优快云

HardwareFP32 can do 2 floating points operations (multiply and add) per clock cycle.if we use tensor core to calculate the computational capacity: (72 SM)x(8 tensor core per SM)x (128 FLOP/cycle) * (1.77Gcycle/sec)=130 TFLOP/secAnd all output elements

AlexNetOutput size:通道数和滤波器数量保持一致，均为64H/W = (H - K + 2P)/S + 1 = (227-11+4)/4+1=56Memory(KB):Number of output elements: C * Ｈ* Ｗ＝64*56 *56=200704; Bytes per element=4 (for 32-bit floating point). KB=200704 * 4/1024 = 784Parameters(k):Weight shape =

why RELU works?Vector DerivationBut Jacobian is sparse: off-diagonal entries are all zero! Never explicitly form Jacobian.结果我们发现，dy/dx_1,1 = [3,2,1,-1]其恰好是w的第一行，故而不需要显性地进行求解。dL/dxi,j=(dy/dxij)⋅(dL/dy)=(wi)⋅(dL/dyi,:)\begin{array}{l}\mathrm{dL} / \m

最近邻算法复杂度Train: O(1)对于最近邻算法，训练阶段只是简单地记忆下所有的训练数据，对于浅拷贝而言只需要记录其指向数据集的指针Test: O(N)此处N为训练集的大小，即每做一次判断需要把测试样例和训练集的每一个样本进行Distance计算，对于test而言，O(N)的复杂度是不可接受的线性分类器的三个视角从代数角度来看，学习了一个线性函数的权重，即一个权重矩阵从视觉角度来看，对于每一个类别，找到一个具有代表性的模板从几何角度来看，寻找的是一个超平面正则化