Deep neural networks in Self-driving cars
Deep neural networks (DNN) have obtained many many breakthroughs recently in important fields of computer science like computer vision. These neural networks that use deep learning research have many different applications. One that is currently on the rise is self-driving cars (Di Palo, 2017). Significant process in deep neural networks in the last decade has enabled the development of safety-critical machine learning systems like autonomous cars. Major car manufacturers are building and actively testing these cars, including Tesla, Ford, GM, BMW and even non-car manufacturers like Google. Recently, autonomous cars have become very efficient in practice and have in fact already driven millions of kilometers without any human intervention (Davies, 2017). Legislation to enable testing and deployment of autonomous vehicles has recently been passed in multiple US states, including California and New York (Autonomous Vehicle Enacted Legislation, 2017). Now, deep neural networks are ready to fully disrupt the car industry as we know it.
What are deep neural networks?
Artificial intelligence (AI) is the simulation of human intelligence processes by machines, more specifically by computer systems. AI is incorporated into a variety of different types of technologies, machine learning being one example. Machine learning is the science behind computers acting without programming. Deep learning is a subset of machine learning, that can be thought of as the automation of predictive analytics (Rouse, 2018).
A standard neural network consist of multiple simple and connected processors called neurons, each of them are producing a sequence of real-valued activations. These input neurons can be activated either through sensor that perceive the environment, or through weighted connections from previously active neurons. Some of these neurons then may influence the environment by triggering actions (Schmidhuber, 2015).
Neural networks are learning by finding weights that make them exhibit desired behaviour, such as driving a car. Depending on how the neurons are connected and the problem to be addressed, getting the desired behaviour may require long causal chains of computational stages, where each stage transforms the aggregate activation of the network. The concept of deep learning is about accurately assigning credit across these stages (Schimdhuber, 2015).
Although shallow neural network models with only a few of such stages have been around for many decades, the training of deep neural networks with many layers had been found to be difficult in practice (Schmidhuber, 2015). DNN have attracted widespread attention since they outperform alternative machine learning methods, e.g. kernel machines, in several important applications (Vapnik, 1995). Also, in limited domains, DNN have achieved the first superhuman visual pattern recognition results and have won several pattern recognition competitions internationally (Schmidhuber, 2015).
DNN for self-driving cars
The key component of a self-driving vehicle is the perception module, which is controlled by the underlying deep neural network. The deep neural network uses input from different sensors, like light detection, cameras and ranging sensor, and an infrared sensor, that measure the environment and produce as output the steering angle, amount of braking, etc. necessary to maneuver the car safely.
Digital Technology: Definition
There is not just one definition of digital technology. A digital technology can be defined as a process that is perceived as new, requires some significant changes on the part of the adopters, and is embodied in IT (Fichman et al., 2014). It can also be defined as leading to the creation of and consequent changes in market offerings, business processes, or models that result from the use of the digital technology (Nambisan et al., 2017). Digital technologies have their own set of characteristics. Namely, homogenization and decoupling, connectivity, reprogrammability, digital traces, and modularity.
Deep neural networks can be defined as a digital technology as they are a new process embodied in IT that have created new market offerings and consequent changes in the business processes and models. Self-driving cars are a good example of how deep learning neural networks have created new market offerings.
Homogenization & decoupling
Homogenization means that all digital information assumes the same form. This means it can be processed by the same technologies, at least in principle. This in turn, enables decoupling. Meaning, digitizing offers the opportunity to remove the tight couplings between types of information and their storage, transmission, and processing technologies (Tilson et al., 2010). Because all digital information assumes the same form, deep neural networks can process all different types of information, like numbers, text, and images. Because homogenization allows information to be separated from its artefacts, this allows deep neural networks to be used in different situations. Deep neural networks also show this homogenization and decoupling, as all different sorts of inputs and outputs can be processed and produced by this technology.
Connectivity
Digital innovations often allow connections with other users, with other applications, and between firm and customer. Deep neural networks allow connections with other applications, for example with a car, as the output of a neural network triggers action in the other device, e.g. steering. In the future, deep neural networks in self-driving cars may be connected to many more other devices. For example, they could be connected to a traffic management system. If this system were to be guided by artificial intelligence, it could produce optimal traffic condition considering the information it receives from the connected self-driving cars about where they are driving, at what speed, and in which direction for example. Moreover, currently self-driving cars can be classified as connected if they can communicate with other vehicles or the infrastructure, e.g. next generation traffic lights. However, most self-driving prototypes do not have this capability, but this does show the opportunity of connectivity for self-driving cars (Bojarski et al., 2016).
Reprogrammability
Digital products can be reprogrammed and edited through for example software updates. This can happen either by the supplier, enabled by the product´s connectivity, or autonomously through machine learning. This is a very clear and important characteristic of DNN. Using sensors, processors, and actuators this technology can autonomously learn and improve, and is therefore qualified as a ´smart´ innovation. As explained before, machine learning is an important characteristic of deep neural networks (in self-driving cars).
Digital traces
Moreover, digital technologies all leave digital traces. Digital traces occur as a consequence of using digital devices or applications. One's unique set of traceable digital activities, actions, contributions, and communication that are manifested on any digital device or the Internet itself is known as a digital footprint. For a DNN, this consists of all their inputs, outputs they produce, and what they have learned while processing these inputs into actions in a digital environment.
Modularity
Furthermore, modularity is a characteristic of digital technologies. “A module is a unit whose elements are powerfully connected among themselves and relatively weakly connected to elements in other units” (Baldwin & Clark, 2000: 63). Modularity can be created by standardizing interfaces between units. Schilling says “Modularity [...] refers to the degree to which a product can be decomposed into components that can be recombined.” (as cited by Yoo et al., 2010). DNN are modular in the sense that a series of individual neural networks can be combined and moderated by some intermediary to create a new (modular) artificial neural network. Then, each independent neural network serves as a module, operating on separate inputs to accomplish a certain sub-task of what the main network aims to perform.
Consequences for products & services
Low marginal cost
Deep learning neural networks can process all different kinds of information, o.a. because of homogenization and decoupling of information, from text to images (Castelvecchi, 2016). Digitized information can be stored, transmitted, and computed in fast and low cost ways. This is in line with Moore´s Law, which states that the speed and costs of computing power will improve exponentially. This has affected deep neural networks as well. DNN are heavy computational and use huge amounts of data and memory. Because of digitization of information in combination with Moore´s law, the marginal costs have decreased to a point that DNN can be in an embedded system on an individual car and compute outputs in real-time. This thus has o.a. created the opportunity of autonomous vehicles at a reasonable cost.
Another opportunity for the development of new products and services is interoperability. Interoperability is a product or system´s ability to work with other products or systems. This is facilitated by standardized and open interfaces. By standardizing the interface for neural networks, they can be used for all different kinds of applications, of which self-driving cars is only one. If neural networks are not just build to control one particular type of car, they can be used to steer other types of autonomous cars as well.
Network effects & Smart
Another consequence for the development of new products and services can be network effects, which mean the value of a product increases as the number of users increases. In the case of deep neural networks this could allow a business ecosystem to arise surrounding the use of DNN is self-driving cars. These ecosystems can help nourish better software updates and thus help improve the quality of DNN behind autonomous vehicles. However, network effects usually apply to the users of the technology, in this case deep neural networks. All the inputs used in a DNN and the outputs produced by the DNN leave a digital trace. This digital data can then be used as information to improve the product or process, e.g. improving image recognition or learning from feedback on the decision. Especially in DNN, all the data it gathers from different streams of digital traces, it can use to learn and improve internal processes and outputs. As the installed base of users increases, the digital traces increase as well, allowing the neural network to learn and improve more and more.
Servitization
The consequences for the development of new products are continuous development of the process and servitization. Servitization is the shift towards the service that a product offers as a products´ value stems from the experience. As is the case for autonomous cars, which are subject to continuous development and have shifted the value of cars from the product itself to the experience and service that comes with it. Deep neural networks created an environment in which it is not just the product, the car itself that is valued, but the process of the DNN behind it and the experience and service that are provided by the self-driving car.
Industry-level dynamics
Intel and research firm Strategy Analytics predicted that in the years 2035 to 2045, worldwide, 585,000 lives would be saved by autonomous vehicle technology, public safety costs would be reduced by $234 billion, and 250 million commuter hours would be saved every year. Also, they predicted that the passenger economy would be worth $7 trillion by 2050 (Strategy Analytics, 2017). These numbers indicate the impact that the rise of autonomous vehicles is expected to have. Among the industries that will be most clearly impacted by autonomous vehicle technology are car manufacturers and ride-sharing services, and insurance companies, as these type of firms are heavily involved with automobile driving.
Automobile industry
Change of dynamics
It is expected that in the future people will no longer need their own cars and will be able to rely on a fleet of driverless vehicles to transfer them to work, home, or any other destination. With fleets of autonomous vehicles to hop in and out of, more and more car users may abandon traditional car ownership models. The car’s driverless characteristic comes with unique features and convenience which may lead to customers needing an on-demand usage model instead of traditionally purchasing a car for their family’s exclusive use. And even when customers do decide to purchase their own cars, the increased efficiency of a driverless car that doesn’t require the time of a dedicated human driver may lead to reducing the need for multiple vehicles per household, increasing one-vehicle households. The result and future of autonomous vehicle technology for traditional car manufactures is unclear, and it will largely depend on how quickly traditional car manufacturers can develop and incorporate the technology and the speed and extent that consumers adopt autonomous vehicles.
Potential disadvantage
One hypothesis is that car manufacturers will be affected negatively for several reasons. As a result of the elimination of drivers and the corresponding costs, fewer consumers will need to own cars since ride-sharing will be very cheap, as well as more convenient. Furthermore, as tech companies enter the autonomous vehicle market to compete, car manufacturers end up fighting for market share with more players, resulting in some companies losing market share and leaving the market.
Potential advantage
On the other hand, the rate with which people will give up car ownership may be exaggerated as consumers are used to the current model of car ownership, especially those in suburban and rural areas. Also for many consumers owning and driving a car may be more of an emotional decision. Thus, the other hypothesis is that the autonomous vehicle market could be an opportunity for car manufacturers as it possibly will expand the market for cars. By alleviating traffic and parking, and eliminating the need for driving, autonomous vehicles will become more attractive than other forms of transportation. Resulting in self-driving vehicles stealing share from other forms of transportation such as subways, buses, trains and air travel. As a consequence, the number of vehicle miles traveled and the number of cars needed may increase.
Insurance industry
Change of dynamics
Car insurance is an enormous business, with a market value in the United States alone of $200 billion (KPMG, 2017). Autonomous vehicles have the potential to significantly reduce the number of accidents on the road. Currently, more than 90% of accidents are caused by human error (Treat, J.R., 1979). With the elimination of the need for a human driver this statistic will be significantly lower, resulting in a smaller need for car insurance. With less accidents in the long run, the insurance industry will spend less money on payouts. Since insurance prices are based on accident rates, prices will fall as accident rates go down. The insurance industry is a competitive market with many companies fighting for drivers' business, giving price reduction even greater momentum.
Potential disadvantage
Berkshire Hathaway’s owner Warren Buffett, who owns several car insurance companies, noted that self-driving cars would hurt the insurance industry. Buffett stated that (CNBC, 2017), "If they're safer, there's less in the way of insurance costs, and that brings down premiums significantly." KPMG (2017) predicted that the insurance market will lose $137 billion of its value and shrink 70% by 2050. For the insurance industry, the implications of autonomous vehicles are clearly negative, and insurance companies will have to adapt to this changing environment.
Taxi industry
Ride-sharing
Ride-sharing services such as Uber and Lyft have had a major impact on the taxi industry. Now that consumers can arrange rides with an app, the value of taxi medallions in major cities has decreased as regulated taxis no longer have a monopoly (CrainsNewYork, 2017). With the rise of autonomous vehicles, the ride-sharing industry possibly experiencing another disruption.
Change of dynamics
Eliminating the driver from the driving process would make taxi services and ride-sharing significantly cheaper. Companies can save on labor costs and pass those savings on to customers. This may result in an expansion of the market as rides become less expensive than public transportation. Currently, Uber and Lyft are the leaders in the ride-sharing market. Competitors wanting to access this market have significant barriers to entry such as mapping technology, applications, and building a user base. Anyhow, autonomous vehicle technology results in the taxi industry getting more competitive. It is predicted that autonomous taxis won’t cost the economy jobs, as self driving taxis will be on the road 24 hours a day requiring more human maintenance (Uber, 2016).
Potential disadvantage
The taxi industry will be negatively affected by autonomous vehicle technology in case the industry doesn’t prepare for change. The industry currently enjoys a large customer base and the challenge for them is how the taxi companies are going to hold on to those customers and change their businesses quickly enough once the time comes for autonomous taxi competitors trying to steal their market share.
DNN and Organizations
The technology of neural networks has also penetrated on organizational level. Companies such as Uber, Google and BMW are focusing on the development of self-driving cars. This all occurs in a very dynamic but also competitive environment, in which each company strives to market the first autonomous car.
Neural network is the key-technology for organizations who are operating in the self-driving car industry. The importance of these networks for autonomous cars can be found in the need to make decisions in split-second. Neural networks ensure right decisions of the self-driving car in all possible cases happening on the road, like road-crossing pedestrians for example. Although incorporating neural networks in autonomous cars is needed for manufacturers to keep up with the competition, it is hard to develop such neural network (Pawsey, 2017).
Challenges for organizations
Development
The ability of using neural networks allows manufacturers to enhance their self-driving systems. As a result, nowadays some companies are testing self-driving cars yet, even in cities with complex infrastructure. Only a few companies are developing self-driving cars. It is quite hard to develop a self-driving system, including neural networks, what implicates that companies should have access to both financial as human capital. For example, Ford invested 1 billion dollar in an AI start-up. Ford invested in this start-up to develop an AI based virtual-driver system, so that Ford is able to deliver a full autonomous vehicle by 2021 (Muoio & DeBord, 2017). The leaders in the self-driving car industry are those who have access to financial assets, for example by having deep-pocketed partners. To develop a good self-driving system, a company should have access to AI knowledge as well. Partnering up with the right organizations and/or recruiting smart people with knowledge of neural networks is a must.
Ethics
When it comes to decision making of autonomous cars, several ethical questions rise. Imagine an autonomous car, driving on a road. At a certain moment, a group of 5 pedestrians appear just in front of the car. The car has to decide autonomously to dodge and hit something that will obviously lead to the car’s passenger’s death or to drive into the pedestrians leading to their death. Although this is an extreme example, this shows that ethics are quite important in building autonomous cars. When people make choices in driving a car, morality plays a certain role. Morality is an factor in taking risk, or not. However, to what extent should autonomous cars take risks? The extent to which an autonomous car takes risks is embedded in the system. One could argue that an autonomous car should not take any risk at all. However, if an autonomous car does not take any risk other problems with regarding to mobility will come into play. For example, if an autonomous car approaches a double-parked car someone would like to see the car pass that double-parked car safely. However, if an autonomous car is programmed to not take any risk, the car will simply wait until the double-parked car has continued driving. No one driving a car, would like to see their car stop and wait for just a double parked car. These examples are just a few, showing how ethics can be challenges for self-driving car manufacturers.
Opportunities for companies
The application of neural networks to enhance the performance of autonomous cars, is not just interesting to the (autonomous) car manufacturers. These developments are interesting for organizations in, for example, the taxi- or delivery sector. In these sectors, completely new business models can pop up (Nambisan et al., 2017). Consider the situation for taxi companies when no drivers are needed to bring customers from A to B. Uber, for instance, is already piloting a program with self-driving cars. One could be matched up with a self-driving car by using UberX in Pittsburgh (Uber, 2018). So, as the car industry is disrupted by neural networks, other industries follow. As a result, many companies should change the way they are currently working. In logistics we see opportunities too. Ford, who is one of world’s leading autonomous car manufacturers, is building a platform which allows businesses to use its autonomous cars for logistics. Businesses can partner up with Ford and use the autonomous Ford cars to deliver their products to customers (Thompson, 2018).
Organizations
With the advent of the neural network, not only new firms arose, incumbent players in the car-industry had to adapt in order to stay competitive with the new generation cars. An example of an incumbent firm that is experimenting with the building of autonomous vehicles is BMW. Traditional car-manufacturers used to focus mainly on the production of the traditional car. By the arrival of the neural network, these incumbent players, like BMW, now developed a complete new sector (Tian, 2018). Moreover, new players are developing rapidly new innovation with the use of the neural network technology. The start-up Wayve, originated in the UK, have developed a neural network which is smart en advanced enough to master a new car within 20 minutes. Due to these extreme improvements, traditional manufacturers are led with no other choice than investing in these technologies.
Incumbent firms
Incumbent firms have realized that neural networks are both a threat and an opportunity for their current business model. In order to stay competitive, they have invested money and devoted time and effort into research projects and experiments to develop this technology themselves and be able to extract the most value of it for their business (Hill and Rothaermel, 2004). A good example of an incumbent player who is adopting neural networks technologies in their current business is Ford. For more than 20 years, Ford is developing and conducting research on artificial intelligence and in specific on neural networks. It is of high importance for companies like Ford to stay ahead on all innovation regarding neural networks, otherwise new players will enter and will take over the market (Argawal, 2017).
New players
However, the advent of neural networks offers also opportunities for new firms in the market. These new players are exploiting the market by developing new technologies and features with the use of neural networks. New players experience an advantage in contrast to incumbent players since they are not constraint by already existing enterprise architecture. These constraints makes incumbent firms less flexible and responsive. Moreover, new players can merely focus on the systematic research of this technology where incumbent players have also other business activities. Another advantage for new players is fact that they can learn from other existing companies (Hill and Rothaermel, 2004). They have plenty of failed and less efficient examples and therefore new players can skip multiple research steps. An example of new player in the market of self-driving cars is Wayve (Wayve, 2018).
References[edit]
Agarwal, M. (2017). Adopting AI in the Enterprise: Ford Motor Company. O’Reilly.
Autonomous Vehicles Enacted Legislation, (2017). http://www.ncsl.org/research/ transportation/autonomous-vehicles-self-driving-vehicles-enacted-legislation. aspx.
Baidu Apollo. Retrieved from: https://github.com/ApolloAuto/apollo. (2017).
Baldwin, C. Y., & Clark, K. B. (2000). Design rules: The power of modularity (Vol. 1). MIT press.
Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., ... & Zhang, X. (2016). End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316.
Castelvecchi, D. (2016) The black box of AI. Nature, 538, 20-23
CNBC, (2017), Buffett has an interesting theory about why self-driving cars will hurt the insurance industry. Retrieved from: https://www.cnbc.com/2017/02/27/buffett-self-driving-cars-will-hurt-the-insurance-industry.html
Crains New York, (2017) Have taxis finally hit rock bottom? retrieved from: http://www.crainsnewyork.com/article/20170411/TRANSPORTATION/170419978/have-taxis-finally-hit-rock-bottom
Davies, A. (2017). The Numbers Don’t Lie: Self-Driving Cars Are Getting Good. Wired, February.
Di Palo, N. (2017). How to Train your Self-Driving Car to Steer. Retrieved from: https://towardsdatascience.com/how-to-train-your-self-driving-car-to-steer-68c3d24bbcb7
Fichman, R. G., Dos Santos, B. L., & Zheng, Z. E. (2014). Digital innovation as a fundamental and powerful concept in the information systems curriculum. MIS quarterly, 38(2).
Heineke, K., Kampshoff, P., Mkrtchyan, A., Shao, E. (2017) Self-driving car technology: When will robots hit the road? Retrieved from: https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/self-driving-car-technology-when-will-the-robots-hit-the-road
Hill, C., & Rothaermel, F. (2004). The Performance of Incumbent Firms in the Face of Radical Technological Innovation. The Academy of Management Review
Nambisan, S., Lyytinen, K., Majchrzak, A., & Song, M. (2017). Digital Innovation Management: Reinventing innovation management research in a digital world. Mis Quarterly, 41(1).
KPMG, (2017), Autonomous Vehicle Technology Could Shrink Auto Insurance Sector by 71% by 2050 & Change Type & Amount of Insurance Sold. retrieved from: https://home.kpmg.com/us/en/home/insights/2017/06/autonomous-vehicle-technology-could-shrink-auto-insurance-sector-by-71-by-2050-change-type-amount-of-insurance-sold.html
Muoio, D., DeBord, M. (2017) Ford just invested $1 billion in a secretive AI startup founded by former Google and Uber execs
Retrieved from: https://www.businessinsider.nl/ford-invests-1-billion-argo-ai-for-self-driving-cars-2017-2/?international=true&r=US
Pawsey, C. (2017) How Artificial Intelligence is key in autonomous vehicle development. Retrieved from: https://www.trucks.com/2017/11/29/artificial-intelligence-key-autonomous-vehicle-development/
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural networks, 61, 85-117.
Self Driving Ubers. Retrieved from: https://www.uber.com/cities/pittsburgh/self-driving-ubers/ (2018)
Strategy Analytics, (2017), Accelerating the Future: The Economic Impact of the Emerging Passenger Economy. retrieved from: https://newsroom.intel.com/newsroom/wp-content/uploads/sites/11/2017/05/passenger-economy.pdf
Tesla Autopilot. Retrieved from: https://www.tesla.com/autopilot (2016).
Thomson, C. (2018) Ford has big plans to cash in on driverless cars, but not in the way you might think, Business Insider Retrieved from: https://www.businessinsider.nl/ford-ceo-driverless-cars-logistics-will-be-enormous-opportunity-2018-1/?international=true&r=US
Tilson, D., Lyytinen, K., & Sørensen, C. (2010). Research commentary—Digital infrastructures: The missing IS research agenda. Information systems research, 21(4), 748-759.
Treat, J.R.;Tumbas, N. S.; McDonald, S. T.; Shinar, D.; Hume, R. D.; Mayer, R. E.; Stansifer, R. L.; Castellan, N. J.Tri-level study of the causes of traffic accidents: final report. Executive summary. retrieved from: https://deepblue.lib.umich.edu/handle/2027.42/64993
Uber, (2016), Self driving uber. Retrieved from: https://www.uber.com/ja-JP/blog/pittsburgh-self-driving-uber/
Vapnik, V., & Cortes, C. (1995). Support-vector networks. Machine learning, 20(3), 273-297.
Wayve. Retrieved from: https://wayve.ai/blog/ (2018).
Welch, D. & Behrman, E.(2018) Who's winning the self-driving car race? Bloomberg
Retrieved from: https://www.bloomberg.com/news/features/2018-05-07/who-s-winning-the-self-driving-car-race
Yoo, Y., Henfridsson, O., & Lyytinen, K. (2010). Research commentary—the new organizing logic of digital innovation: an agenda for information systems research. Information systems research, 21(4), 724-735.
This article "Deep neural networks in Self-driving cars" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:Deep neural networks in Self-driving cars. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
