Graph Computation Research Articles

Memristors with Tunable Volatility for Reconfigurable Neuromorphic Computing.

Neuromorphic computing promises an energy-efficient alternative to traditional digital processors in handling data-heavy tasks, primarily driven by the development of both volatile (neuronal) and nonvolatile (synaptic) resistive switches or memristors. However, despite their energy efficiency, memristor-based technologies presently lack functional tunability, thus limiting their competitiveness with arbitrarily programmable (general purpose) digital computers. This work introduces a two-terminal bilayer memristor, which can be tuned among neuronal, synaptic, and hybrid behaviors. The varying behaviors are accessed via facile control over the filament formed within the memristor, enabled by the interplay between the two active ionic species (oxygen vacancies and metal cations). This solution is unlike single-species ion migration employed in most other memristors, which makes their behavior difficult to control. By reconfiguring a single crossbar array of hybrid memristors, two different applications that usually require distinct types of devices are demonstrated - reprogrammable heterogeneous reservoir computing and arbitrary non-Euclidean graph networks. Thus, this work outlines a potential path toward functionally reconfigurable postdigital computers.

Generating Euler Diagrams Through Combinatorial Optimization

AbstractCan a given set system be drawn as an Euler diagram? We present the first method that correctly decides this question for arbitrary set systems if the Euler diagram is required to represent each set with a single connected region. If the answer is yes, our method constructs an Euler diagram. If the answer is no, our method yields an Euler diagram for a simplified version of the set system, where a minimum number of set elements have been removed. Further, we integrate known wellformedness criteria for Euler diagrams as additional optimization objectives into our method. Our focus lies on the computation of a planar graph that is embedded in the plane to serve as the dual graph of the Euler diagram. Since even a basic version of this problem is known to be NP‐hard, we choose an approach based on integer linear programming (ILP), which allows us to compute optimal solutions with existing mathematical solvers. For this, we draw upon previous research on computing planar supports of hypergraphs and adapt existing ILP building blocks for contiguity‐constrained spatial unit allocation and the maximum planar subgraph problem. To generate Euler diagrams for large set systems, for which the proposed simplification through element removal becomes indispensable, we also present an efficient heuristic. We report on experiments with data from MovieDB and Twitter. Over all examples, including 850 non‐trivial instances, our exact optimization method failed only for one set system to find a solution without removing a set element. However, with the removal of only a few set elements, the Euler diagrams can be substantially improved with respect to our wellformedness criteria.

GEPINN: An innovative hybrid method for a symbolic solution to the Lane–Emden type equation based on grammatical evolution and physics-informed neural networks

In this paper, we present an innovative and powerful combination of grammatical evolution and a physics-informed neural network approach for symbolically solving the Lane–Emden type equation, which is a nonlinear ordinary differential equation. We employ a grammatical evolution algorithm based on a context-free grammar to construct a mathematical expression comprising some parameters. Subsequently, these parameters are determined using the physics-informed neural networks approach. To achieve this, the computational graph of the mathematical expression generated in each iteration of the grammatical evolution is treated as a network. To assess the proposed method, we consider the Lane–Emden type equation. The proposed method demonstrated that it is a capable method for symbolically solving nonlinear ordinary differential equations accurately.

Graph network surrogate model for subsurface flow optimization

The optimization of well locations and controls is an important step in the design of subsurface flow operations such as oil production or geological CO2 storage. These optimization problems can be computationally expensive, however, as many potential candidate solutions must be evaluated. In this study, we propose a graph network surrogate model (GNSM) for optimizing well placement and controls. The GNSM transforms the flow model into a computational graph that involves an encoding-processing-decoding architecture. Separate networks are constructed to provide global predictions for the pressure and saturation state variables. Model performance is enhanced through the inclusion of the single-phase steady-state pressure solution as a feature. A multistage multistep strategy is used for training. The trained GNSM is applied to predict flow responses in a 2D unstructured model of a channelized reservoir. Results are presented for a large set of test cases, in which five injection wells and five production wells are placed randomly throughout the model, with a random control variable (bottom-hole pressure) assigned to each well. Median relative error in pressure and saturation for 300 such test cases is 1-2%. The ability of the trained GNSM to provide accurate predictions for new (geologically similar) permeability realizations is demonstrated. Finally, the trained GNSM is used to optimize well locations and controls with a differential evolution algorithm. GNSM-based optimization results are comparable to those from simulation-based optimization, with a runtime speedup factor of 36. Much larger speedups are expected if the method is used for robust optimization, in which each candidate solution is evaluated on multiple geological models.

Performance Evaluation of Parallel Graphs Algorithms Utilizing Graphcore IPU

Recent years have been characterized by increasing interest in graph computations. This trend can be related to the large number of potential application areas. Moreover, increasing computational capabilities of modern computers allowed turning theory of graph algorithms into explorations of best methods for their actual realization. These factors, in turn, brought about ideas like creation of a hardware component dedicated to graph computation; i.e., the Graphcore Intelligent Processor Unit (IPU). Interestingly, Graphcore systems are a hardware implementation of the Bulk Synchronous Parallel paradigm, which seemed to be a mostly theoretical concept from the end of last century. In this context, the question that has to be addressed experimentally is as follows: how good are Graphcore systems in comparison with standard systems that can be used to run graph algorithms, i.e., CPUs and GPUs. To provide a partial response to this broad question, in this contribution, PageRank, Single Source Shortest Path and Breadth-First Search algorithms are used to compare the performance of IPU-deployed algorithms to other parallel architectures. Obtained results clearly show that the Graphcore IPU outperforms other devices for the studied heterogeneous algorithms and, currently, provides best-in-class execution time results for a range of graph sizes and densities.

A Pseudo-Deterministic Noisy Extremal Optimization algorithm for the pairwise connectivity Critical Node Detection Problem

Abstract The critical node detection problem is a central task in computational graph theory due to its large applicability, consisting in deleting $k$ nodes to minimize a certain graph measure. In this article, we propose a new Extremal Optimization-based approach, the Pseudo-Deterministic Noisy Extremal Optimization (PDNEO) algorithm, to solve the Critical Node Detection variant in which the pairwise connectivity is minimized. PDNEO uses an adaptive pseudo-deterministic parameter to switch between random nodes and articulation points during the search, as well as other features, such as noise induction to preserve diversity, greedy search to better exploit the search space and a greater search space exploration mechanism. Numerical experiments on synthetic and real-world networks show the effectiveness of the proposed algorithm compared with existing methods.

A high-performance dataflow-centric optimization framework for deep learning inference on the edge

Edge computing has been emerging as a popular scenario for model inference. However, the inference performance on edge devices (e.g., Multi-Core DSP, FGPA, etc.) suffers from inefficiency due to the lack of highly optimized inference frameworks. Previous model inference frameworks are mainly developed in an operator-centric way, which provides insufficient acceleration to edge-based inference. Besides, the operator-centric framework incurs significant costs for continuous development and maintenance.Targeting the existing drawbacks of operator-centric frameworks, we design Xenos, which can automatically conduct dataflow-centric optimization of the computation graph and accelerate inference in two dimensions. Vertically, Xenos develops operator linking technique to improve data locality by restructuring the inter-operator dataflow. Horizontally, Xenos develops DSP-aware operator split technique to enable higher parallelism across multiple DSP units. Our evaluation demonstrates the effectiveness of vertical and horizontal dataflow optimization, which reduce the inference time by 15.0%–84.9% and 17.9%–89.9% , respectively. Besides, Xenos also outperforms the widely-used TVM by 1.1×–1.9×. Moreover, we extend Xenos to a distributed solution, which we call d-Xenos. d-Xenos employs multiple edge devices to jointly conduct the inference task and achieves a speedup of 3.68×–3.78× compared with the single device.

A low-latency graph computer to identify metastable particles at the Large Hadron Collider for real-time analysis of potential dark matter signatures

Image recognition is a pervasive task in many information-processing environments. We present a solution to a difficult pattern recognition problem that lies at the heart of experimental particle physics. Future experiments with very high-intensity beams will produce a spray of thousands of particles in each beam-target or beam-beam collision. Recognizing the trajectories of these particles as they traverse layers of electronic sensors is a massive image recognition task that has never been accomplished in real time. We present a real-time processing solution that is implemented in a commercial field-programmable gate array using high-level synthesis. It is an unsupervised learning algorithm that uses techniques of graph computing. A prime application is the low-latency analysis of dark-matter signatures involving metastable charged particles that manifest as disappearing tracks.

Physics-constrained robust learning of open-form partial differential equations from limited and noisy data

Unveiling the underlying governing equations of nonlinear dynamic systems remains a significant challenge. Insufficient prior knowledge hinders the determination of an accurate candidate library, while noisy observations lead to imprecise evaluations, which in turn result in redundant function terms or erroneous equations. This study proposes a framework to robustly uncover open-form partial differential equations (PDEs) from limited and noisy data. The framework operates through two alternating update processes: discovering and embedding. The discovering phase employs symbolic representation and a novel reinforcement learning (RL)-guided hybrid PDE generator to efficiently produce diverse open-form PDEs with tree structures. A neural network-based predictive model fits the system response and serves as the reward evaluator for the generated PDEs. PDEs with higher rewards are utilized to iteratively optimize the generator via the RL strategy and the best-performing PDE is selected by a parameter-free stability metric. The embedding phase integrates the initially identified PDE from the discovering process as a physical constraint into the predictive model for robust training. The traversal of PDE trees automates the construction of the computational graph and the embedding process without human intervention. Numerical experiments demonstrate our framework's capability to uncover governing equations from nonlinear dynamic systems with limited and highly noisy data and outperform other physics-informed neural network-based discovery methods. This work opens new potential for exploring real-world systems with limited understanding.

Multimodal surface wave inversion with automatic differentiation

SUMMARY Investigating subsurface shear wave velocity (vs) structures using surface wave dispersion data involves minimizing a misfit function that is commonly solved through gradient-based optimization. Sensitivity kernels for model updates are commonly estimated using numerical differentiation, variational methods or implicit functions which however, may involve numerical instability and computational challenges when dealing with complex velocity models and large data sets. In this study, we propose a novel surface wave inversion framework in which error-free gradients are calculated by automatic differentiation (AD) and forward modelling is implemented by convenient computational graphs in the state-of-the-art deep learning framework. The AD-based inversion approach is first validated using two synthetic data sets. Then, the subsurface structures at three distinct locations, namely the Great Plains and the Long Beach in the US and Tong Zhou in China, are also derived using this method with seismic ambient noise data, which show nice consistency with those obtained using traditional methods. With the significantly improved computational efficiency, a great number of initial models can be inverted simultaneously to mitigate the impact of local minima and to estimate the uncertainty in the invert models. We have developed a new surface wave inversion package named ADsurf based on automatic differentiation and computational graphs in the deep learning framework, and its computational efficiency is also compared with the traditional finite-difference-based gradient estimation approach. While a great number of intriguing studies on the geophysical inverse problems have been conducted recently using deep learning for end-to-end mapping, the use of AD provided in the in the deep learning frameworks to assist and expedite the gradient computations are still underexploited in geophysics. Thus, it is expected that various geophysical inverse problems in many different areas beyond the surface wave inversion can also be tackled with this new paradigm in the future.

Point cloud registration method for indoor depth sensor acquisition system based on dual graph computation with irregular shape factors

AbstractThe registration performance determines the widespread indoor application of 3D models acquired by depth sensors. Many advanced registration methods lack comprehensive feature aggregation and poor generalization capabilities, which improves the mismatching ratio. Here, a dual graph network is proposed by incorporating irregular shape factors to make point cloud features more expressive. At first, we transform point cloud sets into the stellar graph within the local neighbourhood of each point. The deep feature and shape factor of each point are combined in the directional‐connected irregular projection space. Subsequently, the combined features are modelled as the second graph. By the attention mechanism computation, feature information is continuously aggregated with intra‐graph and inter‐graph. Finally, a loss function is utilized to confirm point correspondence and perform the registration through singular value decomposition. Extensive experiments validate that the proposed point cloud registration method achieves state‐of‐the‐art performance.

Distributed matrix computing system for big data

In order to solve the problem of low computing efficiency in big data analysis and model construction, this paper intended to deeply explore the big data analysis programming model, DAG (Directed Acyclic Graph) and other contents, and on this basis, it adopted a distributed matrix computing system Octopus for big data analysis. Octopus is a universal matrix programming framework that provides a programming model based on matrix operations, which can conveniently analyze and process large-scale data. By using Octopus, users can extract functions and data from multiple platforms and operate through a unified matrix operation interface. The distributed matrix representation and storage layer can design data storage formats for distributed file systems. Each computing platform in OctMatrix provides its own matrix library, and it provides a matrix library written in R language for the above users. SymboMatrix provides a matrix interface to OctMatrix that is consistent with OctMatrix. However, SymboMatrix also retains the flow diagram for matrix operations in the process, and it also supports logical and physical optimization of the flow diagram on a DAG. For the DAG computational flow graph generated by SymbolMatrix, this paper divided it into two parts: logical optimization and physical optimization. This paper adopted a distributed file system based on line matrix, and obtained the corresponding platform matrix by reading the documents based on line matrix. In the evaluation of system performance, it was found that the distributed matrix computing system had a high computing efficiency, and the average CPU (central processing unit) usage reached 70%. This system can make full use of computing resources and realize efficient parallel computing.

Optimizing Job Offloading Schedule for Collaborative DNN Inference

Deep Neural Networks (DNNs) have been widely deployed in mobile applications. DNN inference latency is a critical metric to measure the service quality of those applications. Collaborative inference is a promising approach for latency optimization, where partial inference workloads are offloaded from mobile devices to cloud servers. Model partition problems for collaborative inference have been well studied. However, little attention has been paid to optimizing offloading pipeline for multiple DNN inference jobs. In practice, mobile devices usually need to process multiple DNN inference jobs simultaneously. We propose to jointly optimize the DNN partitioning and pipeline scheduling for multiple inference jobs. We theoretically analyze the optimal scheduling conditions for homogeneous chain-structure DNNs. Based on the analysis, we proposed near-optimal partitioning and scheduling methods for chain-structure DNNs. We also extend those methods for general-structure DNNs. In addition, we extend our problem scenario to handle heterogeneous DNN inference jobs. A layer-level scheduling algorithm is proposed. Theoretical analyses show that our proposed method is optimal when computation graphs are tree-structure. Our joint optimization methods are evaluated in a real-world testbed. Experiment results show that our methods can significantly reduce the overall inference latency of multiple inference jobs compared to partition-only or schedule-only approaches.

Исследование гранулометрического состава измельченного продукта, полученого при измельчении доломита в центробежной мельнице вертикального типа

Introduction. Modern medium (10–200 microns) and fine (0.1–10.0 microns) powders have a high cost, which limits the scale of their production and use. This assumes the implementation the latest achievements in the field of creating new equipment that is not inferior in its technical characteristics to the best foreign models into industry. The use centrifugal mills of vertical type for grinding various raw materials has shown their high efficiency in obtaining materials in the range – 0.2–0.01 mm. Studies of these mills operation under various operating conditions made it possible to create a large data set on the dependence of the crushed material granulo-metric composition on the design features of the mills and technological parameters of the grinding process. The results of a previously study of the granulo-metric composition a finely dispersed crushed product integral distribution according to particle size classes less than 80 microns were presented. They were obtained in a Brookhaven BI-XDC X-ray analytical centrifuge by measuring particle sizes in natural sedimentation mode, when the centrifuge disk was stationary during the experiment. The studies of the particle size differential distribution of a finely dispersed crushed product according to particle size classes less than 80 microns, which have not been carried out to date are of particular interest. Purpose of the study. The differential distribution study of the finely dispersed granulo-metric product composition obtained by grinding lump dolomite in a centrifugal mills of vertical type under different operating modes. Materials and methods. To solve this problem, samples of crushed material were taken from grinding dolomite from the Bosninsky deposit (RNO-Alania) in a centrifugal mill of the vertical type MV-0.3 with a 300 mm rotor diameter with following technological modes of the grinding process: mill rotor speed n = 310; 360 min–1, height of the material column above the rotating rotor H = 180; 270 mm, test time – 6 hours. Selected samples with a size class of –0.08 mm were analyzed in an X-ray analytical centrifuge Brookhaven BI-XDC, designed to determine the particle size distribution in the range from 0.01 to 100 microns. The measurements were performed on a suspension prepared from a mixture of 0.5 g of the selected sample with a 20% solution of sucrose and distilled water. Samples with a size class of 0.08 mm were weighed on laboratory scales LV SP.005.048 RE. The resulting suspension was dispersed in an ultrasonic bath for 3–5 minutes. After dispersion, the suspension was poured into the analyzer disk. While stirring, the lower absorption baseline line of a homogeneous particle suspension was measured. The obtained data from the analysis of the fine fractions size of were processed by the centrifuge computer software and graphs of the differential distribution of particles were constructed for various technological modes of grinding dolomite in a centrifugal mill. During the processing of the experiments, minor bursts in the concentration of the crushed product were not taken into account. The results and discussion. For the first time, a differential analysis of the granulo-metric composition of finely dispersed material with a particle size of –0.08 mm, obtained by grinding lump dolomite in a centrifugal mill of the vertical type under different mill operating modes, was carried out. Under all investigated operating modes of the centrifugal mill, the main part of the fine fraction in the mill outlet was in the range from 1.197 to 4.643 μm, and in steady-state operation it varied from 1.310 to 4.601 μm. The range of fine particle sizes has maximum values when the height of the column of crushed material above the rotating mill rotor and the rotation speed of the mill rotor decrease, and minimum values are observed when the height of the column of crushed material above the rotating mill rotor increases and the rotation speed of the mill rotor increases. Resume. 1. The article presents the results of the granulo-metric composition differential distribution study of a finely dispersed product obtained by grinding lump dolomite in a centrifugal mill of the vertical type. 2. The conducted studies have shown that, despite the positive results in obtaining fine particles with a size of 0.001 microns (nanoparticles) by mechanical means, the industrial production of nano-materials when grinding lump mineral raw materials in a vertical centrifugal mill is not economically profitable due to the low specific productivity and high energy intensity of the grinding process. 3. The obtained research results can be useful in the design of centrifugal mills.

Dynamic Merging for Optimal Onboard Resource Utilization: Innovating Mission Queue Constructing Method in Multi-Satellite Spatial Information Networks

The purpose of constructing onboard observation mission queues is to improve the execution efficiency of onboard tasks and reduce energy consumption, representing a significant challenge in achieving efficient global military reconnaissance and target tracking. Existing research often focuses on the aspect of task scheduling, aiming at optimizing the efficiency of single-task execution, while neglecting the complex dependencies that might exist between multiple tasks and payloads. Moreover, traditional task scheduling schemes are no longer suitable for large-scale tasks. To effectively reduce the number of tasks within the network, we introduce a network aggregation graph model based on multiple satellites and tasks, and propose a task aggregation priority dynamic calculation algorithm based on graph computations. Subsequently, we present a dynamic merging-based method for multi-satellite, multi-task aggregation, a novel approach for constructing onboard mission queues that can dynamically optimize the task queue according to real-time task demands and resource status. Simulation experiments demonstrate that, compared to baseline algorithms, our proposed task aggregation method significantly reduces the task size by approximately 25% and effectively increases the utilization rate of onboard resources.

Binary relations applied to the fuzzy substructures of quantales under rough environment

Abstract Binary relations (BIRs) have many applications in computer science, graph theory, and rough set theory. This study discusses the combination of BIRs, fuzzy substructures of quantale, and rough fuzzy sets. Approximation of fuzzy subsets of quantale with the help of BIRs is introduced. In quantale, compatible and complete relations in terms of aftersets and foresets with the help of BIRs are defined. Furthermore, we use compatible and complete relations to approximate fuzzy substructures of quantale, and these approximations are interpreted by aftersets and foresets. This concept generalizes the concept of rough fuzzy quantale. Finally, using BIRs, quantale homomorphism is used to build a relationship between the approximations of fuzzy substructures of quantale and the approximations of their homomorphic images.

Auto-Prox: Training-Free Vision Transformer Architecture Search via Automatic Proxy Discovery

The substantial success of Vision Transformer (ViT) in computer vision tasks is largely attributed to the architecture design. This underscores the necessity of efficient architecture search for designing better ViTs automatically. As training-based architecture search methods are computationally intensive, there’s a growing interest in training-free methods that use zero-cost proxies to score ViTs. However, existing training-free approaches require expert knowledge to manually design specific zero-cost proxies. Moreover, these zero-cost proxies exhibit limitations to generalize across diverse domains. In this paper, we introduce Auto-Prox, an automatic proxy discovery framework, to address the problem. First, we build the ViT-Bench-101, which involves different ViT candidates and their actual performance on multiple datasets. Utilizing ViT-Bench-101, we can evaluate zero-cost proxies based on their score-accuracy correlation. Then, we represent zero-cost proxies with computation graphs and organize the zero-cost proxy search space with ViT statistics and primitive operations. To discover generic zero-cost proxies, we propose a joint correlation metric to evolve and mutate different zero-cost proxy candidates. We introduce an elitism-preserve strategy for search efficiency to achieve a better trade-off between exploitation and exploration. Based on the discovered zero-cost proxy, we conduct a ViT architecture search in a training-free manner. Extensive experiments demonstrate that our method generalizes well to different datasets and achieves state-of-the-art results both in ranking correlation and final accuracy. Codes can be found at https://github.com/lilujunai/Auto-Prox-AAAI24.

Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training

Recurrent Neural Networks (RNNs) are useful in temporal sequence tasks. However, training RNNs involves dense matrix multiplications which require hardware that can support a large number of arithmetic operations and memory accesses. Implementing online training of RNNs on the edge calls for optimized algorithms for an efficient deployment on hardware. Inspired by the spiking neuron model, the Delta RNN exploits temporal sparsity during inference by skipping over the update of hidden states from those inactivated neurons whose change of activation across two timesteps is below a defined threshold. This work describes a training algorithm for Delta RNNs that exploits temporal sparsity in the backward propagation phase to reduce computational requirements for training on the edge. Due to the symmetric computation graphs of forward and backward propagation during training, the gradient computation of inactivated neurons can be skipped. Results show a reduction of ∼80% in matrix operations for training a 56k parameter Delta LSTM on the Fluent Speech Commands dataset with negligible accuracy loss. Logic simulations of a hardware accelerator designed for the training algorithm show 2-10X speedup in matrix computations for an activation sparsity range of 50%-90%. Additionally, we show that the proposed Delta RNN training will be useful for online incremental learning on edge devices with limited computing resources.

Memory-Efficient Reversible Spiking Neural Networks

Spiking neural networks (SNNs) are potential competitors to artificial neural networks (ANNs) due to their high energy-efficiency on neuromorphic hardware. However, SNNs are unfolded over simulation time steps during the training process. Thus, SNNs require much more memory than ANNs, which impedes the training of deeper SNN models. In this paper, we propose the reversible spiking neural network to reduce the memory cost of intermediate activations and membrane potentials during training. Firstly, we extend the reversible architecture along temporal dimension and propose the reversible spiking block, which can reconstruct the computational graph and recompute all intermediate variables in forward pass with a reverse process. On this basis, we adopt the state-of-the-art SNN models to the reversible variants, namely reversible spiking ResNet (RevSResNet) and reversible spiking transformer (RevSFormer). Through experiments on static and neuromorphic datasets, we demonstrate that the memory cost per image of our reversible SNNs does not increase with the network depth. On CIFAR10 and CIFAR100 datasets, our RevSResNet37 and RevSFormer-4-384 achieve comparable accuracies and consume 3.79x and 3.00x lower GPU memory per image than their counterparts with roughly identical model complexity and parameters. We believe that this work can unleash the memory constraints in SNN training and pave the way for training extremely large and deep SNNs.

SasWOT: Real-Time Semantic Segmentation Architecture Search WithOut Training

In this paper, we present SasWOT, the first training-free Semantic segmentation Architecture Search (SAS) framework via an auto-discovery proxy. Semantic segmentation is widely used in many real-time applications. For fast inference and memory efficiency, Previous SAS seeks the optimal segmenter by differentiable or RL Search. However, the significant computational costs of these training-based SAS limit their practical usage. To improve the search efficiency, we explore the training-free route but empirically observe that the existing zero-cost proxies designed on the classification task are sub-optimal on the segmentation benchmark. To address this challenge, we develop a customized proxy search framework for SAS tasks to augment its predictive capabilities. Specifically, we design the proxy search space based on the some observations: (1) different inputs of segmenter statistics can be well combined; (2) some basic operators can effectively improve the correlation. Thus, we build computational graphs with multiple statistics as inputs and different advanced basis arithmetic as the primary operations to represent candidate proxies. Then, we employ an evolutionary algorithm to crossover and mutate the superior candidates in the population based on correlation evaluation. Finally, based on the searched proxy, we perform the segmenter search without candidate training. In this way, SasWOT not only enables automated proxy optimization for SAS tasks but also achieves significant search acceleration before the retrain stage. Extensive experiments on Cityscapes and CamVid datasets demonstrate that SasWOT achieves superior trade-off between accuracy and speed over several state-of-the-art techniques. More remarkably, on Cityscapes dataset, SasWOT achieves the performance of 71.3% mIoU with the speed of 162 FPS.