″.. we just don’t have very compelling example domains where ML systems understand important things in ways we can’t. ”
I’m guessing you mean in ways humans can’t even in principle?
Regardless, here’s something people might find amusing—researchers found that a simple VGG-like 3D CNN model can look at electron microscope images of neural tissue and do a task that humans don’t know how to do. The network distinguishes neurons that specialize in certain neurotransmitters. From the abstract to this preprint:
“The network successfully discriminates between six types of neurotransmitters (GABA, glutamate, acetylcholine, serotonin, dopamine, and octopamine) with an average accuracy of 87% for individual synapses and 94% for entire neurons, assuming each neuron expresses only one neurotransmitter. This result is surprising as there are often no obvious cues in the EM images that human observers can use to predict neurotransmitter identity.”
They are developing explainability techniques to try to figure out how the CNN does this classification (see the figures in this preprint). In addition to the custom methods they’ve developed I know they have also used more bog-standard activation maximization techniques as well (personal communication with Jan Funke in January year). Jan told me he’s read Chris Olah et al.’s publications in Distill. They think the network may be cuing in on subtle differences in the size/shape of vesicles.
I’m guessing you mean in ways humans can’t even in principle?
Regardless, here’s something people might find amusing—researchers found that a simple VGG-like 3D CNN model can look at electron microscope images of neural tissue and do a task that humans don’t know how to do. The network distinguishes neurons that specialize in certain neurotransmitters. From the abstract to this preprint:
They are developing explainability techniques to try to figure out how the CNN does this classification (see the figures in this preprint). In addition to the custom methods they’ve developed I know they have also used more bog-standard activation maximization techniques as well (personal communication with Jan Funke in January year). Jan told me he’s read Chris Olah et al.’s publications in Distill. They think the network may be cuing in on subtle differences in the size/shape of vesicles.