Recent Anecdotal Evidence Of Practical AI
AI is learning from data. Practical AI is the successful (and unsuccessful) learning from data by enterprises.
“Learning” in the context of AI is computer-based data classification based on examples. Today’s AI (“deep learning”) is a method of statistical analysis that can classify lots and lots of data along many dimensions. In the last decade or so it has been applied successfully to a variety of tasks, with researchers demonstrating its usefulness first in image analysis and later in text analysis (“natural language processing” or NLP).
These breakthroughs have encouraged many researchers to focus their work on AI. The total number of AI publications grew from 162,444 in 2010 to 334,497 in 2021, according to the 2022 AI Index Report. In 2021, a PubMed search of papers with the keyword “deep learning” returned 14,685 citations, up from 107 papers in 2010.
What has been the real-world impact of this considerable research effort? Has it been translated successfully into Practical AI?
A recent comprehensive review of “AI in Health and Medicine” concluded that “Although AI systems have repeatedly been shown to be successful in a wide variety of retrospective medical studies, relatively few AI tools have been translated into medical practice… medical AI remains in an early phase of validation and implementation.”
This conclusion also applies to other activities performed by enterprises in all economic sectors, where the great promise demonstrated in research has not materialized in practice. Today’s vast research effort is by and large focused on developing and perfecting research methods, not on addressing the challenges of implementation and adoption by enterprises.
Regardless of the challenges, however, the promise of learning from data is enticing enterprises to experiment with AI, just as they have been doing—learning from data—since the advent of modern computers more than 70 years ago (and from the development of statistical analysis tools long before that).
As a result, we have some anecdotal evidence regarding the usefulness—and possible misuse—of Practical AI, of real-world AI breakthroughs and breakdowns. In the spirit of today’s AI, we can classify these anecdotes into a number of categories, based on the type of enterprise and real-world environment in which the AI is deployed. It’s a spectrum, from “data-born” (also known as “born digital” or “digital natives”) enterprises to “data-laggards” (also known as “legacy”) enterprises.
The prime example of a data-born enterprise is Google, a company that has contributed greatly not only to AI research but also to implementing AI in the service of its business goals and ambitions. Search has been the core competency of the company (and search-related advertising still accounts for almost all of its revenues) and, around 2015, it started to implement AI (of the text analysis variety) to improve its search results.
Another category includes the companies linking the digital with the physical, whether they are pre-Internet (pre-Web would be a more accurate term) or Internet-born, smartphone-created, enterprises.
Levi Strauss & Co., established in 1853, is using computer vision and AI to enable customers to identify a “look” through images that enables the company to shape and create new products. “Today, Levi’s analyzes more data than ever to transform the product design process,” reports Randy Bean.
Shutterfly, established in 1999, has been collecting lots of data on how its customers design their photo albums—spending hours on sorting photos, cropping them, positioning them, etc. Hilary Schneider, Shutterfly’s CEO and President, tells Calcalist:
“A decade ago we recognized that the album market has insane growth potential, if we just make the product simpler. This is what machine learning staff do here in Haifa [Israel]: they have developed an algorithm that goes over all the albums we have printed before and checks what can be deduced from them… We ‘train’ this algorithm, almost as Google trains its search algorithm, and we do it based on the vast body of knowledge we have accumulated, to predict what will make the album better. The result is that last year we launched a service which designs an album for you in 90 seconds. You just upload your photos to our server, without selecting, and our system alone chooses the photos that have the most chance of being the ones you want in your album.”
This is Practical AI. Measurable improvement directly related to the application of learning from data.
Then there are companies whose core competencies were built on analog technologies. Bose Corporation was established in 1964 when sound was analog as it has always been in nature, acoustic waves moving through air or other media. But in the 1970s and 1980s, digital audio technology, using audio signals that have been encoded in digital (numerical) form, has gradually replaced analog audio.
Bose’s CEO Lila Snyder sees the application of AI to sound as the third stage in the evolution of audio technology. “We think the future of noise cancellation is hearing what you want to hear,” Snyder said at the inaugural event of the Institute for Experiential AI at Northeastern University. “So typically you don’t want to hear everything or nothing. There are things that you want to hear and things you don’t want to hear. The power of AI and data is that we can start to discern the difference between the two.”
Data helps sound engineers get a better sense of the different environments in which people use Bose’s products, triggering noise-cancellation when the user moves from a quiet setting to a noisy one and smooth out volume spikes or dips. Data also helps Bose’s engineers develop technology that amplifies certain sounds while dynamically adjusting noise cancellation and to create immersive audio environments. “We can render and recreate the music in [a specific] environment so you feel like you’re sitting there… but we can’t do that without AI and without data,” Snyder said.
Finally, examples of Practical AI could come from “laggards,” economic sectors such as healthcare that have been traditionally slow to adopt new digital tools or developing and rural areas that have not had the means to do so. For example, Rwanda.
The government of Rwanda, where 83% of the population lives in rural areas, signed six years ago a contract for blood delivery with drone maker Zipline. An analysis of 13,000 drone deliveries between 2017 and 2019 found that half of the orders took 41 minutes or less to deliver by drone. The median time for these deliveries by car would be at least two hours. A reduction in the number of wasted blood donations was also reported.
Practical AI can also fail miserably and measurably. The tax authorities in The Netherlands had used machine learning to identify fraudulent claims of childcare allowances. Based on its training data, the fraud-detection system figured out that immigration and ethnic background and level of income were good predictors of potential culprits. The tax authorities “baselessly ordered thousands of families to pay back their claims, pushing many into onerous debt and destroying lives in the process,” reported IEEE Spectrum.
In this case, the discovery of the failure of Practical AI forced the government to resign.