This post is an adaptation of a talk I gave the Global Airline Training & Simulation virtual conference (Global ATS-V). I provided some background and history of the field of artificial intelligence, followed by examples of how AI is currently being applied to aviation in exciting ways.
It’s quickly becoming a cliché to say that AI is disrupting every industry on the planet. It’s been almost a year now since Marc Andreessen’s famous essay in the Wall Street Journal about why software is eating the world.
Software is eating the world. — Marc Andreessen
That was true in 2011, and it’s still true today. Everywhere we look, software is being embedded into the things that we use every day. Even toaster ovens are being outfitted with touchscreen displays.
Given the way that people breathlessly describe their new products using the words “artificial intelligence,” “machine learning,” and “predictive analytics” interchangeably (I’m somewhat guilty of this myself), it seems to me that people don’t necessarily understand the distinctions between these terms and may be using them more for their hype value than anything else.
Beyond the hype, though, is real technology that has been gaining traction for a long time and is poised to “eat the world.”
Applied AI, in most industry and product categories, mostly comes down to doing predictions. Given some input data, predict the desired output. That’s it.
Some examples
- Take X audio, and predict Y words as text (speech-to-text engine)
- Take X text, and predict Y synthesized audio (text-to-speech engine)
- Take X text, and predict the next word in the sentence (GPT-3)
- Take X image, and predict Y contents (computer vision, image segmentation)
- Take X time series, and predict Y future values (forecasting)
- Take X user reviews (👍,👎), and predict Y preferred song/show/product (recommendation engine)
- Take X user query, and predict Y table of results (search engine)
Virtually all machine learning tasks are prediction tasks. And the winner is the business that makes the best predictions.
Most businesses do not care about artificial general intelligence or whether consciousness can be instantiated in a machine. They care about solving real business problems.
I’m currently the CTO at an aerospace software company called Paladin AI, committed to melding human intelligence with machine intelligence in the area of pilot training.
With fully autonomous cars expected within years, it’s tempting to assume fully autonomous passenger aircraft are next. That’s mistaken.
Quite apart from the fact that Level 5 autonomy in cars is turning out to be much harder than we thought, given the current state of deep learning research, human brains seem to be uniquely qualified at dealing with the unexpected.
Dealing with the unexpected is at the core of much pilot training today: threat and error management, upset recovery and prevention, situational awareness, problem solving and decision making.
Instead of full autonomy, many automakers are opting for a stepped approach with human-augmenting automation systems that make driving safer, while still leaving the driver the responsibility of navigating through complex environments (any major city) and emergent situations (jaywalkers, unexpected obstacles, inclement weather).
A similar approach has been used in aviation, where automation has steadily overtaken many aircraft systems and processes.
Dealing with the unexpected is at the core of much pilot training today.
In the best case, this increased automation reduces the pilots’ cognitive burden of managing an aircraft. However, in the worst cases, pilots become too dependent on automation and lose their manual flying skills. This dependency has contributed to a number of high-profile accidents over the years.
Machine learning thrives wherever there is lots of data. And aviation is one of the most data-intensive industries on the planet. “Black box” data recorders were an early innovation that tracked a handful of parameters that were essential in accident investigations. The more modern Quick Access Recorder (QAR) tracks 2000 or more different parameters, and at much higher sampling rates. A multitude of other sensors are collecting data for predictive maintenance and future performance improvements.
Here is another example:
Most airlines employ highly complex machine learning models that predict the price a given consumer is willing to pay for their flight, allowing an airline to maximize the revenue per seat-km. This is why ticket prices are constantly changing right up until just before boarding.
The Covid-19 pandemic has thrown off these models. Since they’re based on historic data, and a disruption of this magnitude has never occurred before, airlines are having a difficult time setting and adjusting prices for flights.
The holy grail of “evidence-based training” would be if you could take all of your airline operations data, and have some very intelligent software design a perfectly tailored recurrent training curriculum for you. This program would perfectly target any weaknesses and not waste any time on redundant training.
What if such an AI could personalize down to the individual level, and start measuring competency directly?
For most of our industry’s history, pilot training has been task-based training, which has some obvious problems:
- The number and variety of training tasks keeps increasing in response to new threats, and…
- Rote execution of a task from memory does not, by itself, guarantee competency.
So the concern is that we’re making training more complex and more expensive, but not necessarily safer.
For these reasons, ICAO and others have been advocating for a shift to competency-based training. Now regulators are starting to push in that direction as well.
Implementing competency-based training, if done manually, comes with many upfront costs and increased instructor burden. Hence the need for automation.
Programs like the Multi-crew Pilot License (MPL) are already designed around competency-based training. These programs can help address the pilot shortage. But competency-based training is hard to implement. There’s an upfront cost to overhauling curricula. Instructors need to be retrained, because now they’re tasked with assessing pilots across even more dimensions.
So instead of adding more instructor overhead, we’re using AI to look for the competency indicators in the data, and do this automatically.