Blurry Vision

It looks like the robot brains are beating the sensory prostheses. Telescopes and microscopes are way better at seeing than humans and our smartphones. But we may not need to put these advanced lenses in our phones or our future robots in order to make them superhuman. Instead, we need to fill them with fuzzier brains. 

Artificial intelligence may seem like a hyper-concise, over-literal, braniac, but not these days. The new generation of AI, known simply as deep learning, is the opposite of this. It is less like a calculator and more like a guess. (Although technically it's both). It favors approximation over concision.

Relative to us humans, however, its results are more concise than we could ever attain. A research group via UCLA has outfitted regular lenses found on a smartphone with a 3D-printed microscope attachment and an AI that makes a really good, phenomenally good guess at what it sees. Their invention gives us back an image of the same precision as a lab-grade microscope.

Their fuzzy AI brain "learns" how to see under high resolution by being fed both images taken with the regular smartphone, and with lab-grade microscopes. Using thousands of examples, the brain compares the one to the other, and eventually learns how we get from the one to the other - if I give you this crappy fuzzy image, how do you make it into that sharp-shaped product? It learns how to do that, using algorithms that we don't program (the brain programs itself).

It's only superficially ironic that the loose AI analogy of 'blurry vision' is used by these deep learning techniques to see in high res. The real story here is that we're using the brains, or the software, of our robots to liberate us from the hardware restrictions.


Deep learning transforms smartphone microscopes into laboratory-grade devices
Apr 2018, phys.org

Yair Rivenson et al. Deep Learning Enhanced Mobile-Phone Microscopy, ACS Photonics (2018). DOI: 10.1021/acsphotonics.8b00146
Provided by: University of California, Los Angeles