Extrapolation of results of Fig 1b to the number of synapses in macaque V1.

Naive, unrestricted backprop learning would use 105 × more synaptic updates and 103 × more M1 energy than minimally required.

محفوظ في:

التفاصيل البيبلوغرافية
المؤلف الرئيسي:	Mark C. W. van Rossum (19959623) (author)
مؤلفون آخرون:	Aaron Pache (19959626) (author)
منشور في:	2024
الموضوعات:	Neuroscience Biological Sciences not elsewhere classified Information Systems not elsewhere classified significant metabolic cost often much larger task like mnist results might lead feedforward neural networks div >< p parsimonious energy model 8 </ sup learning </ p task requires biology networks save energy little energy energy required energy needed synaptic updates synaptic plasticity restrict plasticity might require likely evolved least 10 learn using large updates large savings fuel computation energetic costs competitive plasticity biological plasticity biological learning better understanding better match already carries accuracy appears
الوسوم:	إضافة وسم لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!

مواد مشابهة