
Deep learning and competition
Competition and cooperation are typically viewed as being at odds with one another. This is especially the case within centers of learning, such as colleges and universities. Cooperation pays its dividends to the extent that individual interests are not impacted, and as examinations draw closer, a significant majority take on competitive behaviors, trying to better one another as they battle it out for the best grades. All is fair in love and war, as the saying goes.
However, this doesn’t always have to be the norm. Competition and co-operation can co-exist, as we will observe in this blog through the utilization of deep-learning tech, and agent-based modelling.
Agent-based modelling has a close relationship with computational thinking. In the University Scholars Program at the National University of Singapore, the syllabus covers altruistic and cooperative habits with the use of an agent-based model, that mimics the actions of altruistic and selfish agents.
This model is initially random with each portion in a field occupied by an agent. During every time step, an agent will attempt to sow “seeds” in its surrounding patches, which will subsequently grow and evolve into an altruistic or selfish agent. The odds of each fresh agent being altruistic or selfish is dependent on the expenses and advantages of acting in an altruistic manner, which is a factor modelled between 0 and 1. In this simulation, researchers at NUS executed a couple of simulations to depict the spread of altruism and selfishness in this created world.
Simulating scenarios such as these are an amazing asset for comprehending relationships within intricate systems. With them, we can see how changing variables can impact results and outcomes, which can be more informative than studying by hearing or reading. Within this simulation, NUS researchers altered the stress inflicted on agents and discovered the following:
Comparing these couple of simulations helped researchers come to a somewhat curious conclusion. The agents that were deemed selfish were naturally more competitive, but the caveat was that this competitive disposition did not always mean good results or outcomes particularly when stress was a factor. That is, competitiveness did not lend well to stress-inducing conditions. What are the implications for productive competition? When we enter into society as part of productive organizations and communities, does it pay off to be cooperative, or competitive?
Researchers desiring different perspectives, dug deeper, and this led them to a research piece by artificial intelligence research scientist Ian Goodfellow. Goodfellow suggests an innovative approach to developing generative models with mathematical abstractions of the human brain. The numbers behind these Generative Adversarial Networks (GANs) is not so vital, but its worth knowing that it’s a direct approach consisting of an adversarial process.
Where previous progress in deep learning were around discriminative models, GANs were generative, naturally. On one side, discriminative jobs such as image recognition consist of machines attempting to differentiate between images of objects, and tech progressions has rendered such jobs relatively trivial. On the other side, generative jobs are considerably more difficult as they need a machine to be creative, and to generate unique images of fresh objects.
GANs produce fresh images by learning and generalizing aspects from current images, prior to stringing them together much like a painter would. Researchers, by a stroke of luck, came upon a project by users of the open source community where they attempted to develop new Pokemon designs using GANs, and the outcome left researchers pretty spellbound.
While the Pokemon that were created were far from being complete representations, the GANs capacity to learn and generalize the patterns shows a considerable evolution over its predecessors. The relation between the generator and discriminator runs in parallel to other historical rivalries, like that between legendary football players Cristiano Ronaldo and Lionel Messi. They are typically viewed as all time greats, and are always on each others heels to prove who the numero uno is. Through this competition, they indirectly help each other and assist each other in becoming the best possible versions of themselves.
The Generator and Discriminator can mimic this relationship that we just saw between Messi and Ronaldo. They work against one another and try to best each other, upgrading themselves very quickly in the process. Over the course of time, generators produce more lifelike images that are tougher for discriminators to differentiate, who then have no other option but to improve to capture the generators.
While the physical world is more complicated than can be created by an adversarial process, the fact that competition and cooperation can indeed coexist is a powerful realization, a fact that can be utilized by organizational management to optimize, enhance, and increase the cohesion and value of their production teams.
References
https://machinelearnings.co/what-deep-learning-taught-me-about-competition-5157e9eb464