Such a great post Edem! And many thanks for the mention 🙏. I would also add "training/development bias" to the list.
Sometimes even if the human bias and data bias is accounted for, engineers can bias the training process e.g. by splitting the data into train/test in ways which unknowingly induce bias. Also untested pre-processing pipelines might introduce unwanted bias in performance too!
You’re right man. I’ll add that in. I’m also looking to add to in examples of what biased data looks like and examples of real life results of biased data.
I like this form for Aurum Finds - it seems appropriate, and like it will reward your readers once they finish your piece. Good call, and thank you for the shout!
"Data is essentially gathered from human interaction so doesn’t it make sense that the biases which exist in the real world would be found in datasets?" one minor point of order: data doesn't have to involve human interactions at all, but of course we are always the ones deciding what data to consider in the first place. And, ultimately, a human will evaluate said data, so you're right in asserting this, but some readers might be confused and think data only measures what humans do or how they interact. Very minor point, but my OCD forced me to weigh in!
There's at least three layers of Bias in AI/ML.
1. Structural (what we might call ethical bias)
2. As you mentioned, Data Bias
3. The Algorithm itself is mathematical bias.
In fact, you can't ever eliminate it....you can only bias bias to bias toward the bias you want.
https://www.polymathicbeing.com/p/eliminating-bias-in-aiml
Interesting that we touched upon many of the same issues on the same day!
I read your piece Rudy, you hit the nail right on the head with bias as a huge delimiter or equal AI progress.
Such a great post Edem! And many thanks for the mention 🙏. I would also add "training/development bias" to the list.
Sometimes even if the human bias and data bias is accounted for, engineers can bias the training process e.g. by splitting the data into train/test in ways which unknowingly induce bias. Also untested pre-processing pipelines might introduce unwanted bias in performance too!
You’re right man. I’ll add that in. I’m also looking to add to in examples of what biased data looks like and examples of real life results of biased data.
Come to think of it though. I think serves as sub-category under human bias. What do you think?
You're right - it could be a sub category - depending on how smart the human is ;)
😂
I like this form for Aurum Finds - it seems appropriate, and like it will reward your readers once they finish your piece. Good call, and thank you for the shout!
"Data is essentially gathered from human interaction so doesn’t it make sense that the biases which exist in the real world would be found in datasets?" one minor point of order: data doesn't have to involve human interactions at all, but of course we are always the ones deciding what data to consider in the first place. And, ultimately, a human will evaluate said data, so you're right in asserting this, but some readers might be confused and think data only measures what humans do or how they interact. Very minor point, but my OCD forced me to weigh in!