I can't help but shake my head at this whole story.
Someone (who by the way describes themselves more as an artist than anything else in public bios) builds a ML "bullshit artist" (literally) that hallucinates portraits from ridiculously undersampled data. Someone else runs a low resolution image of Obama through a hallucinating computer program, and gets a hallucination.
The story should end here at "these hallucinated portraits have basically nothing to do with the input data, they're meaningless". I would bet that input images that are this low resolution are difficult or impossible to determine the race of, no matter how diverse the training data for this model could have been. Note that the skin color of the low resolution Obama and the random "white" guy it generated are basically the same!? I doubt this pipeline will ever be capable of not confusing the race of its input and output.
What is on display here is the uselessness of this ML pipeline, not its racist nature. I have seen probably a dozen of examples of this ML pipeline failing hilariously in ways that have nothing to do with race. Yet this failure is suddenly fascinating to the entire tech journalism space as an example of ML algorithms being racist?!
And somehow, despite the absurdity of this situation, even ML "luminaries" get drawn into a destructive whirlpool of social justice debate about a hallucinating computer program that doesn't even work sometimes? What the fuck are we all doing here?
Someone (who by the way describes themselves more as an artist than anything else in public bios) builds a ML "bullshit artist" (literally) that hallucinates portraits from ridiculously undersampled data. Someone else runs a low resolution image of Obama through a hallucinating computer program, and gets a hallucination.
The story should end here at "these hallucinated portraits have basically nothing to do with the input data, they're meaningless". I would bet that input images that are this low resolution are difficult or impossible to determine the race of, no matter how diverse the training data for this model could have been. Note that the skin color of the low resolution Obama and the random "white" guy it generated are basically the same!? I doubt this pipeline will ever be capable of not confusing the race of its input and output.
What is on display here is the uselessness of this ML pipeline, not its racist nature. I have seen probably a dozen of examples of this ML pipeline failing hilariously in ways that have nothing to do with race. Yet this failure is suddenly fascinating to the entire tech journalism space as an example of ML algorithms being racist?!
And somehow, despite the absurdity of this situation, even ML "luminaries" get drawn into a destructive whirlpool of social justice debate about a hallucinating computer program that doesn't even work sometimes? What the fuck are we all doing here?