# Is AI research culture predisposed to adversarialism while even the mathematics is more friendly?

Language is how we communicate to one other in the lab, at home, and in educational settings.

Word selection matters. Sometimes we don't select well. If we are writers, we may edit a word later or the editor may, but once picked and published, the something we have given a particular name may stick, and we can't change it easily. An example is the GAN (generative adversarial networks).

The generative artificial network ideas predated Ian J. Goodfellow et. al.'s 2014 paper introducing GANs. Some of those ideas are listed in the paper's bibliography. In the paper, they state,

Generative adversarial networks has been sometimes confused with the related concept of adversarial examples.

Why, after admitting that it is a cause for confusion, did they pick the term adversarial for the second adjective in the name of their design?

They also write,

In the proposed adversarial nets framework, the generative model is pitted against an adversary: a discriminative model ...

What does it mean to be pitted against an adversary? The straightforward answer is that we place two entities $$A$$ and $$B$$ such that the pursuit of the objectives of $$A$$ by $$A$$ hurts or destroys $$B$$ and vice versa. Even if not a game of annihilation or at least a zero sum game, there is clearly no win-win scenario, where $$A$$ and $$B$$ both get much of what they want.

Mutual benefit and the bargaining or negotiation process, in game theory, is called a Nash Equilibrium1. It has been refined2 since Nash, but the basic idea of positive sum games as being the basis of bargaining in purchases and treaty negotiations remains a primary political and economic concept. When trade occurs or agreements are reached, such is done so that both parties benefit, even though there is a push-pull process in reaching the equilibrium.

In the GAN design, each network's objective is represented by its loss function, and nothing else. The networks cannot die, kill, dismember, hate, or even get angry at one another. There is no adrenaline or testosterone. They are not proud of themselves or their country. Isn't the term blatantly anthropomorphic?

The mathematics reveals, as so often it does, the truth. The paper further explains,

If $$G$$ and $$D$$ have enough capacity, and at each step of Algorithm 1, the discriminator is allowed to reach its optimum given $$G$$, and p_g is updated so as to improve the criterion

$$\mathbb{E}_{x ∼ p_{data}} \Big[ \log D_G^∗ (x) \Big] + \mathbb{E}_{x ∼ p_g} \Big[ \log (1 − D_G^∗ (x)) \Big] \; \text{,}$$

then p_g converges to p_{data}.

Note that, in this convergence, neither $$G$$ nor $$D$$ are hurt or destroyed as in Chess, where pieces are lost and concession is symbolized by tipping one's own King, representing that the King has fallen, meaning killed. Both networks are in Nash Equilibrium. Some might call the interaction between $$G$$ and $$D$$ negotiation and the result harmony.

In straight mathematical and unbiased terms, GANs are successful because of convergence collaboration. If each network concedes a portion of its loss optimality so that an equilibrium can be reached, then we have a Nash Equilibrium, which is not zero sum. It is win-win. If both converge fully, then the win-win is even stronger.

The collaboration mushrooms to a win-win-win-win scenario once the AI engineer and the project stakeholder are added in, since the consequence of the GAN equilibrium is the generated output the stakeholder wants and for which the AI engineer receives monetary and professional appreciation.

What causes these militant terms and negatively charged, anthropomorphic ways of looking at benign processes?

For the case of the 2014 paper, is the answer related to the fact that, since the cyber conflicts between Russia and Estonia, the term adversarial networks was common in papers arising from or vying for military funding?3

Could that word association been subliminally introduced from Ian Goodfellow's subconscious? This is not to fault him or his colleagues but to get to the bottom of the conflict between the mathematics which proves non-adversarialism in a paper that names the algorithm arising from the math as adversarial.

In the case of Chess, the Huns attacked the Gupta Empire (now northern India) in the fifth century and Chaturanga, the forerunner of Chess, was invented in that same region in the sixth century. That would be immaterial if Chess wasn't the thing most associated with AI until GANs started generating images.

Stepping back from anthropomorphism and mathematics, from a sociology perspective, enemies are just regular people that resent others with greater abilities, social grace, food supply, and other goods. Some believe that, from God's perspective, enemies are people sent so we can learn how to love and show compassion unconditionally. If we apply this particular belief to populations, the purpose of adversarialism is solely and precisely to lay it down, along with its weapons.

Life for most on planet earth starts with some difficulty and increases gradually, with perhaps a decade or two of reprieve, but then increases more quickly until it ends. One interesting and multidimensional question for each individual is how we act and speak to one another under those universal conditions. And that intelligence is ancient and far from artificial. Which brings us back to the primary question.

Is AI research culture predisposed to adversarialism while even the mathematics is more friendly?

The selection of words is how we communicate, and the choices we make perpetuate things that are good for AI and the world and things that may not be so good.

Footnotes

[1] Even Nash's paper, Non-cooperative Games, Annals of Mathematics, 1951, describes an equilibrium with mutual advantage but uses the term non-cooperative in the paper title instead of Game Advantage Equilibria.

[2] K Binmore, A Rubinstein, and A Wolinsky, in their The Nash bargaining solution in economic modelling, RAND Journal of Economics, 1986, "The players are induced to reach an agreement by their impatience for the [mutually beneficial] outcomes."

[3] With regard to the cyber conflicts between Russia and Estonia, the BBC news stated, "But since [the 2007 cyber Estinian incident,] cyber warfare has been used all over the world, including in Russia's war with Georgia in 2008, and in Ukraine. "Cyber has become a really serious tool in disrupting society for military purposes," says Tanel Sepp, ... a cyber security official at Estonia's Ministry of Defense."

Short Answer: Although I agree that words can mean different things, the names of new ideas in AI are not predisposed to be adversarial but are intended to be relatable simplifications of an idea.

Word selection matters. Sometimes we don't select well. If we are writers, we may edit a word later or the editor may, but once picked and published, the something we have given a particular name may stick, and we can't change it easily.

Like any writer, AI researchers try to find names that give an intuitive meaning towards things. A related example is Collaborative Learning. In this approach, multiple classifiers are trained on the same data and the authors highlight 2 advantages:

First, the consensus of multiple views from different classifier heads on the same example provides supplementary information as well as regularization to each classifier, thereby improving generalization.

Second, intermediate-level representation (ILR) sharing with backpropagation rescaling aggregates the gradient flows from all heads, which not only reduces training computational complexity, but also facilitates supervision to the shared layers.