The Fight for Ethics in AI's Lossy Compression Problem(doxa.substack.com)
doxa.substack.com
The Fight for Ethics in AI's Lossy Compression Problem
https://doxa.substack.com/p/googles-colosseum
11 comments
Professor Pedro Domingos, comes to mind. He's been fighting them lately through twitter (@pmddomingos). They've tried to ruin his career by reporting him to his employer, IIRC. But yes, unfortunately, academics rarely get involved for obvious reasons. This reminds me of the Intelligent Design, Creationism, "Teach the controversy" etc debate from the late 2000s, during Richard Dawkins' God Delusion era. It's amazing how everything changed.
The Facct (Fairness, Accountability and Transparency) conference ended recently (https://facctconference.org/). it might be a good start to find different research contributions.
Even if LeCun is totally right, he was a fool for allowing himself to be drawn to the center of such a debate.
Serious academics flee from such topics for good reason. See for example Eugenics.
Serious academics flee from such topics for good reason. See for example Eugenics.
So nobody should ever be the fool? You’re describing a world where people are afraid to speak academically about things that matter.
LeCun is quoted twice in this article and the quotes are logically inconsistent with each other. So he can't be right, unless you just pick one.
And since they're positive claims we can't know either of them are right without evidence, which is not provided.
And since they're positive claims we can't know either of them are right without evidence, which is not provided.
The statements are logically consistent. You're incorrectly assuming he meant "only" in the first statement.
Those two statements are part of the same thread and LeCun makes it more clear in that thread. And the article says, right after the second statement: "If you read that thread, you’ll see that LeCun seems well aware of everything I’ve laid out in this post"
Those two statements are part of the same thread and LeCun makes it more clear in that thread. And the article says, right after the second statement: "If you read that thread, you’ll see that LeCun seems well aware of everything I’ve laid out in this post"
> Fixing racial bias in AI is not just a matter of infusing the training data with more melanin (for example), the AI Ethics crowd argues — the actual models are being developed by white guys, and their insular, white-guy priorities somehow surface as bias in the algorithms that go to work on the training data.
I recently created a natural language generation model(built with LSTM layers mostly) that was trained on east-asian zen books. Do you think that my result could have been better if I would've used an architecture not designed by white germans?
This idea seems is to me like anthropomorphizing model architecture for no real reason.
I do feel like there might be issues of ingrained bias in the model itself when using trained NLP embeddings or even some facial feature recognition algorithms that were tested on racially homogeneous groups.
I recently created a natural language generation model(built with LSTM layers mostly) that was trained on east-asian zen books. Do you think that my result could have been better if I would've used an architecture not designed by white germans?
This idea seems is to me like anthropomorphizing model architecture for no real reason.
I do feel like there might be issues of ingrained bias in the model itself when using trained NLP embeddings or even some facial feature recognition algorithms that were tested on racially homogeneous groups.
> I recently created a natural language generation model(built with LSTM layers mostly) that was trained on east-asian zen books. Do you think that my result could have been better if I would've used an architecture not designed by white germans?
Potentially. Ways this might happen: most SOTA architectures are dependent on larger datasets and abundant compute resources for training (as well as related tasks such as hyperparameter optimization or architecture search). Few architectures are designed or evaluated based on smaller (fixed) corpora sizes and smaller (fixed) training budgets. Even few-shot learning tasks typically still require a huge amount of pre-training on large datasets. So researchers and practitioners constrained by fewer resources and smaller datasets (which may not apply to you specifically) trying to adapt popular architectures to their needs are disadvantaged. Compare the attention being given to energy budgets and similar constraints for inference as opposed to training and the disparity becomes fairly obvious.
So, yes, adapting an architecture created in the first place with the kinds of constraints you are likely to face in mind, by folks that are more likely to be facing similar constraints themselves, may very well lead to you achieving as good or better results with similar or less effort and expense.
Potentially. Ways this might happen: most SOTA architectures are dependent on larger datasets and abundant compute resources for training (as well as related tasks such as hyperparameter optimization or architecture search). Few architectures are designed or evaluated based on smaller (fixed) corpora sizes and smaller (fixed) training budgets. Even few-shot learning tasks typically still require a huge amount of pre-training on large datasets. So researchers and practitioners constrained by fewer resources and smaller datasets (which may not apply to you specifically) trying to adapt popular architectures to their needs are disadvantaged. Compare the attention being given to energy budgets and similar constraints for inference as opposed to training and the disparity becomes fairly obvious.
So, yes, adapting an architecture created in the first place with the kinds of constraints you are likely to face in mind, by folks that are more likely to be facing similar constraints themselves, may very well lead to you achieving as good or better results with similar or less effort and expense.
I think i agree with your general sentiment, but specifically LSTMs were first proposed in a 1997 paper[1], thus my guess is that they didn't have vast resources compared to today. I mean, i trained my model on a GPU for a few days, which im guessing is actually more resources than they had.
>Few architectures are designed or evaluated based on smaller (fixed) corpora sizes and smaller (fixed) training budgets. Even few-shot learning tasks typically still require a huge amount of pre-training on large datasets. So researchers and practitioners constrained by fewer resources and smaller datasets (which may not apply to you specifically) trying to adapt popular architectures to their needs are disadvantaged. Compare the attention being given to energy budgets and similar constraints for inference as opposed to training and the disparity becomes fairly obvious.
that is an interesting point, and i feel like it generalizes to the fact that using more efficient architecture that was perhaps was designed by someone with a lesser training budget. Although I must say that from my limited DL paper reading, efficient small-scale novel architecture doesn't necessarily comes from cash-strapped researchers, as a more efficient(energy and time) would be of huge economic value also to companies like OpenAI, who have spent huge amounts on training GPTs.
[1] - https://www.bioinf.jku.at/publications/older/2604.pdf
>Few architectures are designed or evaluated based on smaller (fixed) corpora sizes and smaller (fixed) training budgets. Even few-shot learning tasks typically still require a huge amount of pre-training on large datasets. So researchers and practitioners constrained by fewer resources and smaller datasets (which may not apply to you specifically) trying to adapt popular architectures to their needs are disadvantaged. Compare the attention being given to energy budgets and similar constraints for inference as opposed to training and the disparity becomes fairly obvious.
that is an interesting point, and i feel like it generalizes to the fact that using more efficient architecture that was perhaps was designed by someone with a lesser training budget. Although I must say that from my limited DL paper reading, efficient small-scale novel architecture doesn't necessarily comes from cash-strapped researchers, as a more efficient(energy and time) would be of huge economic value also to companies like OpenAI, who have spent huge amounts on training GPTs.
[1] - https://www.bioinf.jku.at/publications/older/2604.pdf
I guess it is all about which questions you are trying to answer. Generally speaking at the OpenAI and Deepmind end of the scale the question is "(How) can we escape the diminishing returns of applying ever more data and compute to improve upon the SOTA?" rather than "(How) can we improve upon the SOTA without applying more data and compute?".
My two cents: I think LeCun is totally right about this one, but it appears perhaps politically incorrect to be scientifically correct.