We consider finding flat, local minimizers by adding average weight perturbations. Given a nonconvex function $f: \mathbb{R}^d \rightarrow \mathbb{R}$ and a $d$-dimensional distribution $\mathcal{P}$ which is symmetric at zero, we perturb the weight of $f$ and define $F(W) = \mathbb{E}[f({W + U})]$, where $U$ is a random sample from $\mathcal{P}$. This injection induces regularization through the Hessian trace of $f$ for small, isotropic Gaussian perturbations. Thus, the weight-perturbed function biases to minimizers with low Hessian trace. Several prior works have studied settings related to this weight-perturbed function by designing algorithms to improve generalization. Still, convergence rates are not known for finding minima under the average perturbations of the function $F$. This paper considers an SGD-like algorithm that injects random noise before computing gradients while leveraging the symmetry of $\mathcal{P}$ to reduce variance. We then provide a rigorous analysis, showing matching upper and lower bounds of our algorithm for finding an approximate first-order stationary point of $F$ when the gradient of $f$ is Lipschitz-continuous. We empirically validate our algorithm for several image classification tasks with various architectures. Compared to sharpness-aware minimization, we note a 12.6% and 7.8% drop in the Hessian trace and top eigenvalue of the found minima, respectively, averaged over eight datasets. Ablation studies validate the benefit of the design of our algorithm.
相關內容
We introduce TeraHAC, a $(1+\epsilon)$-approximate hierarchical agglomerative clustering (HAC) algorithm which scales to trillion-edge graphs. Our algorithm is based on a new approach to computing $(1+\epsilon)$-approximate HAC, which is a novel combination of the nearest-neighbor chain algorithm and the notion of $(1+\epsilon)$-approximate HAC. Our approach allows us to partition the graph among multiple machines and make significant progress in computing the clustering within each partition before any communication with other partitions is needed. We evaluate TeraHAC on a number of real-world and synthetic graphs of up to 8 trillion edges. We show that TeraHAC requires over 100x fewer rounds compared to previously known approaches for computing HAC. It is up to 8.3x faster than SCC, the state-of-the-art distributed algorithm for hierarchical clustering, while achieving 1.16x higher quality. In fact, TeraHAC essentially retains the quality of the celebrated HAC algorithm while significantly improving the running time.
We generalize the leverage score sampling sketch for $\ell_2$-subspace embeddings, to accommodate sampling subsets of the transformed data, so that the sketching approach is appropriate for distributed settings. This is then used to derive an approximate coded computing approach for first-order methods; known as gradient coding, to accelerate linear regression in the presence of failures in distributed computational networks, \textit{i.e.} stragglers. We replicate the data across the distributed network, to attain the approximation guarantees through the induced sampling distribution. The significance and main contribution of this work, is that it unifies randomized numerical linear algebra with approximate coded computing, while attaining an induced $\ell_2$-subspace embedding through uniform sampling. The transition to uniform sampling is done without applying a random projection, as in the case of the subsampled randomized Hadamard transform. Furthermore, by incorporating this technique to coded computing, our scheme is an iterative sketching approach to approximately solving linear regression. We also propose weighting when sketching takes place through sampling with replacement, for further compression.
A function $f : U \to \{0,\ldots,n-1\}$ is a minimal perfect hash function for a set $S \subseteq U$ of size $n$, if $f$ bijectively maps $S$ into the first $n$ natural numbers. These functions are important for many practical applications in computing, such as search engines, computer networks, and databases. Several algorithms have been proposed to build minimal perfect hash functions that: scale well to large sets, retain fast evaluation time, and take very little space, e.g., 2 - 3 bits/key. PTHash is one such algorithm, achieving very fast evaluation in compressed space, typically several times faster than other techniques. In this work, we propose a new construction algorithm for PTHash enabling: (1) multi-threading, to either build functions more quickly or more space-efficiently, and (2) external-memory processing to scale to inputs much larger than the available internal memory. Only few other algorithms in the literature share these features, despite of their big practical impact. We conduct an extensive experimental assessment on large real-world string collections and show that, with respect to other techniques, PTHash is competitive in construction time and space consumption, but retains 2 - 6$\times$ better lookup time.
A package query returns a package - a multiset of tuples - that maximizes or minimizes a linear objective function subject to linear constraints, thereby enabling in-database decision support. Prior work has established the equivalence of package queries to Integer Linear Programs (ILPs) and developed the SketchRefine algorithm for package query processing. While this algorithm was an important first step toward supporting prescriptive analytics scalably inside a relational database, it struggles when the data size grows beyond a few hundred million tuples or when the constraints become very tight. In this paper, we present Progressive Shading, a novel algorithm for processing package queries that can scale efficiently to billions of tuples and gracefully handle tight constraints. Progressive Shading solves a sequence of optimization problems over a hierarchy of relations, each resulting from an ever-finer partitioning of the original tuples into homogeneous groups until the original relation is obtained. This strategy avoids the premature discarding of high-quality tuples that can occur with SketchRefine. Our novel partitioning scheme, Dynamic Low Variance, can handle very large relations with multiple attributes and can dynamically adapt to both concentrated and spread-out sets of attribute values, provably outperforming traditional partitioning schemes such as KD-tree. We further optimize our system by replacing our off-the-shelf optimization software with customized ILP and LP solvers, called Dual Reducer and Parallel Dual Simplex respectively, that are highly accurate and orders of magnitude faster.
The quadrotor is a $6$ degrees-of-freedom (DoF) system with underactuation. Adding a spherical pendulum on top of a quadrotor further complicates the task of achieving any output tracking while stabilizing the rest. In this report, we present different types of controllers for the nonlinear dynamical system of quadrotor and pendulum combination, utilizing feedback-linearization and control Lyapunov function with quadratic programming (CLF-QP) approaches. We demonstrated trajectory tracking for quadrotor-only case as well as quadrotor-pendulum-combined case.
A dictionary data structure maintains a set of at most $n$ keys from the universe $[U]$ under key insertions and deletions, such that given a query $x \in [U]$, it returns if $x$ is in the set. Some variants also store values associated to the keys such that given a query $x$, the value associated to $x$ is returned when $x$ is in the set. This fundamental data structure problem has been studied for six decades since the introduction of hash tables in 1953. A hash table occupies $O(n\log U)$ bits of space with constant time per operation in expectation. There has been a vast literature on improving its time and space usage. The state-of-the-art dictionary by Bender, Farach-Colton, Kuszmaul, Kuszmaul and Liu [BFCK+22] has space consumption close to the information-theoretic optimum, using a total of \[ \log\binom{U}{n}+O(n\log^{(k)} n) \] bits, while supporting all operations in $O(k)$ time, for any parameter $k \leq \log^* n$. The term $O(\log^{(k)} n) = O(\underbrace{\log\cdots\log}_k n)$ is referred to as the wasted bits per key. In this paper, we prove a matching cell-probe lower bound: For $U=n^{1+\Theta(1)}$, any dictionary with $O(\log^{(k)} n)$ wasted bits per key must have expected operational time $\Omega(k)$, in the cell-probe model with word-size $w=\Theta(\log U)$. Furthermore, if a dictionary stores values of $\Theta(\log U)$ bits, we show that regardless of the query time, it must have $\Omega(k)$ expected update time. It is worth noting that this is the first cell-probe lower bound on the trade-off between space and update time for general data structures.
Efficient differential equation solvers have significantly reduced the sampling time of diffusion models (DMs) while retaining high sampling quality. Among these solvers, exponential integrators (EI) have gained prominence by demonstrating state-of-the-art performance. However, existing high-order EI-based sampling algorithms rely on degenerate EI solvers, resulting in inferior error bounds and reduced accuracy in contrast to the theoretically anticipated results under optimal settings. This situation makes the sampling quality extremely vulnerable to seemingly innocuous design choices such as timestep schedules. For example, an inefficient timestep scheduler might necessitate twice the number of steps to achieve a quality comparable to that obtained through carefully optimized timesteps. To address this issue, we reevaluate the design of high-order differential solvers for DMs. Through a thorough order analysis, we reveal that the degeneration of existing high-order EI solvers can be attributed to the absence of essential order conditions. By reformulating the differential equations in DMs and capitalizing on the theory of exponential integrators, we propose refined EI solvers that fulfill all the order conditions, which we designate as Refined Exponential Solver (RES). Utilizing these improved solvers, RES exhibits more favorable error bounds theoretically and achieves superior sampling efficiency and stability in practical applications. For instance, a simple switch from the single-step DPM-Solver++ to our order-satisfied RES solver when Number of Function Evaluations (NFE) $=9$, results in a reduction of numerical defects by $25.2\%$ and FID improvement of $25.4\%$ (16.77 vs 12.51) on a pre-trained ImageNet diffusion model.
We show that the minimax sample complexity for estimating the pseudo-spectral gap $\gamma_{\mathsf{ps}}$ of an ergodic Markov chain in constant multiplicative error is of the order of $$\tilde{\Theta}\left( \frac{1}{\gamma_{\mathsf{ps}} \pi_{\star}} \right),$$ where $\pi_\star$ is the minimum stationary probability, recovering the known bound in the reversible setting for estimating the absolute spectral gap [Hsu et al., 2019], and resolving an open problem of Wolfer and Kontorovich [2019]. Furthermore, we strengthen the known empirical procedure by making it fully-adaptive to the data, thinning the confidence intervals and reducing the computational complexity. Along the way, we derive new properties of the pseudo-spectral gap and introduce the notion of a reversible dilation of a stochastic matrix.
We consider structural equation models (SEMs), in which every variable is a function of a subset of the other variables and a stochastic error. Each such SEM is naturally associated with a directed graph describing the relationships between variables. When the errors are homoscedastic, recent work has proposed methods for inferring the graph from observational data under the assumption that the graph is acyclic (i.e., the SEM is recursive). In this work, we study the setting of homoscedastic errors but allow the graph to be cyclic (i.e., the SEM to be non-recursive). Using an algebraic approach that compares matroids derived from the parameterizations of the models, we derive sufficient conditions for when two simple directed graphs generate different distributions generically. Based on these conditions, we exhibit subclasses of graphs that allow for directed cycles, yet are generically identifiable. We also conjecture a strengthening of our graphical criterion which can be used to distinguish many more non-complete graphs.
We present a linear-time algorithm that, given as input (i) a bipartite Pfaffian graph $G$ of minimum degree three, (ii) a Hamiltonian cycle $H$ in $G$, and (iii) an edge $e$ in $H$, outputs at least three other Hamiltonian cycles through the edge $e$ in $G$. This linear-time complexity of finding another Hamiltonian cycle given one is in sharp contrast to the problem of deciding the existence of a Hamiltonian cycle, which is NP-complete already for cubic bipartite planar graphs; such graphs are Pfaffian. Also, without the degree requirement, we show that it is NP-hard to find another Hamiltonian cycle in a bipartite Pfaffian graph. We present further improved algorithms for finding optimal traveling salesperson tours and counting Hamiltonian cycles in bipartite planar graphs with running times that are not known to hold in general planar graphs. We prove our results by a new structural technique that efficiently witnesses each Hamiltonian cycle $H$ through an arbitrary fixed anchor edge $e$ in a bipartite Pfaffian graph using a two-coloring of the vertices as advice that is unique to $H$. Previous techniques -- the Cut&Count technique of Cygan et al. [FOCS'11, TALG'22] in particular -- were able to reduce the Hamiltonian cycle problem only to essentially counting problems; our results show that counting can be avoided by leveraging properties of bipartite Pfaffian graphs. Our technique also has purely graph-theoretical consequences; for example, we show that every cubic bipartite Pfaffian graph has either zero or at least six distinct Hamiltonian cycles; the latter case is tight for the cube graph.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191