Read on Twitter
Ai memorizes patterns. It doesn't generalize well.
To generalize better, blend the results of a few Ai's. Google blends 8 deep learning results.
One person is very wrong about the number of marbles in the jar. But the average of a bunch of people's guesses is pretty damn close.
To generalize better, blend the results of a few Ai's. Google blends 8 deep learning results.
One person is very wrong about the number of marbles in the jar. But the average of a bunch of people's guesses is pretty damn close.
This works similar to the Central Limit Theorem (if you ask me what this is and don't google it, I'll steal your house).
Basically if the Ai's make DIFFERENT TYPES of mistakes, then they tend to average out to the truth.
I think Ai engineers should blend hundreds of DL results.
Basically if the Ai's make DIFFERENT TYPES of mistakes, then they tend to average out to the truth.
I think Ai engineers should blend hundreds of DL results.
Ai's should be input into a genetic algorithm with survival of the fittest.
Basically, 2 Ai's compete on maybe 100 new datapoints. The one that gets more wrong is deleted.
Repeat the natural selection continuously.
When the user wants a prediction, average the surviving ones.
Basically, 2 Ai's compete on maybe 100 new datapoints. The one that gets more wrong is deleted.
Repeat the natural selection continuously.
When the user wants a prediction, average the surviving ones.
In the field, each Ai would be trained on a small subset of data.
Each one is called a "weak learner".
Algorithms such as Adaboost intelligently combine weak learners.
A system can be continuously retraining in the background. Never stopping. With random subsets of data pool.
Each one is called a "weak learner".
Algorithms such as Adaboost intelligently combine weak learners.
A system can be continuously retraining in the background. Never stopping. With random subsets of data pool.
This can be turned into a democracy or an oligarchy of the best Ai's, all averaging out to a more ROBUST final prediction.
Most Ai engineers try to tinker with making single great Ai. An Ai dictatorship.
I think this is the wrong approach.
Scale horizontally, not vertically.
Most Ai engineers try to tinker with making single great Ai. An Ai dictatorship.
I think this is the wrong approach.
Scale horizontally, not vertically.
And by having the genetic algorithm continuously competing in the background, it keeps the Ai fresh with new data from the environment. They keep getting stronger. Natural Ai selection.
And this can be built to fit to available memory/processing, making it somewhat future-proof.
And this can be built to fit to available memory/processing, making it somewhat future-proof.
With each weak learner just using a subset of data, an individual Ai will likely be very wrong. But on average the Ai hive mind will be both robust and accurate.
You can take the regression layer right before the final classification layer and just blend that one.
You can take the regression layer right before the final classification layer and just blend that one.
Overall, Ai engineers should be focusing on robustness, on a low Sharpe Ratio (do NOT ask me what this is or I'll steal your first-born child) of the full system instead of squeezing out the best single Ai.
And focusing on future-proofing systems with continuous training.
Fin.
And focusing on future-proofing systems with continuous training.
Fin.
