Search

Numerical nonlinear algebra is a computational paradigm that uses numerical analysis to study polynomial equations. Its origins were methods to solve systems of polynomial equations based on the classical theorem of B\'ezout. This was decisively linked to modern developments in algebraic geometry by the polyhedral homotopy algorithm of Huber and Sturmfels, which exploits the combinatorial structure of the equations and led to efficient software for solving polynomial equations. Subsequent growth of numerical nonlinear algebra continues to be informed by algebraic geometry and its applications. These include new approaches to solving, algorithms for studying positive-dimensional varieties, certification, and a range of applications both within mathematics and from other disciplines. With new implementations, numerical nonlinear algebra is now a fundamental computational tool for algebraic geometry and its applications. We survey some of these innovations and some recent applications., Comment: 40 pages, many figures

Muntjac is an open-source collection of components which can be used to build a multicore, Linux-capable system-on-chip. This includes a 64-bit RISC-V core, a cache subsystem, and TileLink interconnect allowing cache-coherent multicore configurations. Each component is easy to understand, verify, and extend, with most being configurable enough to be useful across a wide range of applications., Comment: To be published in the First Workshop on Open-Source Computer Architecture Research

We present a novel method for using Neural Networks (NNs) for finding solutions to a class of Partial Differential Equations (PDEs). Our method builds on recent advances in Neural Radiance Field research (NeRFs) and allows for a NN to converge to a PDE solution much faster than classic Physically Informed Neural Network (PINNs) approaches.

Deep Graph Neural Networks (GNNs) show promising performance on a range of graph tasks, yet at present are costly to run and lack many of the optimisations applied to DNNs. We show, for the first time, how to systematically quantise GNNs with minimal or no loss in performance using Network Architecture Search (NAS). We define the possible quantisation search space of GNNs. The proposed novel NAS mechanism, named Low Precision Graph NAS (LPGNAS), constrains both architecture and quantisation choices to be differentiable. LPGNAS learns the optimal architecture coupled with the best quantisation strategy for different components in the GNN automatically using back-propagation in a single search round. On eight different datasets, solving the task of classifying unseen nodes in a graph, LPGNAS generates quantised models with significant reductions in both model and buffer sizes but with similar accuracy to manually designed networks and other NAS results. In particular, on the Pubmed dataset, LPGNAS shows a better size-accuracy Pareto frontier compared to seven other manual and searched baselines, offering a 2.3 times reduction in model size but a 0.4% increase in accuracy when compared to the best NAS competitor. Finally, from our collected quantisation statistics on a wide range of datasets, we suggest a W4A8 (4-bit weights, 8-bit activations) quantisation strategy might be the bottleneck for naive GNN quantisations.

The high energy costs of neural network training and inference led to the use of acceleration hardware such as GPUs and TPUs. While this enabled us to train large-scale neural networks in datacenters and deploy them on edge devices, the focus so far is on average-case performance. In this work, we introduce a novel threat vector against neural networks whose energy consumption or decision latency are critical. We show how adversaries can exploit carefully crafted $\boldsymbol{sponge}~\boldsymbol{examples}$, which are inputs designed to maximise energy consumption and latency. We mount two variants of this attack on established vision and language models, increasing energy consumption by a factor of 10 to 200. Our attacks can also be used to delay decisions where a network has critical real-time performance, such as in perception for autonomous vehicles. We demonstrate the portability of our malicious inputs across CPUs and a variety of hardware accelerator chips including GPUs, and an ASIC simulator. We conclude by proposing a defense strategy which mitigates our attack by shifting the analysis of energy consumption in hardware from an average-case to a worst-case perspective., Comment: Accepted at 6th IEEE European Symposium on Security and Privacy (EuroS&P)

Deep convolutional neural networks (CNNs) are powerful tools for a wide range of vision tasks, but the enormous amount of memory and compute resources required by CNNs pose a challenge in deploying them on constrained devices. Existing compression techniques, while excelling at reducing model sizes, struggle to be computationally friendly. In this paper, we attend to the statistical properties of sparse CNNs and present focused quantization, a novel quantization strategy based on power-of-two values, which exploits the weight distributions after fine-grained pruning. The proposed method dynamically discovers the most effective numerical representation for weights in layers with varying sparsities, significantly reducing model sizes. Multiplications in quantized CNNs are replaced with much cheaper bit-shift operations for efficient inference. Coupled with lossless encoding, we built a compression pipeline that provides CNNs with high compression ratios (CR), low computation cost and minimal loss in accuracy. In ResNet-50, we achieved a 18.08x CR with only 0.24% loss in top-5 accuracy, outperforming existing compression methods. We fully compressed a ResNet-18 and found that it is not only higher in CR and top-5 accuracy, but also more hardware efficient as it requires fewer logic gates to implement when compared to other state-of-the-art quantization methods assuming the same throughput., Comment: To appear in NeurIPS 2019, this is the same paper adapted for viewing on arXiv. TL;DR: Better size/accuracy trade-off of compressed sparse models with focused quantization. 11 pages, 5 figures, 4 tables

Numerical algebraic geometry provides a number of efficient tools for approximating the solutions of polynomial systems. One such tool is the parameter homotopy, which can be an extremely efficient method to solve numerous polynomial systems that differ only in coefficients, not monomials. This technique is frequently used for solving a parameterized family of polynomial systems at multiple parameter values. Parameter homotopies have recently been useful in several areas of application and have been implemented in at least two software packages. This article describes Paramotopy, a new, parallel, optimized implementation of this technique, making use of the Bertini software package. The novel features of this implementation, not available elsewhere, include allowing for the simultaneous solutions of arbitrary polynomial systems in a parameterized family on an automatically generated (or manually provided) mesh in the parameter space of coefficients, front ends and back ends that are easily specialized to particular classes of problems, and adaptive techniques for solving polynomial systems near singular points in the parameter space. This last feature automates and simplifies a task that is important but often misunderstood by non-experts., Comment: Long version of ICMS extended abstract

The problem of geolocation of a transmitter via time difference of arrival (TDOA) and frequency difference of arrival (FDOA) is given as a system of polynomial equations. This allows for the use of homotopy continuation-based methods from numerical algebraic geometry. A novel geolocation algorithm employs numerical algebraic geometry techniques in conjunction with the random sample consensus (RANSAC) method. This is all developed and demonstrated in the setting of only FDOA measurements, without loss of generality. Additionally, the problem formulation as polynomial systems immediately provides lower bounds on the number of receivers or measurements required for the solution set to consist of only isolated points., Comment: This work has been submitted to the IEEE for possible publication

The memory system of a modern embedded processor consumes a large fraction of total system energy. We explore a range of different configuration options and show that a reconfigurable design can make better use of the resources available to it than any fixed implementation, and provide large improvements in both performance and energy consumption. Reconfigurability becomes increasingly useful as resources become more constrained, so is particularly relevant in the embedded space. For an optimised architectural configuration, we show that a configurable cache system performs an average of 20% (maximum 70%) better than the best fixed implementation when two programs are competing for the same resources, and reduces cache miss rate by an average of 70% (maximum 90%). We then present a case study of AES encryption and decryption, and find that a custom memory configuration can almost double performance, with further benefits being achieved by specialising the task of each core when parallelising the program., Comment: Presented at HIP3ES, 2016

Researchers working with mathematical models are often confronted by the related problems of parameter estimation, model validation, and model selection. These are all optimization problems, well-known to be challenging due to non-linearity, non-convexity and multiple local optima. Furthermore, the challenges are compounded when only partial data is available. Here, we consider polynomial models (e.g., mass-action chemical reaction networks at steady state) and describe a framework for their analysis based on optimization using numerical algebraic geometry. Specifically, we use probability-one polynomial homotopy continuation methods to compute all critical points of the objective function, then filter to recover the global optima. Our approach exploits the geometric structures relating models and data, and we demonstrate its utility on examples from cell signaling, synthetic biology, and epidemiology., Comment: References added, additional clarifications

We give a Descartes'-like bound on the number of positive solutions to a system of fewnomials that holds when its exponent vectors are not in convex position and a sign condition is satisfied. This was discovered while developing algorithms and software for computing the Gale transform of a fewnomial system, which is our main goal. This software is a component of a package we are developing for Khovanskii-Rolle continuation, which is a numerical algorithm to compute the real solutions to a system of fewnomials., Comment: 22 pages, 4 figures

Numerical algebraic geometry is the field of computational mathematics concerning the numerical solution of polynomial systems of equations. Bertini, a popular software package for computational applications of this field, includes implementations of a variety of algorithms based on polynomial homotopy continuation. The Macaulay2 package Bertini.m2 provides an interface to Bertini, making it possible to access the core run modes of Bertini in Macaulay2. With these run modes, users can find approximate solutions to zero-dimensional systems and positive-dimensional systems, test numerically whether a point lies on a variety, sample numerically from a variety, and perform parameter homotopy runs.

By using a result from the numerical algebraic geometry package Bertini we show that (up to high numerical accuracy) a specific set of degree 6 and degree 9 polynomials cut out the secant variety $\sigma_{4}(\mathbb{P}^{2}\times \mathbb{P} ^{2} \times \mathbb{P} ^{3})$. This, combined with an argument provided by Landsberg and Manivel (whose proof was corrected by Friedland), implies set-theoretic defining equations in degrees 5, 6 and 9 for a much larger set of secant varieties, including $\sigma_{4}(\mathbb{P}^{3}\times \mathbb{P} ^{3} \times \mathbb{P} ^{3})$ which is of particular interest in light of the salmon prize offered by E. Allman for the ideal-theoretic defining equations., Comment: 15 pages. Updated to reflect the referees' suggestions. Also added ancillary files to the arXiv in this version

We use Gale duality for polynomial complete intersections and adapt the proof of the fewnomial bound for positive solutions to obtain the bound (e^4+3) 2^(k choose 2) n^k/4 for the number of non-zero real solutions to a system of n polynomials in n variables having n+k+1 monomials whose exponent vectors generate a subgroup of Z^n of odd index. This bound exceeds the bound for positive solutions only by the constant factor (e^4+3)/(e^2+3) and it is asymptotically sharp for k fixed and n large., Comment: 5 pages

A path tracking algorithm that adaptively adjusts precision is presented. By adjusting the level of precision in accordance with the numerical conditioning of the path, the algorithm achieves high reliability with less computational cost than would be incurred by raising precision across the board. We develop simple rules for adjusting precision and show how to integrate these into an algorithm that also adaptively adjusts the step size. The behavior of the method is illustrated on several examples arising as homotopies for solving systems of polynomial equations.