technology had been standardized, from voltages to outlet plug shapes, and ensuring that a new investment in electrification would pay off was far from certain. In Ng’s telling, “a hundred years ago, electricity was really complicated. You had to choose between AC and DC power, different voltages, different levels of reliability, pricing, and so on. And it was hard to figure out how to use electricity: should you focus on building electric lights? Or replace your gas turbine with an electric motor?” According to Andrew Ng, “thus many companies hired a VP of Electricity” (Ng 2016).

Similarly, DS/ML is, today, “really complicated”. Data can be local or distributed in the cloud. It is difficult to know whether and why to use a random forest algorithm or a neural network, or how to evaluate the success of any particular implementation. Furthermore, it is difficult to anticipate the costs of a project; the reliability and cost of machines, data storage, and engineering talent vary widely. And it is difficult to know where to focus one’s efforts; should one build an audience segmentation model first or a churn model?

While most commentators gloss Ng’s story under breathless headlines like “Artificial Intelligence Is the New Electricity?” (Eckert 2016), the story that Ng tells is more nuanced than one of simple metaphor-making when one focuses on the importance of the “VP of Electricity” to Ng’s narrative. It is also more nuanced than the modular addition of capabilities. Through this lens, his story is one of complexity in emerging technologies that requires dedicated expertise to construct new interfaces that mediate between the different needs of the different parts of an organization. For DS/ML, this means preparing data in a way that is easily ingestible, and constructing tools that simplify the underlying complexity but offer affordances for making use of tools that had been previously beyond the reach of non-experts. AlphaGo Zero is presented by its creators as a persuasive proof of reinforcement learning and its capacity to solve ‘complex’ problems.However, reinforcement learning works best on problems that have been adequately abstracted to sufficiently resemble the kind of closed problems that reinforcement learning can solve. That is, real-world problems must be made sufficiently deterministic, observable, discrete, simulatable, short, evaluable, and well-documented before they can be addressed by the emergent technologies embedded in AlphaGo Zero, as Karpathy points out above. Furthermore, these emergent technologies must also be reshaped to accommodate real-world problems, even in their abstracted conditions, as inputs. This work of abstraction and accommodation is drastically different than the work of software development attuned to understanding DS/ML as the modular addition of capabilities. And yet, most promotional materials for DS/ML tend to further this narrative, evoking the sense of magic that Elish and boyd identify in their work (2017). In business settings, these narratives fulfil specific functions; modular capabilities are easier to sell as products, and they are easier to explain to customers as discrete technologies. Furthermore, they lend themselves to the very same spectacular demonstrations that we have discussed above. These spectacular demonstrations are countered by Ng’s metaphor, which argues that until DS/ML can be utilized as easily as a lamp can be plugged into a standard wall outlet, a dedicated form of expertise will be required to make it have any value for an organization.

CONSEQUENCES FOR PATTERNS OF ADOPTION

Metaphors matter, they guide adoption of emerging technology (Sturken & Thomas, 2004). And, they shaped how the audience of stakeholders communicated with the DS/ML expert consultants that had gathered in that Midtown Manhattan conference room. Those stakeholders and expert consultants were gathered to develop and implement a data strategy (see above) for the entertainment media client. The goal of a data strategy is to help companies realize the potential, and potential value, of their data for their organization. Recently, over the past couple years or so, companies have started offering consulting services to help craft such data strategies, responding to a need in the market thereby acknowledging the difficulty of translating the spectacular successes of DS/ML into industry applications, from Amazon’s ML Solutions Lab¹ and Google’s ML Advanced Solutions Lab² to the startup Element AI³ (to mention the more prominent players).

As a producer of original content, from written text to short-form video, for a variety of different audiences, the client was interested in natural language generation, from de novo generation of content, from text to video, to the automatic tailoring of existing content to appeal to different sets audiences. In addition, they were interested in internal-facing chat bots to increase operational efficiency (e.g., a bot that suggests to re-publish existing content). This set of projects, while feasible, suggests a modular view of DS/ML. Furthermore, in preliminary meetings prior to the workshop, there was little to no concern for the conditions that allow DS/ML to succeed within organizations, yet another hallmark of the modular view of DS ML.

Over the course of the day, the consultants met with business stakeholders, content creators, software and data engineers, and data analysts. They started the onsite with a presentation on data science, machine learning, and artificial intelligence (AI) designed, one the one hand, to define a common language, and one the other, to set realistic expectations for what can be accomplished with DS/ML and the work it takes to achieve these possible accomplishments. In particular, the presentation was designed to create awareness for the conditions that need to be created for a long-term, successful, in-house DS/ML practice that could develop text generation algorithms and smart bots for internal efficiency.

Throughout this paper, we use the following definitions of data analytics, data science, machine learning, and artificial intelligence. We do not claim that our definitions are better than their alternatives, there are many competing definitions, in part because definitions suggest and drive the particulars of the adoption of new technology. Our definition of AI, for example, sidesteps a thorny issue (the definition of “intelligence”, which is highly political). Here, we merely define our use of terms for the purposes of this paper to avoid confusion.