Something interesting in Cryo-EM

Something interesting.

1.其实早在2013年,PRIME的作者就想过用GD去做三维重构,但是他说:

In our hands, gradient-based methods fail to produce acceptable solutions to the problem.

在我们手中,基于梯度的方法无法为该问题提供可接受的解决方案。

That gradient-based optimization methods fail indicates either that the goal function is nondifferentiable or that the goal function landscape is nonconvex. A nonconvex goal function landscape could be explained by two image properties: the many rotational autocorrelation peaks and the low SNR obtained in the low-dose regime. The errors in the goal function landscape increase as the SNRs of the images decrease. Gradient-based optimizers become trapped in poor local optima due to the noise. We therefore use a direct search-based optimization method that does not require gradient information

基于梯度的优化方法失败表明目标函数是不可微的,或者目标函数格局是非凸的。 非凸目标函数的格局可以由两个图像属性来解释:许多旋转自相关峰和在低剂量方案中获得的低SNR。 随着图像的SNR降低,目标函数景观中的误差增加。 基于梯度的优化器由于噪声而陷入了较差的局部最优状态。 因此,我们使用不需要梯度信息的基于直接搜索的优化方法

其实蛮可惜,如果多想一步,用SGD的话,可能就没有后来的cryoSPARC啥事儿了

2.

0%