Legion, introduced in the science fiction video game Mass Effect 2, is the only killer robot I know that makes all its decisions by consensus. What does that mean? Well, let me explain that what appears to be a single being is in fact a mobile platform run by no fewer than 1,183 individual sentient programs. As digital life forms coexisting in the same digital interface, these programs communicate with each other at light speed and agree on each decision Legion itself makes.
Well, that’s a neat, whimsical idea, eh? Not so fast. Many models working together to collectively make one decision is a real thing. It’s called an ensemble model. Ensemble models don’t tend to communicate with each other to achieve consensus, but they do each cast a vote, with the result either going to the most popular decision or the decision of another model that takes the various lower models’ confidence levels into account.
Does that sound goofy? Maybe. Inefficient? Well, let’s say you wouldn’t want to have 1,183 2000-layer neural networks voting on whether they’re looking at a dog, a pig, or a loaf of bread. But with relatively simple models that run fast individually, you can find big boosts in accuracy by running them in groups.
But wait, wouldn’t a thousand of the same model trained on the same data all give the same result? Not exactly, since there are random elements in training, but pretty close, yeah. So how does it get more accurate? Well, in training, we don’t give every model the same data. Each of the hundred or thousand models gets a random subset of the data. At runtime, when we want the best possible answer, they all see the same data, but instead of a thousand identical clones observing the input, you have a thousand models each with its own unique experiences and perspective.
Why would having more models and giving each of them less information lead to a better result? Well, for the answer to that, let me take you away from the distant future where a robot comes to consensus with itself to shoot you in the face. In fact, let me take you all the way back to 1907.
Sir Francis Galton is tallying guesses on the weight of an ox submitted at a fair. He assumes that since these are the unwashed masses, their answer will be way off, and, individually, he’s right. Each guess ranges dramatically. When he calculates the median, however, he finds a different story. The median prediction, taken from this gaggle of uneducated folk, is less than 1% removed from the real weight of the animal. It’s almost exactly right. Why is that?
Each individual guessing has their own biases and distortions that take them away from the correct answer. But then when we combine them together using a median or a mean, they tend to cancel each other out. The more individuals involved, the closer we can filter out the noise and get to the underlying signal.
Pooling votes from a crowd only works to get to the right answer if the distortions of each individual are independent. That is, there’s not a common factor confusing people or models in the same direction, which you can encourage by making sure to maximize the diversity.
That’s right, maximizing diversity in the voting body is something we actually worry about in machine learning. Not because we’re bleeding hearts that want to make sure every model feels good, but because it actually gets a better result. Some food for thought.
Origin: Mass Effect 2 (2010)
Likely Architecture: Transformers for speech and language, Convolutional networks for vision. Each input gets passed to a voting collective of 1,183 reinforcement learning networks that deliberate and come to consensus.
Possible Training Domains: Years and years of experience coming to consensus and making decisions in the “Legion” body and more in other previous bodies.
I take requests. If you have a fictional AI and wonder how it could work, or any other topic you’d like to see me cover, mention it in the comments or on my Facebook page.