Process and Results Orientation
Does gravity move water or does the water flow to the ocean?
In AI right now there is a problem in that most AI companies are very much focused on results. Mostly they care about doing well on the MMLU or some other dataset that is designed to test the limits of the model. Most algorithms work this way as well, they focus only on optimizing the loss function or improving accuracy to 100%, this is called results orientation. Results orientation is teleological, it’s what the ancient Greeks called Telos or purpose. It’s focused on finding the best, fastest, quickest, way to a goal. This is admirable at times, however, there is something to be said for another way of looking at things.
When water moves downstream is it moving because of gravity or is it moving downhill toward the open ocean? This is the difference between process and results orientation. Results-oriented are focused on the end result, the metric, and the teleological reason for the entire thing. This is what is called supervised learning in artificial intelligence. Unsupervised learning is what we might call process-oriented. It’s not focused on some end goal but rather focused on the process to get there. The modern world is heavily results-oriented, performance, efficiency, end goals, competitions, etc. This is valuable in some sense because it allows you to perform well, however when trying to creatively explore a problem space or understand the problem it doesn’t do the right thing at all.
For example, a process-oriented algorithm like t-SNE dimensionality reduction doesn’t help us improve on any metric but it does help us understand the problem space better. Art classes are process-oriented, unlike most math classes which are focused on test results. Art focuses on creative exploration rather than getting the right answer, maybe math should go back to being process-oriented. A process-oriented dating algorithm would try to find matches by exploring the search space of matches and determining which they do and do not like rather than trying to optimize the perfect match. The many world interpretation of quantum mechanics is process-oriented, every possible universe is tried like a simulation.
There are downsides to process orientation, rarely do you get the best, often you get a subpar solution but you try a lot more solutions than you would with a results-orientation and you explore the problem space a lot more. Disney movies are results-oriented, they want to get you to some happily ever after ending as quickly as possible and packaged up with merchandise. Indie flicks and obscure films tend to be process-oriented, exploring the spaces of the unknown. Process orientation is comfortable with the unknown, they don’t fear it. Star Trek is process-oriented and Marvel is results-oriented. The superhero saves the day and everything is gonna be ok, nothing is left open-ended. Compare that to cliffhangers like the first part of The Best of Both Worlds which leaves us on an unfulfilling note.
Process orientation is the journey as the main goal and the goal as the sideshow. When looking for a place to eat, process-oriented might mull over deciding for long periods of time, and results-oriented might want to find a place quickly that has the highest rating. There are upsides to results orientation, efficiency, satisfaction, and finding the best option, however, there are significant downsides. One major downside is a lack of diversity, everything looks the same, and everything is optimized, this is the downside of communism, corporate capitalism, and fascism, everything is gray. Another problem is a lack of exploration of the problem and solution space. We don’t visit places that might prove better in the long run because we focus on the quickest efficient goal. Large swaths of the solution space are unexplored and never get seen. In machine learning algorithms like stochastic gradient descent were designed to mitigate against this but even the best still fall victim to it.
The best process-oriented learner is nature itself, evolutionary change explores the solution space bit by bit by creatively modifying genes. Likewise, genetic algorithms and evolutionary computing may take a long time but explore a lot of possible solutions. Evolution produces interesting designs like this weird-looking plant.
A results-oriented process could never generate this only a natural artist could. Nature is the best at process orientation, however, it’s not the only process-oriented process out there. Here are some examples of process and results-oriented things.
Results Orientation: Competitive sports, gradient descent, Common Core, finance, Disney, post-it notes, SCRUM boards, uniforms, sitcoms with laugh tracks, GDP, fast cars, Dijkstra’s algorithm, modern air travel, deep neural networks, Michelin restaurants, Overwatch, SAT, drug laws, chess.
Process Orientation: Nature, genetics, space exploration, art, indie films, Avante grade music, science fiction, fantasy, hiking, boat rides to nowhere, leisurely train travel, clustering, PCA, obscure bookstores, flash fiction, quantum mechanics, No Man’s Sky, Minecraft, interdimensional cable, simulations, permaculture.
Notice the process-oriented list isn’t necessarily “the best” but instead is “the most explored.” Are indie films the most visually appealing or do they have the best writing? No not necessarily but they are the most expansive in broad reach covering topics that mainstream films never would. Does taking a train through the European countryside lead to the fastest route? No, but it does lead to more interesting views. I often ask myself why nothing goes the way I wanted it to, maybe I should ask, why we should not explore what is around us. It seems cruel to tell a grieving mother who lost her child to an illness that she’s “exploring the space of possible existences,” why didn’t they live the longest? However, exploration is a part of discovery, even something unexpected that may be more valuable than what we thought we lost.
The modern world is focused heavily on results-orientation, we care about how much money we’re making or whether the unemployment rate is low enough. However, this leads to one drawback, we can’t control everything, and this leads to frustration. Results orientation is a frustrating dilemma when the results don’t live up to what we hope. Nature does not abide by a results orientation, rather it effectively utilizes a process orientation. For example, evolution explores different kinds of lifeforms and physical bodies moving around due to gravity. They are not reaching for a goal, but rather selectively exploring the space by using different natural processes. In hydrodynamics, water flows from a high point to a low point based on the process of gravity itself rather than because of some goal it is necessarily trying to reach near the bottom.
I was thinking about this distinction regarding AI competitions like the ARC prize. I started off trying the genetic algorithm approach and disregarding it because it was taking too long. However, sometimes good things take a long time. The results approach which uses conventional deep learning or DSLs which have done fairly well takes a lot less time but tends to overfit and not generalize well to unique problems not covered by the DSL or model. The process orientation might be a useful approach when thinking of general artificial intelligence. It remains to be seen whether it works, it may not score the highest but it might generalize like genes do. In choosing a process orientation you might discover results you never knew about.
