Page 9 - february_013019abab.indd
P. 9

FEBRUARY 2019      TAP TIMES          -9-           Humans identify objects primarily from
                                                                their shape, Dr. Kellman said. The re-
              Artifi cial Intelligence (from 8)                  searchers suspected the computer net-
                                                                works were using a different method.
        VGG-19 thought there was 0 percent
        chance that the elephant was an elephant
        and only a 0.41 percent chance the teapot               Networks did not discern an elephant
        was a teapot.                                           In another experiment, using a different
                                                                learning network, AlexNet and VGG-19,
                                                                the programs were unable to recognize a
        Network looks for texture                               glass fi gurine of an elephant.
        Its fi rst choice for the teapot was a golf
        ball, which shows that the artifi cial intel-            In another experiment, the researchers
        ligence network looks at the texture of an              showed both networks 40 images in solid
        object more so than its shape, said lead                black.  The networks did better identify-
        author Nicholas Baker, a UCLA psychology                ing these images.  The researchers think
        graduate student.                                       the networks did better with black images
                                                                because the items lack edges that confuse
        “It’s absolutely reasonable for the golf ball           the machines.
        to come up, but alarming that the teapot
        doesn’t come up anywhere among the                      Researchers concluded that humans see
        choices,” Dr. Kellman said. “It’s not picking           the entire object, while AI networks iden-
        up shape.”                                              tify fragments. ucla.edu









                                                        ISO 9001:2015
   4   5   6   7   8   9   10   11   12   13   14