Language models employ a very large number of trainable parameters. Despite being highly overparameterized, these networks often achieve good out-of-sample test performance on the original task and easily fine-tune to related tasks. Recent observations involving, for example, intrinsic dimension of the objective landscape and the lottery ticket hypothesis, indicate that often training actively involves only a small fraction of the parameter space. Thus, a question remains how large a parameter space needs to be in the first place -- the evidence from recent work on model compression, parameter sharing, factorized representations, and knowledge distillation increasingly shows that models can be made much smaller and still perform well. Here, we focus on factorized representations of matrices that underpin dense, embedding, and self-attention layers. We use low-rank factorized representation of a reshaped and rearranged original matrix to achieve space efficient and expressive linear layers. We prove that stacking such low-rank layers increases their expressiveness, providing theoretical understanding for their effectiveness in deep networks. In Transformer models, our approach leads to more than tenfold reduction in the number of total trainable parameters, including embedding, attention, and feed-forward layers, with little degradation in on-task performance. The approach operates out-of-the-box, replacing each parameter matrix with its compact equivalent while maintaining the architecture of the network.

In this section, we prove the Theorem 2.1, which states a problem P and its' orthogonal transformed problem Q(P) = {{Qxi}Ni=1,f}have identical optimal solutions if Qis orthogonal matrix: QQT = QTQ = I. As we mentioned in Section 2.2, reward R is a function of a1:T (solution sequences), ||xi xj||i,j {1,...N} (relative distances) and f (nodes features). And Let R (P)is optimal value of problem P: i.e. Then, the remaining proof is to show Q(P)has an identical solution set with P. Let optimal solution set Π (P) = {πi(P)}Mi=1, where πi(P)indicates optimal solution of P and M is the number of heterogeneous optimal solution. Conversely, For any πi(P) Π (P), they have sample optimal value with Q(P): R(πi(P);P) = R (P) = R (Q(P)) Thus, πi(P) Π (Q(P)).

Joint matrix triangularization is often used for estimating the joint eigenstructure of a set M of matrices, with applications in signal processing and machine learning. We consider the problem of approximate joint matrix triangularization when the matrices in M are jointly diagonalizable and real, but we only observe a set M' of noise perturbed versions of the matrices in M. Our main result is a first-order upper bound on the distance between any approximate joint triangularizer of the matrices in M' and any exact joint triangularizer of the matrices in M. The bound depends only on the observable matrices in M' and the noise level. In particular, it does not depend on optimization specific properties of the triangularizer, such as its proximity to critical points, that are typical of existing bounds in the literature. To our knowledge, this is the first a posteriori bound for joint matrix decomposition. We demonstrate the bound on synthetic data for which the ground truth is known.

Neural Information Processing Systems http://nips.cc/

This transformation naturally implements a parameterization for the relaxation of permutation matrices, allowing for gradient-based optimization of problems involving permutations.

HR-based orthogonal fine-tuning is equivalent to an adaptive low-rank adaptation.

In this paper, we introduce a new class of structured matrices, which unifies and generalizes structured classes from previous works.

However, only the smooth setting was considered. To close this gap, we introduce and analyze O-ZD, the first structured finite-difference algorithm for non-smooth black-box optimization.