杨立昆(Yann LeCun)教授列出了一系列机器学习方法与物理学的相似之处,或者说关联( ML methods connected with physics)。Insightful👍
物理 vs. 人工智能,两者的后面都是数学。
– variational Bayesian inference is the math as thermodynamics.
—@ylecun
– Bayesian ensembling is like path integrals.
– backprop can be derived through Lagrangian saddle point method.
– convergence in non-convex objectives is spontaneous symmetry breaking.
– ConvNet pooling is like renormalization group theory and MERA tensor networks.
– and then, ConvNets can be Lorentz-equivariant
——
– 变分贝叶斯推理在数学上类似于热力学。
– 贝叶斯集成类似于路径积分。
– 反向传播可以通过拉格朗日鞍点方法推导出来。
– 非凸目标的收敛是自发对称破缺。
– 卷积神经网络的池化类似于重整化群理论和MERA张量网络。
– 然后,卷积神经网络可以是洛伦兹协变的。