| Popis: |
We consider a gradient-based optimization method applied to a function $\mathcal{L}$ of a vector of variables $\boldsymbol{\theta}$, in the case where $\boldsymbol{\theta}$ is represented as a tuple of tensors $(\mathbf{T}_1, \cdots, \mathbf{T}_S)$. This framework encompasses many common use-cases, such as training neural networks by gradient descent. First, we propose a computationally inexpensive technique providing higher-order information on $\mathcal{L}$, especially about the interactions between the tensors $\mathbf{T}_s$, based on automatic differentiation and computational tricks. Second, we use this technique at order 2 to build a second-order optimization method which is suitable, among other things, for training deep neural networks of various architectures. This second-order method leverages the partition structure of $\boldsymbol{\theta}$ into tensors $(\mathbf{T}_1, \cdots, \mathbf{T}_S)$, in such a way that it requires neither the computation of the Hessian of $\mathcal{L}$ according to $\boldsymbol{\theta}$, nor any approximation of it. The key part consists in computing a smaller matrix interpretable as a "Hessian according to the partition", which can be computed exactly and efficiently. In contrast to many existing practical second-order methods used in neural networks, which perform a diagonal or block-diagonal approximation of the Hessian or its inverse, the method we propose does not neglect interactions between layers. Finally, we can tune the coarseness of the partition to recover well-known optimization methods: the coarsest case corresponds to Cauchy's steepest descent method, the finest case corresponds to the usual Newton's method. |
