O laboratório de pesquisa Google X, que desenvolve tecnologias como os óculos de realidade aumentada conhecidos como 'Glass Project', conectou 1 mil computadores, total de 16 mil processadores, para formar uma rede neural com 1 bilhão de conexões e, assim, criar uma simulação do cérebro humano.
Conforme reportagem do “The New York Times”, os cientistas apresentaram a essa rede 10 milhões de imagens retiradas de vídeos aleatórios do YouTube para que ela aprendesse sozinha o que é um gato durante três dias.
“Durante o treinamento, nós nunca dissemos: ‘isso é um gato’”, explicou Jeff Dean, da equipe do Google. "Ela basicamente inventou o conceito de gato”. Os resultados da pesquisa revelaram que é possível criar um detector de face sem precisar "rotular" imagens.
Segundo o jornal, a rede neural do Google aprendeu sozinha a reconhecer gatos. O “cérebro do Google” montou uma imagem digital de um gato ao selecionar características gerais depois de ser exposto a milhões de imagens.
“A ideia é que, em vez de pesquisadores tentarem descobrir como encontrar as beiradas, jogue uma tonelada de dados no algoritmo e deixe falar, e assim o software automaticamente aprende a partir desses dados”, explicou Andrew Y. Ng, cientista da Universidade de Stanford.
Nesta semana, os pesquisadores vão apresentar os resultados do trabalho em uma conferência em Edimburgo, na Escócia.
The Google research laboratory X, which develops technologies like augmented reality glasses known as 'Glass Project', 1000 connected computers, totaling 16,000 processors to form a neural network with a billion connections and thus create a simulation the human brain.
As reported in "The New York Times," scientists presented to the network 10 million images taken from random videos from YouTube to herself that she learned what a cat for three days.
"During training, we never said, 'This is a cat,'" said Jeff Dean, Google team. "She basically invented the concept of cat." The survey results revealed that it is possible to create a face detector without "tag" images.
According to the newspaper, the neural network Google taught herself to recognize cats. The "Google Brain" set up a digital image of a cat when selecting general characteristics after being exposed to millions of images.
"The idea is that instead of researchers trying to figure out how to find the edges, you play a ton of data in the algorithm and let them talk, and so the software automatically learns from these data," explained Andrew Y. Ng, a scientist at Stanford University.
This week, researchers will present the results of work at a conference in Edinburgh, Scotland.
As reported in "The New York Times," scientists presented to the network 10 million images taken from random videos from YouTube to herself that she learned what a cat for three days.
"During training, we never said, 'This is a cat,'" said Jeff Dean, Google team. "She basically invented the concept of cat." The survey results revealed that it is possible to create a face detector without "tag" images.
According to the newspaper, the neural network Google taught herself to recognize cats. The "Google Brain" set up a digital image of a cat when selecting general characteristics after being exposed to millions of images.
"The idea is that instead of researchers trying to figure out how to find the edges, you play a ton of data in the algorithm and let them talk, and so the software automatically learns from these data," explained Andrew Y. Ng, a scientist at Stanford University.
This week, researchers will present the results of work at a conference in Edinburgh, Scotland.
Imagem de um gato que a rede neural do Google aprendeu a reconhecer sozinha
Picture of a cat that the neural network learned to recognize Google alone
Sem comentários:
Enviar um comentário