I suppose this all depends on how you are framing/thinking about the question.
There is a certain amount of bias in being designed/architected/engineered/constructed by humans - the artifact is 'in our image', as it were.
Are you talking about feeding a machine learning system data/information to be turned into information/knowledge by the system that is created by humans? That has our fingers all in it as well.
Cognitive bias? Well, are we talking about what is an 'acceptable' result to a human, or the underlying process? Take Google's machine vision system that sees the world as dogs - the cognition certainly is different, and the results of value to humans for an illustration/visualization of the different approach as well as a certain aesthetic I suppose. The bias is in the entire point of the question and the specialization of the system to provide an answer. But as to the details of the analysis and results? I don't think that has bias in the sense you may be thinking of.
Can you be more specific about what is meant by bias?
Well not when our biases are based in some sort of reality.
Take the classic example of racial bias in crime algorithms or something like that. Let's say you have a neural network that you feed details of someone and it generates a weighting to use to determine whether to 'randomly' screen them at a train station or something.
Now sure, statistically, it might produce better results to screen more people of skin tone X. But how much is that actually true and how much does it appear to be true because the existing data you have is coloured by the bias of previous actors? If the police routinely 'randomly' screen people of a particular skin colour, then they're going to be found to be worthy of being screened more often, simply because you're screening more of them. So the algorithm goes 'hey screen that guy, he's skin colour X' and the cycle continues.
But at the same time, if that's the best data you have then it's the best data you have.
I'd argue that human bias is actually a source of truth. "higher" levels of truth are often codified in a way very similar to logical constructs. Human knowledge is biased but is flexible and adaptable. Codified law is fairer and has less bias but can only deal with exact situations.
I wonder if black box AI internally has an isomorphism to what is basically a gigantic codification of logic and the training process just discovers then or if it's fundamentally something very different.
3 comments
[ 0.19 ms ] story [ 19.0 ms ] threadThere is a certain amount of bias in being designed/architected/engineered/constructed by humans - the artifact is 'in our image', as it were.
Are you talking about feeding a machine learning system data/information to be turned into information/knowledge by the system that is created by humans? That has our fingers all in it as well.
Cognitive bias? Well, are we talking about what is an 'acceptable' result to a human, or the underlying process? Take Google's machine vision system that sees the world as dogs - the cognition certainly is different, and the results of value to humans for an illustration/visualization of the different approach as well as a certain aesthetic I suppose. The bias is in the entire point of the question and the specialization of the system to provide an answer. But as to the details of the analysis and results? I don't think that has bias in the sense you may be thinking of.
Can you be more specific about what is meant by bias?
Take the classic example of racial bias in crime algorithms or something like that. Let's say you have a neural network that you feed details of someone and it generates a weighting to use to determine whether to 'randomly' screen them at a train station or something.
Now sure, statistically, it might produce better results to screen more people of skin tone X. But how much is that actually true and how much does it appear to be true because the existing data you have is coloured by the bias of previous actors? If the police routinely 'randomly' screen people of a particular skin colour, then they're going to be found to be worthy of being screened more often, simply because you're screening more of them. So the algorithm goes 'hey screen that guy, he's skin colour X' and the cycle continues.
But at the same time, if that's the best data you have then it's the best data you have.
I wonder if black box AI internally has an isomorphism to what is basically a gigantic codification of logic and the training process just discovers then or if it's fundamentally something very different.