Focusing on Machine Learning 2020: Augmentation Instead of Automation
In less than last two years the interest in machine learning has tripled and in five years it has more than quadrupled. Gartner’s famous hype cycle put it at the peak of inflated expectations on their 2017 report. During this overwhelming wave of hype we have looked at machine learning as a tool to fix everything from customer service, sales, human resources and marketing to health, energy and traffic — to name just a few.
When betting on hype many things can, and usually will, go wrong. We’re seeing massive problems in the healthcare industry trying to adopt new technologies (see [1], [2], [7] as examples), technologies taken too soon to fuel autonomous vehicles [6] and businesses trying to automate customer service, especially within the area of chatbots [3]. At the same time, some highly valued players are trying to save energy on a global scale ([4], [5]) and their plans seem actually plausible. All of these examples, and the vast majority of current AI endeavours, are interested in the automation of tasks. The automation of driving, automation of mundane tasks, automation of customer service, automation of marketing.
Since automation is not particularly hard, many of these areas are now filled with different players with similar offerings. As an example, virtually every possible marketing-related vendor is offering their own recommendation engines and personalization tools. Funnily enough, the underlying techniques of these systems are 95% of the same stuff — that being collaborative filtering — independent of vendor and price tag. You can buy a basic “machine learning tool” to fuel your banner placement much like you can buy a box of cereal down at the supermarket. It’s becoming a saturated market.
That said, automation is key wherein volume and scale are involved. Online retailers may sell hundreds of millions of different products [10] and thus new ways of automation are required. Recommendation engines relying on stale data (thus, ones that do not learn online, see [11] for details in online learning) or do not know how to explore new possibilities (see reinforcement learning in comparison to bandits in general [12], [13]) will quickly run out of steam when competing on AI. Furthermore, the world of automation for anything that’s being sold on the internet is now a world of machine learning. Not just personalization and recommendation of products, but prices, colours, copy-texts and heck, even the order of product reviews (some personas deem to like certain kinds of reviews, so we’ll put those first). So there’s automation and there’s automation. The latter being smart, autonomous, constantly learning, adapting everything we see. Like really smart.
Still, automation is the easy part in many ways. It will continue to create value beyond our current understanding but much more interestingly, we’ll see a surge of adoption in machine learning and AI being used for augmenting human capabilities. The concept itself is quite old, going back to Engelbart in 1962 who stated that such augmentation strives for “increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems” [9]. This means a lot for every conceivable part of human capability. Decision making, creativity, health, senses, imagination, emotions and so on. Both automation and augmentation have a significant part to play as new players and domains follow up on AI, but augmentation just holds a more versatile promise to us as a species.
The holy grail of augmentation can be easily seen as the pursuit of creativity but there are many other areas of interest as well. Strategic decision making, such as choosing where to build new skyscrapers, where to build new infrastructure (bridges, roads, facilities), what type of aircraft should we buy to maximize profitability and growth and what routes should we fly —counting in sustainability. These questions are still largely thought out with excel sheets, BI-tools and GIS-systems, and maybe some legacy statistics software (SAS, SPSS) with some custom analysis. While that may be sufficient for some industries, many of these problems have so many attributes that it’s impossible for us as humans to make optimal decisions — hence welcoming optimization and machine learning to help as augmenting features of decision making. And despite the fact that it’s still quite early to tell, deep learning may well be of use here (as it’s not limited to historical data but can play against itself [14]).
Machine learning in 2020 will be about augmenting our capabilities as humans, as workers, as decision makers, as engineers, as designers. It will also be about the next wave of automation, where we’ll see completely autonomous and magic-like optimization of logistics, processes, retail-ecosystems and even robotics & space. Pricing will see whole new dynamics and we will finally reach out to strategic decision making too. At the same time batch-style customer churn-prediction, recommendation engines, chatbots, marketing AIs, robotic process automation, fraud detection and up & cross-sell scoring (~tactical decision making), to name a few, will become more and more standard operations across industries and thus make driving competitive advantage much more trickier in their respective areas. It’s likely that this part of machine learning will be done mostly by software engineers, not data scientists.
To gain more from automation, machine learning engineers will look to reinforcement learning and online learning as well as refresh on operations research and mathematics for optimization tools and techniques. NLP and NLU will continue to flourish as speech recognition systems and smart assistants spread to wider use. And finally, we’ll start to create completely new things — be that sliced bread or sneakers — through the use of creative AI. It’s going to be fun.
Regards, the Fourkind machine learning & AI team*
Max Pagels, Maria Pusa, Tomas Heiskanen, Tony Hämmäinen, Jan Kokko, Jan Hiekkaranta, Jarno Kartela, Elli Taimela, Lilli Nevanlinna, Henri Poikonen
[2] https://gizmodo.com/why-everyone-is-hating-on-watson-including-the-people-w-1797510888
[3] https://venturebeat.com/2017/06/27/your-chatbot-lacks-empathy-and-thats-a-problem/
[4] https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/
[5] https://www.ft.com/content/27c8aea0-06a9-11e7-97d1-5e720a26771b
[7] https://hbr.org/2018/01/artificial-intelligence-for-the-real-world
[8] https://www.ft.com/content/eabf70e8-6318-11e8-90c2-9563a0613e56
[9] http://www.dougengelbart.org/pubs/augment-3906.html
[10] https://www.gsb.stanford.edu/sites/gsb/files/mkt_10_17_misra.pdf
[11] https://medium.com/value-stream-design/online-machine-learning-515556ff72c5
[12] https://web.stanford.edu/class/psych209/Readings/SuttonBartoIPRLBook2ndEd.pdf
[13] http://www0.cs.ucl.ac.uk/staff/d.silver/web/Teaching_files/intro_RL.pdf
[14] https://deepmind.com/blog/alphago-zero-learning-scratch/
*We’re growing quite quickly, so this will be old news in a short time. Follow Fourkind to keep track.
