학습영상: 메타코드-딥러닝 강의 컴퓨터 비전 인식모델 개발 1편
출처: 메타코드M
velocity = momentum * velocity - learning_rate * gradient
w = w + velocity
cache = decay_rate * cache + (1 - decay_rate) * gradient^2
w = w - (learning_rate / sqrt(cache + epsilon)) * gradient
m = beta1 * m + (1 - beta1) * gradient
v = beta2 * v + (1 - beta2) * gradient^2
m_hat = m / (1 - beta1^t)
v_hat = v / (1 - beta2^t)
w = w - (learning_rate / (sqrt(v_hat) + epsilon)) * m_hat