Salta il testo dell’articolo
Nina wanted one pile of photos : every picture of her cat , Miso , and none of the empty rooms . The folder held years of shots , and she wanted an album from it . Writing a rule for every pose was hopeless . A cat can be a shadow on a sofa , a tail under a chair , or a face turned away . She turned to machine learning , which means the computer learns from examples instead of following a rule she writes for each case . She labeled a stack of pictures by hand . Cat . Not a cat . Then she let a neural network train on them . A network like this is built in layers of small units , sometimes called neurons . Each link between units has a weight . When the network guessed wrong , those weights shifted a little , toward the label she had given . The first layer took the raw picture . Middle layers worked on what the earlier ones passed forward . The last layer answered : cat , or not . Picture networks often learn in a stack . Early layers respond to edges and patches of color . Deeper layers respond to shapes . Later ones can respond to an animal or a face . Nina could not open the network and point to the step that meant ear . The decision sat in the weights , not in a sentence she could read . The early results annoyed her . Bright photos of the kitchen were marked as Miso even when the cat was not in them . Dark photos of Miso on the sofa were missed . She looked back at her labels . Most of the cat photos had been taken in that kitchen , in the afternoon light . The network had learned the room . The labels had taught it a shortcut , and a shortcut in the training data can send the answer the wrong way . She went out with the camera . She photographed Miso in the hallway , on the night street , and on the gray sofa . She photographed the empty kitchen too , and labeled those as not a cat . She trained the network again . This time the bright empty kitchen stayed out of the pile , and the dark sofa photo came in . The album was the pile from that second training . Nina printed it . The network had not been given a definition of a cat . It had been given her labels , including the mistake she had almost left in place .
Continua a leggere con le traduzioni Crea un account gratuito: 7 giorni di Pro, poi 20 traduzioni di parole al giorno oppure un piano a pagamento.
