We present hls4ml , a free and open source platform that translates machine learning (ML) models from modern deep learning frameworks into high-level synthesis (HLS) code that can be integrated into full designs for field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs). With its flexible and modular design, hls4ml supports a large number of deep learning frameworks and can target HLS compilers from several vendors, including Vitis HLS, Intel oneAPI and Catapult HLS. Together with a wider eco-system for software-hardware co-design, hls4ml has enabled the acceleration of ML inference in a wide range of commercial and scientific applications where low latency, resource usage, and power consumption are critical. In this article, we describe the structure and functionality of the hls4ml platform. The overarching design considerations for the generated HLS code are discussed, together with selected performance results.

Extreme data rate scientific experiments create massive amounts of data that require efficient ML edge processing. This leads to unique validation challenges for VLSI implementations of ML algorithms: enabling bit-accurate functional simulations for performance validation in experimental software frameworks, verifying those ML models are robust under extreme quantization and pruning, and enabling ultra-fine-grained model inspection for efficient fault tolerance. We discuss approaches to developing and validating reliable algorithms at the scientific edge under such strict latency, resource, power, and area requirements in extreme experimental environments. We study metrics for developing robust algorithms, present preliminary results and mitigation strategies, and conclude with an outlook of these and future directions of research towards the longer-term goal of developing autonomous scientific experimentation methods for accelerated scientific discovery.

We develop an end-to-end workflow for the training and implementation of co-designed neural networks (NNs) for efficient field-programmable gate array (FPGA) hardware. Our approach leverages Hessian-aware quantization of NNs, the Quantized Open Neural Network Exchange intermediate representation, and the hls4ml tool flow for transpiling NNs into FPGA firmware. This makes efficient NN implementations in hardware accessible to nonexperts in a single open sourced workflow that can be deployed for real-time machine-learning applications in a wide range of scientific and industrial settings. We demonstrate the workflow in a particle physics application involving trigger decisions that must operate at the 40-MHz collision rate of the CERN Large Hadron Collider (LHC). Given the high collision rate, all data processing must be implemented on FPGA hardware within the strict area and latency requirements. Based on these constraints, we implement an optimized mixed-precision NN classifier for high-momentum particle jets in simulated LHC proton-proton collisions.

The global throughput of an 802.11e WLAN is determined by EDCA (Enhanced Distributed Channel Access) parameters, among other aspects, that are usually configured with predetermined and static values. This study carefully evaluates the Quality of Service (QoS) of Wi-Fi with EDCA in several realistic scenarios with noise and a blend of wireless traffic (e.g., voice, video, and best effort, with Pareto distribution). The metrics of the benefits obtained in each case are compared, and the differentiated impact of network dynamics on each case is quantified. This study proposes a new experimental scenario based on the relative proportion of traffic present in the network. Stations have been implemented using HSANs (Hierarchical Stochastic Activity Networks) and simulated using the Möbius tool.

The time Petri net with firing frequency intervals (TPNF) is a modeling formalism used to specify system behavior under timing and frequency constraints. Efficient techniques exist to evaluate the performance of TPNF models based on the computation of bounds of performance metrics (e.g., transition throughput, place marking). In this paper, we propose a min-max problem to compute the cycle time of a transition under optimistic assumptions. That is, we are interested in computing the lower bound. We will demonstrate that such a problem is related to a maximization linear programming problem (LP-max) previously stated in the literature, to compute the throughput upper bound of the transition. The main advantage of the min-max problem compared to the LP-max is that, in addition to the optimal value, the optimal solutions provide useful feedback to the analyst on the system behavior (e.g., performance bottlenecks). We have implemented two solution algorithms, using CPLEX APIs, to solve the min-max problem, and have compared their performance using a benchmark of TPNF models, several of these being case studies. Finally, we have applied the min-max technique for the vulnerability analysis of a critical infrastructure, i.e., the Saudi Arabian crude-oil distribution network.

Software systems that do not meet their timing constraints can cause risks. In this work, we propose a comprehensive method for assessing the risk of timing failure by evaluating the software design. We show how to apply best practises in software engineering and well-known Time Petri Net (TPN) modeling and analysis techniques, and we demonstrate the effectiveness of the method with reference to a case study in the domain of real-time embedded systems. The method customizes the Australian standard risk management process, where the system context is the UML-based software specification, enriched with standard MARTE profile annotations to capture nonfunctional system properties. During the risk analysis, a TPN is derived, via model transformation, from the software design specification and TPN bound techniques are applied to estimate the probability of timing failure. TPN bound techniques are also exploited, within the risk evaluation and treatment steps, to identify the risk causes in the software design.

Time Petri nets (TPNs) have been widely used for the verification and validation of real-time systems during the software development process. Their quantitative analysis consists in applying enumerative techniques that suffer the well known state space explosion problem. To overcome this problem, several methods have been proposed in the literature, that either provide rules to obtain equivalent nets with a reduced state space or avoid the construction of the whole state space. In this paper, we propose a method that consists in computing performance bounds to predict the average operational behavior of TPNs by exploiting their structural properties and by applying operational laws. Performance bound computation was first proposed for timed (Timed PNs) and stochastic Petri nets (SPNs). We generalize the results obtained for Timed PNs and SPNs to make the technique applicable to TPNs and their extended stochastic versions: TPN with firing frequency intervals (TPNFs) and extended TPNs (XTPNs). Finally, we apply the proposed bounding techniques on the case study of a robot-control application taken from the literature.

A general iterative technique for approximate throughput computation of stochastic live and bounded weighted T-systems (WTS) is presented. It generalizes a previous technique on stochastic marked graphs. The approach has two basic foundations. First, a deep understanding of the qualitative behavior of WTS leads to a general decomposition technique. Second, after the decomposition phase, an iterative response-time approximation method is applied for the throughput computation. Existence of convergence points for the iterative approximation method can be proved. Experimental results generally have an error of less than 5%. The state space is usually reduced by more than one order of magnitude; therefore, the analysis of otherwise intractable systems is possible

In this paper we address the mobile robot localization using some techniques borrowed from the computational biology community. The specific problem studied here is also known as the kidnapped robot problem. Our proposal is to solve this problem by string matching algorithms, which have experienced a large advance in the last years due to (for example) the Genoma Project. The paper uses three different algorithms to solve the mentioned problem and shows their advantages, such as the robustness of the results and the memory and time efficiency. These results are validated by real experimentation using panoramic images of indoor buildings, and compared and discussed with existing techniques that have been used over the same test-bed.

Nowadays, there exist web sites that allow users to retrieve and install software in an easy way. The performance of these sites may be poor if they are used in wireless networks; the reason is the inadequate use of the net resources they need. If this kind of systems are designed using mobile agent technology the previous problem might be avoided. In this paper, we present a comparison between the performance of a software retrieval system especially designed to be used in wireless networks (e.g., mobile computers) and the performance of a software retrieval system similar to the well-known Tucows.com or Download.com web sites.

A dynamics inversion compensation scheme is designed for control of nonlinear discrete-time systems with input backlash. The compensator uses the backstepping technique with neural networks (NN) for inverting the backlash nonlinearity in the feedforward path. The technique provides a general procedure for using NN to determine the dynamics pre-inverse of an invertible discrete time dynamical system. A discrete-time tuning algorithm is given for the NN weights so that the backlash compensation scheme becomes adaptive, guaranteeing bounded tracking and backlash errors, and also bounded parameter estimates. A rigorous proof of stability and performance is given and a simulation example verifies the performance. Unlike standard discrete-time adaptive control techniques, no certainty equivalence assumption is needed.

A fuzzy logic (FL) compensator is designed for control of nonlinear discrete-time systems with input deadzone. The classification property of FL systems makes them a natural candidate for the rejection of errors induced by the deadzone, which has regions in which it behaves differently. A discrete-time tuning algorithm is given for the FL parameters so that the deadzone compensation scheme becomes adaptive, guaranteeing bounded tracking errors and parameter estimates. A rigorous proof of stability and performance is given and a simulation example verifies performance. Unlike standard discrete-time adaptive control techniques, no certainty equivalence assumption is needed.

Asynchronously communicating stochastic modules (SAM) are Petri nets that can be seen as a set of modules that communicate through buffers, so they are not (yet another) Petri net subclass, but they complement a net with a structured view. This paper considers the problem of exploiting the compositionality of the view to generate the state space and to find the steady-state probabilities of a stochastic extension of SAM in a net-driven, efficient way. Essentially we give an expression of an auxiliary matrix, G, which is a supermatrix of the infinitesimal generator of a SAM. G is a tensor algebra expression of matrices of the size of the components for which it is possible to numerically solve the characteristic steady-state solution equation /spl pi//spl middot/G=0, without the need to explicitly compute G. Therefore, we obtain a method that computes the steady-state solution of a SAM without ever explicitly computing and storing its infinitesimal generator, and therefore without computing and storing the reachability graph of the system. Some examples of application of the technique are presented and compared to previous approaches.

We concentrate on a family of discrete event systems obtained from a simple modular design principle that include in a controlled way primitives to deal with concurrency, decisions, synchronization, blocking, and bulk movements of jobs. Due to the functional complexity of such systems, reliable throughput approximation algorithms must be deeply supported on a structure based decomposition technique. We present a decomposition technique and a fixed-point search iterative process based on response time preservation of subsystems. Extensive numerical experiments have shown that the error is less than 3%, and that the state space is usually reduced by more than one order of magnitude.

A general iterative technique for approximate throughput computation of stochastic strongly connected marked graphs is presented. It generalizes a previous technique based on net decomposition through a single input-single output cut, allowing the split of the model through any cut. The approach has two basic foundations. First, a deep understanding of the qualitative behavior of marked graphs leads to a general decomposition technique. Second, after the decomposition phase, an iterative response time approximation method is applied for the computation of the throughput. Experimental results on several examples generally have an error of less than 3%. The state space is usually reduced by more than one order of magnitude; therefore, the analysis of otherwise intractable systems is possible.>

Ergodicity and throughput bound characterization are addressed for a subclass of timed and stochastic Petri nets, interleaving qualitative and quantitative theories. The nets considered represent an extension of the well-known subclass of marked graphs, defined as having a unique consistent firing count vector, independently of the stochastic interpretation of the net model. In particular, persistent and mono-T-semiflow net subclasses are considered. Upper and lower throughput bounds are computed using linear programming problems defined on the incidence matrix of the underlying net. The bounds proposed depend on the initial marking and the mean values of the delays but not on the probability distributions (thus including both the deterministic and the stochastic cases). From a different perspective, the considered subclasses of synchronized queuing networks; thus, the proposed bounds can be applied to these networks.>