深度学习技术趋势报告_Yan_LeCun.ppt

Y LeCun Deep Learning Yann Le Cun Facebook AI Research, Center for Data Science, NYU Courant Institute of Mathematical Sciences, NYU Y LeCun Should We Copy the Brain to Build Intelligent Machines? The brain is an existence proof of intelligent machines – The way birds and bats were an existence proof of heavier-than-air flight Shouldnt we just copy it? – Like Clément Ader copied the bat? The answer is no! But we should draw inspiration from it. LAvion III de Clément Ader, 1897 (Musée du CNAM, Paris) His “Eole” took off from the ground in 1890, 13 years before the Wright Brothers. 85x10 neurons 10 synapses/neuron → 10 250 million neurons per mm . Y LeCun The Brain 9 4 15 synapses 1.4 kg, 1.7 liters 2 180,000 km of “wires” 3 All animals can learn Learning is inherent to intelligence Learning modifies the efficacies of synapses Learning causes synapses to strengthen or weaken, to appear or disappear. [John A Beal] [Viren Jain et al.] Y LeCun The Brain: an Amazingly Efficient “Computer” 10 11 neurons, approximately 4 10 “spikes” go through each synapse per second on average 10 16 “operations” per second 25 Watts Very efficient 1.4 kg, 1.7 liters 2500 cm2 Unfolded cortex [Thorpe Fabre-Thorpe 2001] Y LeCun Fast Processors Today Intel Xeon Phi CPU 2x10 12 operations/second 240 Watts 60 (large) cores $3000 NVIDIA Titan-Z GPU 8x10 12 operations/second 500 Watts 5760 (small) cores $3000 Are we only a factor of 10,000 away from the power of the human brain? Probably more like 1 million: synapses are complicated A factor of 1 million is 30 years of Moores Law 2045? Y LeCun Can we build AI systems by copying the brain? Are computers only a factor of 10,000 away from the power of the human brain? Probably more like 1 million: synapses are complicated A factor of 1 million is 30 years of Moores Law Will computers be as intelligent as human by 2045? Compute power is not the whole story Moores Law may not continue for that long W