Thread by @ylecun

@timnitGebru If I had wanted to "reduce harms caused by ML to dataset bias", I would have said "ML systems are biased *only* when data is biased".
But I'm absolutely *not* making that reduction.
1/N

@timnitGebru I'm making the point that in the *particular* *case* of *this* *specific* *work*, the bias clearly comes from the data.
2/N

@timnitGebru There are many causes for *societal* bias in ML systems
(not talking about the more general inductive bias here).
1. the data, how it's collected and formatted.
2. the features, how they are designed
3. the architecture of the model
4. the objective function
5. how it's deployed

@timnitGebru (that was 3/N).
When you use raw inputs with no hand-crafted features, as is common in modern DL system, #2 becomes a considerably less important source of designer-caused bias.
E.g. Modern image reco systems work directly from pixels, and generative models produce raw pixels
4/N

@timnitGebru Now, if you use someone else's pre-trained model as a feature extractor, your features will contain the biases of that system (as @soumithchintala correctly pointed out in a comment to my tweet).
5/N

@timnitGebru @soumithchintala Concerning #3, the architecture of the model regularizer included, must have at least *some* level of inductive bias, owing to no-free-lunch theorems
And that bias could very well cause societal bias in the result too.
6/N

@timnitGebru @soumithchintala But one may ask whether modern *generic* model architectures like logistic regression, fully-connected neural nets, or ConvNet have significant intrinsic, built-in societal biases.
My guess is no.
I'm ready to change my opinion in front of theoretical or empirical evidence.
7/N

@timnitGebru @soumithchintala Concerning #4, the objective *can* be biased, of course.
But again, one may ask whether a *generic* objective (like mean squared error) has built-in societal bias.
My guess is "not much".
But again, I'm ready to change opinion in front of evidence to the contrary.
8/N

@timnitGebru @soumithchintala So, in modern DL systems working from raw input (no hand-crafted features) with generic architectures, regularizers and objectives, the *main* source of bias will be in the training data.
The point of my tweet was that this is the case for the face super-res work in question.
9/N

@timnitGebru @soumithchintala Now, there can be much worse causes of societal bias and more dire effects of societal bias in the *deployment* of the system.
This was the point of my comment about the *relative* importance of paying attention to bias in development and deployment, versus research.
10/N

@timnitGebru @soumithchintala Minimizing the effect of bias in data can be done in a number of ways.
DL systems have this peculiarity (due the non-convexity of the loss) that they will not develop features for categories of samples that are rare.
This does not happen with logistic regression and such.
11/N

@timnitGebru @soumithchintala An easy way is to be aware of imbalance in the data, and equalize the frequencies of samples from the various categories.
I've been using this method for decades.
FB uses this in its facrec system.
12/N

@timnitGebru @soumithchintala My metaphor for this is: if you go to medical school, you don't spend time studying the common flu and some rare disease in proportion to their frequencies in patients.
You spend comparatively more time on rare diseases because you need to develop the "features" for them.
13/N

@timnitGebru @soumithchintala Another very new way to identify and minimize bias is the recent work on "Invariant Risk Minimization" by Arjovsky et al. co-authored by my dear friend and colleague Léon Bottou arxiv.org/abs/1907.02893
14/N

@timnitGebru @soumithchintala The idea is to train an ML system to be invariant to changes on the proportion of samples with features that are spuriously correlated with the desired output.
I find this very promising, though clearly not the be-all, end-all.
15/N

@timnitGebru @soumithchintala Post-scriptum: I think people like us should strive to discuss the substance of these questions in a non-emotional and rational manner.
It's hard sometimes.
But that's why we can call ourselves scientists.
16/N

@timnitGebru @soumithchintala It's also important to avoid assuming bad intent from your interlocutor.
It only serves to inflame emotions, to hurt people who could be helpful, to mask the real issues, to delay the development of meaningful solutions, and to delay meaningful action.
17/N
N=17.