Artem Kaznatcheev, Konrad Paul Kording, Nothing makes sense in deep learning, except in the light of evolution, arXiv:2205.10320
This is an interesting and imaginative article. I am particularly pleased that they regard the cultural object, in this case DL models, as the beneficiary of cultural evolution and not the human creators of the models. I believe this is the correct approach, and it seems to be what Dawkins had in mind when he first advanced the idea of memes in The Selfish Gene (1976), though memetics has not developed well as an intellectual discipline.[1] I have included the article's abstract at the end of these notes.
I want to take up two issues:
- randomness, and
- identifying roles in the evolutionary process.
Randomness
From the paper, p. 3:
As we consider the “arrival of the fittest”, the history of deep learning might seem quite different from biological evolution in one particular way: new mutations in biology are random but new ideas in deep learning do not seem to be random.
What matters, though, are what ideas become embedded in and survive in practice over the long term, for whatever value of “long” is appropriate, which is not at all obvious.
Consider a case I know better, that of music. To a first approximation no one releases a song to the marketplace with the expectation that it will fail to find an audience. Rather, they intend to reach an audience and craft the song with that intention. Audiences do not, however, care about the artist’s intentions, not the intentions of their financial backers. They care only about the music they hear. Whether or not a song will be liked, much less whether or not it will become a hit, cannot be predicted.
A similar case exists with movies. The business is notoriously fickle, but producers do everything in their power to release films that will return a profit. This has been studied by Arthur De Vany in Hollywood Economics (2004).[2] By the time a film is released we know the producer, director, screen writer, principal actors, and their records. None of those things, taken individually or collectively, allow us to predict how a film will perform at the box office. De Vany shows that at about three or four weeks into circulation, the trajectory of movie dynamics (that is, people coming to theaters to watch a movie) hits a bifurcation. Most movies enter a trajectory that leads to diminishing attendance and no profits. A few enter a trajectory that leads to continuing attendance and, eventually, a profit. Among these, a very few become block busters. We cannot predict the trajectory of an individual movie in advance.
Few objects are more deliberately crafted that movies. All the deliberation is insufficient to predict audience response. Films are too complex to allow that.
Thus I am, in principle, skeptical of Kaznatcheev’s and Kording’s claim that the evolution of DL models is not random in the way that biological evolution is. Yes, developers act in a deliberate and systematic way, but it is not at all clear to me how closely coupled those intentions are to the overall development of the field. What if, for example, the critics of deep learning, such as Gary Marcus, are proven correct at some time in the future? What happens to these models then? Do they disappear from use entirely, indicating evolutionary failure? Or perhaps they continue, but in the context of a more elaborate and sophisticated system – perhaps analogous to the evolution of eukaryotic cells from the symbiosis of simpler types of cells. That of course counts as evolutionary success.
More closely at the home, the performance of DL models seems somewhat unpredictable. For example, it is my impression that the performance of GPT-3 surprised everyone, including the people who created it. Other models have had unexpected outcomes as well. I know nothing about the expectations DL researchers may have about how traits included in a new architecture are going to affect performance metrics. But I would be surprised if very precise prediction is possible.
I don’t regard these considerations as definitive. But I do think they are reason to be very careful about claims made on the basis of developer intentions. Further investigation is needed.
Roles in the evolutionary process
It is my understanding that biological evolution involves a number of roles:
- the environment in which an organism must live and survive,
- the phenotypic traits the organism presents to that environment,
- the genetic elements that pass from one generation to the next, and
- the developmental process that leads from genetic elements to mature phenotypes.
How do Kaznatcheev and Kording assign aspects of deep learning development to parallel roles?
They explicitly assert, p. 6:
In computer science, we will consider a general specification of a model or algorithm as the scientist-facing description – usually as pseudocode or text. And we will use ‘development’ to mean every process downstream of the general specification. For a clear example – all processes during compilation or runtime would be under ‘development’. We might even consider as ‘development’ the human process of transforming pseudocode in a paper into a programming language code.
That roughly speaking is the development process.
Am I to take it then that the genetic elements are to be found in “the scientist-facing description – usually as pseudocode or text”? I don’t know. But let me be clear, I am asking out of open curiosity, not out of a desire to find fault. They know the development process far better than I do. Given what they’ve said, that scientist-facing description seems to be analogous to an organism’s genome.
Correlatively, the mature phenotype would be the code that executes the learning process. Do we think of the data on which the process is executed as part of the phenotype as well? If so, interesting, very interesting.
That leaves us with the environment in which the DL model must function. I take that to be both the range of specific metrics to which the model is subjected and the range of open-ended commentary directed toward it. Here’s a question: How is performance on specific metrics traced back to specific ‘phenotypic’ traits?
Consider a different and, it seems to me, more tractable example: automobiles. One common measure of performance is acceleration, say, from zero to 60mph. We’ve got a particular car and we want to improve its acceleration. What do we do? There is of course an enormous body of information, wisdom, and lore on this kind of thing. There are things we can do to specific automobiles once they’ve been manufactured, but there are also things we can do to redesign the car.
Where do we focus our attention? On the cylinder bore and stroke? The electrical system? The transmission. Axel, wheels, and tires? Lighter, but more expensive, materials? Perhaps we make the shape more aerodynamic? Why not all of the above.
So, we do all of the above and our new car now does 0-60 in four seconds, while the old one did it in 5.5. How do we attribute the improvement over all the differences between the new and the old models? If we can’t do that with a fair amount of accuracy, then how are we to know which design changes were important and which we not? If we don’t know that, then how do we determine which traits to keep in play in further development?
What does this imply about the role of deliberate designer intention in the evolutionary process of complex technical artifacts?
* * * * *
Finally, I note that Kaznatcheev and Kording development a major section of the article to considerations derived from EvoDevo. I have been aware of EvoDevo for years, but no little about it. So this (kind of) material is new to me.
I like what they’re doing with it. They make the point that organisms, and complex technical assemblages, have an internal coherence and dynamic the constrains how they can be modified successfully. Changes must be consistent with existing structures and mechanisms. That does enforce order on the evolutionary process.
Abstract of the Article
Deep Learning (DL) is a surprisingly successful branch of machine learning. The success of DL is usually explained by focusing analysis on a particular recent algorithm and its traits. Instead, we propose that an explanation of the success of DL must look at the population of all algorithms in the field and how they have evolved over time. We argue that cultural evolution is a useful framework to explain the success of DL. In analogy to biology, we use ‘development’ to mean the process converting the pseudocode or text description of an algorithm into a fully trained model. This includes writing the programming code, compiling and running the program, and training the model. If all parts of the process don't align well then the resultant model will be useless (if the code runs at all!). This is a constraint. A core component of evolutionary developmental biology is the concept of deconstraints – these are modification to the developmental process that avoid complete failure by automatically accommodating changes in other components. We suggest that many important innovations in DL, from neural networks themselves to hyperparameter optimization and AutoGrad, can be seen as developmental deconstraints. These deconstraints can be very helpful to both the particular algorithm in how it handles challenges in implementation and the overall field of DL in how easy it is for new ideas to be generated. We highlight how our perspective can both advance DL and lead to new insights for evolutionary biology.
References
[1] I have prepared a brief sketch laying out various approaches that are being taken to study cultural evolution: A quick guide to cultural evolution for humanists, Working Paper, November 14, 2019, 4 pp., https://www.academia.edu/40930224/A_quick_guide_to_cultural_evolution_for_humanists
[2] Arthur De Vany, Hollywood Economics: How Extreme Uncertainty Shapes the Film Industry, Routledge, 2004. I’ve written a brief review: Chaos in the Movie Biz: A Review of Hollywood Economics, New Savanna, December 9, 2018, https://new-savanna.blogspot.com/2012/05/chaos-in-movie-biz-review-of-hollywood.html
