Don't miss out on anything new and exciting in the world of
 deep learning. Sign up to get our weekly digest in your inbox.

Deep Learning and the Innovator's Dilemma

Mostly everyone seems to agree that AI, carried by a wave of deep learning breakthroughs, will 'disrupt' industries left and right. Of course, the term disruption is often somewhat carelessly and imprecisely bandied about to refer to technological or business model innovations that threaten industry incumbents. But what exactly enables an innovation to be disruptive given the fact that incumbents, generally, have both the know-how (in fact they are often the source of the innovation) and the resources to get a head start on any entrant is often glossed over.

Before we turn to the case of disruption through deep learning innovations, let us briefly explore an answer to this question presented by Clay Christensen and his seminal thesis on the Innovator's Dilemma. The central thesis is that incumbent firms operate in a certain context of customer needs, suppliers, target markets and competitors that form a value network which sets the standard of value for any strategy or business decision including where to allocate resources and which innovations to pursue. In practice this means that firms will often pursue sustaining innovations, that is innovations which improve and sustain the firm's position within the established value network.

A Value Network is the context within which a firm identifies and responds to customer's needs, solves problems, procures input, reacts to competitors and strives for profit

A concrete example, cited in the book, are disk drive manufacturers in the late 80ies whose value network dictated innovation along the axes of capacity and not size, since their customers such as IBM's PC Division had no use for smaller drives that came at the expense of capacity. A value network mirrors this nested physical product architecture, since each part is associated with a supplier, a corresponding market and a distribution network. This means that the profits at the level of an established manufacturer of disk drives depend on providing what is valued most in the context of this network. Taking the innovation of smaller drives out of this value network and into the nascent and unproven market of laptop manufacturing was a risky and in the short term unprofitable proposition. This left the opportunity wide open to new entrants who were not locked in by the need to maintain high levels of profit and revenue or bound by a value network that had been built up over the company's lifetime. As the laptop market grew beyond early adopters and PCs became smaller, however, smaller drives began to penetrate ade the value network of the incumbents, leaving them vulnerable to disruption by those startups who had perfected those technologies.

Illustration of value network

This is a really rough sketch of Christensen's framework which hardly does it justice and I highly recommend you consult one of his articles or the book for a thorough treatment. It should suffice, however, for our purpose of identifying vectors of disruption that are made possible by ongoing innovations in deep learning.

The Automotive Industry

Let's first look at the automotive industry. In the U.S. Tesla, Google, Uber, Nvidia and a host of startups are hacking away at the autonomous driving technology vying to take the lead in this fledgling market. Meanwhile, German car giants seem to be chronically behind the curve, first giving the lead in electric mobility to Tesla. And now, realizing that electric mobility is a stepping stone towards autonomous driving, they are about to miss out on what is shaping up to be the greatest revolution in mobility since the invention of the car.

True to Christensen's framework, German companies have been the cradle of the innovation that is poised to disrupt them. As early as 1986 Mercedes Benz launched a project pioneering practical self driving car technology with impressive results; the reengineered Mercedes Benz W140 S-Class we see in the video drove almost entirely by itself over 1,678 kilometers (1,043 miles) from Munich to Copenhagen back in 1995. The project wasn't just abandoned thereafter, but led directly to the semi-autonomous driving capabilities Mercedes boats today. Since then Germany has hardly been a laggard on the research front; about 60% of patents for autonomous driving worldwide have been registered by the R&D department of a German auto giant (Volkswagen, Daimler or BMW) who collectively spend some €22 billion on R&D per year. On the flip side early investment in computer vision research produced a generation of researcher who are heavily invested in traditional computer vision approaches and manual feature engineering, a difficult environment for an upstart engineer to pitch some neural network (didn't NNs turn out to be a dead end anyways) black box which would obsolesce senior colleagues' decades of research effort.

Across the pond in North America, however, research into computer vision was unencumbered by the biases and pratical consideration of any one particular industry. Similarly, deep learning found its way to the car industry not through the GM R&D department, but through various Silicon Valley AI labs. More to the point the car company to first take a serious stab at autonomous driving is not ensnared in an ossified web of suppliers or the whims and preferences of a mass market. Which brings us back to the German auto industry which is characterized by a complex and deeply nested automotive supply chain with thousands of OEMs, many of which are family businesses that are almost as old as the automobile itself, who excel at classical mechanical engineering producing world-class components like the combustion engine or gearbox rather than cutting edge software.

On the consumer side of the equation, companies are faced with the difficulty that their brands are wedded to the combustion engine and the steering wheel. German consumers still very much value their 'Fahrvergnügen'; the recreational or life style component of driving a car. The sense of freedom on the Autobahn, the combusting motion of the pistons and roaring engine; driving as a Lebensgefühl (roughly 'sense of life') holds an overall more positive position in the German consciousness than it does in the American where distances are greater and driving is a more mundane and pervasive aspect of life, too much of which is spent on congested freeways. While self-driving capabilities could well be popular as premium feature of high end cars, there is no obvious mass market strategy for German OEMs. Quite the oppsite is the case in the U.S. where Ride hailing services like Lyft and Uber form the logical point of entry for self-driving car technology and the only change in consumer behavior is an increased tolerance for machine over human control. The groundwork for this commercialization has already been laid through a partnership between Lyft and Google's Waymo. Since there will be predictable demand by ride hailing services and indifference by the enduser to traditional differentiating car features (as long as it's clean, safe and air conditioned do you care about what car your Uber driver picks you up in?) the mass produced self driving fleet-car will be a high turnover commodity item most likely sold in bulk and on a recurring basis. Should such a triumvirate of ride hailing service, AI company and car manufacturer form, Ford or GM are much more likely to be a part of it than Volkswagen or BMW.

The reason there are no ride hailing services in Germany or more broadly Europe brings us to the last node of the value network; the regulatory environment. Not only are autonomous driving systems not permitted by the German Road Traffic Regulation, Uber is effectively banned in Germany. One can easily envision an alternative reality where a German car company like Mercedes partners with a German Uber clone (after all if there is anything German tech entrepreneurs excel at it's copycatting) to develop self driving car capabilities in a market outside of their traditional value network. Given that this venue is barred until even the regulatory state realizes that 'the times they are a changin' it will by definition be too late. When the time has come for the U.S. victory of the self driving car race to send its fleet on the roads of Frankfurt, the German auto industry's bacon will have been taken or more formally in the terminology of Christensen they will be stuck on the wrong one of two intersecting technology S-curves.

TLDR: So far the particular value network of German car companies composed of their own AI research departments, the supplier network, customer preferences and the regulatory environment has provided strong incentives for neglecting autonomous driving technology relative to prospective entrants from the U.S. who are poised to disrupt them.

CRM Software

In order to illustrate that there is a multitude of ways in which deep learning can bring about an innovator's dilemma, let us shift our attention to the example of traditional CRM Software. As an example from the software industry Christensen cites Intuit's accounting software Quickbooks which set itself apart by providing a radically simplified interface, in contrast to incumbents whose software required domain knowledge in accounting and was overloaded with complicated analyses, reports and audit trails. While this was the feature set desired by their most profitable enterprise accounts, it overshot the functionality required by the growing market of small businesses and single proprietors. Quickbooks managed to capture 70% of its market within two years of its introduction.

Now imagine what would happen if you completely transformed the user interface of a CRM, when instead of manual data entry, lead qualification, scoring and nurturing your CRM learned to qualify inbound leads both by analyzing the raw text of their request and external data sources, prioritized opportunities for you and simply plugged right into Slack to let you know whom to talk to next. There is a host of similar deep learning facilitated features imaginable which would move the classical web form based CRM in the direction of a specialized virtual assistant. Large enterprises with legacy customer databases and huge sales forces are unlikely to be early adopters of such a radically new interface which makes this model less interesting to Salesforce than to a startup seeking to disrupt it.

This briefly sketched example is instructive because it highlights the user side of the value network, where a significant change in the way users interact with a technology favors companies who are not locked in by a large, profitable customer base that is already used to and invested in a particular way of interfacing with a product.

Broadly, the pattern that emerges is that there are two types of industries and companies who are affected by advances in deep learning; those where AI is a sustaining innovation and those where it is an all or nothing proposition. In search, ad placement or e-commerce recommendations and matching, deep learning might bring about huge performance leaps but there is nothing disruptive about it. With self driving cars and virtual assistants incremental gains would be either lethal or extremely annoying respectively. Companies in this latter category where there is a clear performance threshold to the practical use of AI are the ones who are most vulnerable to disruption by deep learning.