Inference Framework Research Articles

Sustainability under Active Inference

In this paper, we explore the known connection among sustainability, resilience, and well-being within the framework of active inference. Initially, we revisit how the notions of well-being and resilience intersect within active inference before defining sustainability. We adopt a holistic concept of sustainability denoting the enduring capacity to meet needs over time without depleting crucial resources. It extends beyond material wealth to encompass community networks, labor, and knowledge. Using the free energy principle, we can emphasize the role of fostering resource renewal, harmonious system–entity exchanges, and practices that encourage self-organization and resilience as pathways to achieving sustainability both as an agent and as a part of a collective. We start by connecting active inference with well-being, building on existing work. We then attempt to link resilience with sustainability, asserting that resilience alone is insufficient for sustainable outcomes. While crucial for absorbing shocks and stresses, resilience must be intrinsically linked with sustainability to ensure that adaptive capacities do not merely perpetuate existing vulnerabilities. Rather, it should facilitate transformative processes that address the root causes of unsustainability. Sustainability, therefore, must manifest across extended timescales and all system strata, from individual components to the broader system, to uphold ecological integrity, economic stability, and social well-being. We explain how sustainability manifests at the level of an agent and then at the level of collectives and systems. To model and quantify the interdependencies between resources and their impact on overall system sustainability, we introduce the application of network theory and dynamical systems theory. We emphasize the optimization of precision or learning rates through the active inference framework, advocating for an approach that fosters the elastic and plastic resilience necessary for long-term sustainability and abundance.

Bayesian Estimation of Inverted Beta Mixture Models With Extended Stochastic Variational Inference for Positive Vector Classification.

The finite inverted beta mixture model (IBMM) has been proven to be efficient in modeling positive vectors. Under the traditional variational inference framework, the critical challenge in Bayesian estimation of the IBMM is that the computational cost of performing inference with large datasets is prohibitively expensive, which often limits the use of Bayesian approaches to small datasets. An efficient alternative provided by the recently proposed stochastic variational inference (SVI) framework allows for efficient inference on large datasets. Nevertheless, when using the SVI framework to address the non-Gaussian statistical models, the evidence lower bound (ELBO) cannot be explicitly calculated due to the intractable moment computation. Therefore, the algorithm under the SVI framework cannot directly use stochastic optimization to optimize the ELBO, and an analytically tractable solution cannot be derived. To address this problem, we propose an extended version of the SVI framework with more flexibility, namely, the extended SVI (ESVI) framework. This framework can be used in many non-Gaussian statistical models. First, some approximation strategies are applied to further lower the ELBO to avoid intractable moment calculations. Then, stochastic optimization with noisy natural gradients is used to optimize the lower bound. The excellent performance and effectiveness of the proposed method are verified in real data evaluation.

Online causal inference with application to near real-time post-market vaccine safety surveillance.

Streaming data routinely generated by social networks, mobile or web applications, e-commerce, and electronic health records present new opportunities to monitor the impact of an intervention on an outcome via causal inference methods. However, most existing causal inference methods have been focused on and applied to static data, that is, a fixed data set in which observations are pooled and stored before performing statistical analysis. There is thus a pressing need to turn static causal inference into online causal learning to support near real-time monitoring of treatment effects. In this paper, we present a framework for online estimation and inference of treatment effects that can incorporate new information as it becomes available without revisiting prior observations. We show that, under mild regularity conditions, the proposed online estimator is asymptotically equivalent to the offline oracle estimator obtained by pooling all data. Our proposal is motivated by the need for near real-time vaccine effectiveness and safety monitoring, and our proposed method is applied to a case study on COVID-19 vaccine safety surveillance.

Multidimensional Data Processing With Bayesian Inference via Structural Block Decomposition.

How to handle large multidimensional datasets, such as hyperspectral images and video information, efficiently and effectively plays a critical role in big-data processing. The characteristics of low-rank tensor decomposition in recent years demonstrate the essentials in describing the tensor rank, which often leads to promising approaches. However, most current tensor decomposition models consider the rank-1 component simply to be the vector outer product, which may not fully capture the correlated spatial information effectively for large-scale and high-order multidimensional datasets. In this article, we develop a new novel tensor decomposition model by extending it to the matrix outer product or called Bhattacharya-Mesner product, to form an effective dataset decomposition. The fundamental idea is to decompose tensors structurally in a compact manner as much as possible while retaining data spatial characteristics in a tractable way. By incorporating the framework of the Bayesian inference, a new tensor decomposition model on the subtle matrix unfolding outer product is established for both tensor completion and robust principal component analysis problems, including hyperspectral image completion and denoising, traffic data imputation, and video background subtraction. Numerical experiments on real-world datasets demonstrate the highly desirable effectiveness of the proposed approach.

Lyα Emission at z = 7–13: Clear Evolution of Lyα Equivalent Width Indicating a Late Cosmic Reionization History

We present the evolution of Lyα emission derived from 53 galaxies at z = 6.6–13.2, which have been identified by multiple JWST/NIRSpec spectroscopy programs of Early Release Science, General Observer, Director's Discretionary Time, and Guaranteed Time Observations. These galaxies fall on the star formation main sequence and are typical star-forming galaxies with UV magnitudes of −22.5 ≤ M UV ≤ −17.0. We find that 15 out of 53 galaxies show Lyα emission at the >3σ level, and we obtain Lyα equivalent width (EW) measurements and stringent 3σ upper limits for the 15 and 38 galaxies, respectively. Confirming that Lyα velocity offsets and line widths of our galaxies are comparable to those of low-redshift Lyα emitters, we investigate the redshift evolution of the Lyα EW. We find that Lyα EWs statistically decrease toward high redshifts on the Lyα EW versus the M UV plane for various probability distributions of the uncertainties. We then evaluate neutral hydrogen fractions x H I with the redshift evolution of the Lyα EW and the cosmic reionization simulation results on the basis of a Bayesian inference framework, and obtain x H I < 0.79, = 0.62 − 0.36 + 0.15 ?> , and 0.93 − 0.07 + 0.04 ?> at z ∼ 7, 8, and 9–13, respectively. These moderately large x H I values are consistent with the Planck cosmic microwave background optical depth measurement and previous x H I constraints from galaxy and QSO Lyα damping wing absorption and strongly indicate a late reionization history. Such a late reionization history suggests that major sources of reionization would emerge late and be hosted by moderately massive halos compared with the widely accepted picture of abundant low-mass objects for the sources of reionization.

Projections of Single-level and Multilevel Spinal Instrumentation Procedure Volume and Associated Costs for Medicare Patients to 2050.

Instrumented spinal fusions can be used in the treatment of vertebral fractures, spinal instability, and scoliosis or kyphosis. Construct-level selection has notable implications on postoperative recovery, alignment, and mobility. This study sought to project future trends in the implementation rates and associated costs of single-level versus multilevel instrumentation procedures in US Medicare patients aged older than 65 years in the United States. Data were acquired from the Centers for Medicare & Medicaid Services from January 1, 2000, to December 31, 2019. Procedure costs and counts were abstracted using Current Procedural Terminology codes to identify spinal level involvement. The Prophet machine learning algorithm was used, using a Bayesian Inference framework, to generate point forecasts for 2020 to 2050 and 95% forecast intervals (FIs). Sensitivity analyses were done by comparing projections from linear, log-linear, Poisson and negative-binomial, and autoregressive integrated moving average models. Costs were adjusted for inflation using the 2019 US Bureau of Labor Statistics' Consumer Price Index. Between 2000 and 2019, the annual spinal instrumentation volume increased by 776% (from 7,342 to 64,350 cases) for single level, by 329% (from 20,319 to 87,253 cases) for two-four levels, by 1049% (from 1,218 to 14,000 cases) for five-seven levels, and by 739% (from 193 to 1,620 cases) for eight-twelve levels (P < 0.0001). The inflation-adjusted reimbursement for single-level instrumentation procedures decreased 45.6% from $1,148.15 to $788.62 between 2000 and 2019, which is markedly lower than for other prevalent orthopaedic procedures: total shoulder arthroplasty (-23.1%), total hip arthroplasty (-39.2%), and total knee arthroplasty (-42.4%). By 2050, the number of single-level spinal instrumentation procedures performed yearly is projected to be 124,061 (95% FI, 87,027 to 142,907), with associated costs of $93,900,672 (95% FI, $80,281,788 to $108,220,932). The number of single-level instrumentation procedures is projected to double by 2050, while the number of two-four level procedures will double by 2040. These projections offer a measurable basis for resource allocation and procedural distribution.

Damage to tropical forests caused by cyclones is driven by wind speed but mediated by topographical exposure and tree characteristics.

Each year, an average of 45 tropical cyclones affect coastal areas and potentially impact forests. The proportion of the most intense cyclones has increased over the past four decades and is predicted to continue to do so. Yet, it remains uncertain how topographical exposure and tree characteristics can mediate the damage caused by increasing wind speed. Here, we compiled empirical data on the damage caused by 11 cyclones occurring over the past 40 years, from 74 forest plots representing tropical regions worldwide, encompassing field data for 22,176 trees and 815 species. We reconstructed the wind structure of those tropical cyclones to estimate the maximum sustained wind speed (MSW) and wind direction at the studied plots. Then, we used a causal inference framework combined with Bayesian generalised linear mixed models to understand and quantify the causal effects of MSW, topographical exposure to wind (EXP), tree size (DBH) and species wood density (ρ) on the proportion of damaged trees at the community level, and on the probability of snapping or uprooting at the tree level. The probability of snapping or uprooting at the tree level and, hence, the proportion of damaged trees at the community level, increased with increasing MSW, and with increasing EXP accentuating the damaging effects of cyclones, in particular at higher wind speeds. Higher ρ decreased the probability of snapping and to a lesser extent of uprooting. Larger trees tended to have lower probabilities of snapping but increased probabilities of uprooting. Importantly, the effect of ρ decreasing the probabilities of snapping was more marked for smaller than larger trees and was further accentuated at higher MSW. Our work emphasises how local topography, tree size and species wood density together mediate cyclone damage to tropical forests, facilitating better predictions of the impacts of such disturbances in an increasingly windier world.

Does the effect of adolescent health behaviours on adult cardiometabolic health differ by socioeconomic background? Protocol for a population-based cohort study

IntroductionAdolescence is a sensitive period for cardiometabolic health. Yet, it remains unknown if adolescent health behaviours, such as alcohol use, smoking, diet and physical activity, have differential effects across socioeconomic...

Bayesian simultaneous factorization and prediction using multi-omic data

Integrative factorization methods for multi-omic data estimate factors explaining biological variation. Factors can be treated as covariates to predict an outcome and the factorization can be used to impute missing values. However, no available methods provide a comprehensive framework for statistical inference and uncertainty quantification for these tasks. A novel framework, Bayesian Simultaneous Factorization (BSF), is proposed to decompose multi-omics variation into joint and individual structures simultaneously within a probabilistic framework. BSF uses conjugate normal priors and the posterior mode of this model can be estimated by solving a structured nuclear norm-penalized objective that also achieves rank selection and motivates the choice of hyperparameters. BSF is then extended to simultaneously predict a continuous or binary phenotype while estimating latent factors, termed Bayesian Simultaneous Factorization and Prediction (BSFP). BSF and BSFP accommodate concurrent imputation, i.e., imputation during the model-fitting process, and full posterior inference for missing data, including “blockwise” missingness. It is shown via simulation that BSFP is competitive in recovering latent variation structure, and demonstrate the importance of accounting for uncertainty in the estimated factorization within the predictive model. The imputation performance of BSF is examined via simulation under missing-at-random and missing-not-at-random assumptions. Finally, BSFP is used to predict lung function based on the bronchoalveolar lavage metabolome and proteome from a study of HIV-associated obstructive lung disease, revealing multi-omic patterns related to lung function decline and a cluster of patients with obstructive lung disease driven by shared metabolomic and proteomic abundance patterns.

Association between TB delay and TB treatment outcomes in HIV-TB co-infected patients: a study based on the multilevel propensity score method.

HIV-tuberculosis (HIV-TB) co-infection is a significant public health concern worldwide. TB delay, consisting of patient delay, diagnostic delay, treatment delay, increases the risk of adverse anti-TB treatment (ATT) outcomes. Except for individual level variables, differences in regional levels have been shown to impact the ATT outcomes. However, few studies appropriately considered possible individual and regional level confounding variables. In this study, we aimed to assess the association of TB delay on treatment outcomes in HIV-TB co-infected patients in Liangshan Yi Autonomous Prefecture (Liangshan Prefecture) of China, using a causal inference framework while taking into account individual and regional level factors. We conducted a study to analyze data from 2068 patients with HIV-TB co-infection in Liangshan Prefecture from 2019 to 2022. To address potential confounding bias, we used a causal directed acyclic graph (DAG) to select appropriate confounding variables. Further, we controlled for these confounders through multilevel propensity score and inverse probability weighting (IPW). The successful rate of ATT for patients with HIV-TB co-infection in Liangshan Prefecture was 91.2%. Total delay (OR = 1.411, 95% CI: 1.015, 1.962), diagnostic delay (OR = 1.778, 95% CI: 1.261, 2.508), treatment delay (OR = 1.749, 95% CI: 1.146, 2.668) and health system delay (OR = 1.480 95% CI: (1.035, 2.118) were identified as risk factors for successful ATT outcome. Sensitivity analysis demonstrated the robustness of these findings. HIV-TB co-infection prevention and control policy in Liangshan Prefecture should prioritize early treatment for diagnosed HIV-TB co-infected patients. It is urgent to improve the health system in Liangshan Prefecture to reduce delays in diagnosis and treatment.

Multilevel Latent Differential Structural Equation Model with Short Time Series and Time-Varying Covariates: A Comparison of Frequentist and Bayesian Estimators

Continuous-time modeling using differential equations is a promising technique to model change processes with longitudinal data. Among ways to fit this model, the Latent Differential Structural Equation Modeling (LDSEM) approach defines latent derivative variables within a structural equation modeling (SEM) framework, thereby allowing researchers to leverage advantages of the SEM framework for model building, estimation, inference, and comparison purposes. Still, a few issues remain unresolved, including performance of multilevel variations of the LDSEM under short time lengths (e.g., 14 time points), particularly when coupled multivariate processes and time-varying covariates are involved. Additionally, the possibility of using Bayesian estimation to facilitate the estimation of multilevel LDSEM (M-LDSEM) models with complex and higher-dimensional random effect structures has not been investigated. We present a series of Monte Carlo simulations to evaluate three possible approaches to fitting M-LDSEM, including: frequentist single-level and two-level robust estimators and Bayesian two-level estimator. Our findings suggested that the Bayesian approach outperformed other frequentist approaches. The effects of time-varying covariates are well recovered, and coupling parameters are the least biased especially using higher-order derivative information with the Bayesian estimator. Finally, an empirical example is provided to show the applicability of the approach.

Post-selection Inference in Multiverse Analysis (PIMA): An Inferential Framework Based on the Sign Flipping Score Test.

When analyzing data, researchers make some choices that are either arbitrary, based on subjective beliefs about the data-generating process, or for which equally justifiable alternative choices could have been made. This wide range of data-analytic choices can be abused and has been one of the underlying causes of the replication crisis in several fields. Recently, the introduction of multiverse analysis provides researchers with a method to evaluate the stability of the results across reasonable choices that could be made when analyzing data. Multiverse analysis is confined to a descriptive role, lacking a proper and comprehensive inferential procedure. Recently, specification curve analysis adds an inferential procedure to multiverse analysis, but this approach is limited to simple cases related to the linear model, and only allows researchers to infer whether at least one specification rejects the null hypothesis, but not which specifications should be selected. In this paper, we present a Post-selection Inference approach to Multiverse Analysis (PIMA) which is a flexible and general inferential approach that considers for all possible models, i.e., the multiverse of reasonable analyses. The approach allows for a wide range of data specifications (i.e., preprocessing) and any generalized linear model; it allows testing the null hypothesis that a given predictor is not associated with the outcome, by combining information from all reasonable models of multiverse analysis, and provides strong control of the family-wise error rate allowing researchers to claim that the null hypothesis can be rejected for any specification that shows a significant effect. The inferential proposal is based on a conditional resampling procedure. We formally prove that the Type I error rate is controlled, and compute the statistical power of the test through a simulation study. Finally, we apply the PIMA procedure to the analysis of a real dataset on the self-reported hesitancy for the COronaVIrus Disease 2019 (COVID-19) vaccine before and after the 2020 lockdown in Italy. We conclude with practical recommendations to be considered when implementing the proposed procedure.

Altered topological structure of the brain white matter in maltreated children through topological data analysis.

Childhood maltreatment may adversely affect brain development and consequently influence behavioral, emotional, and psychological patterns during adulthood. In this study, we propose an analytical pipeline for modeling the altered topological structure of brain white matter in maltreated and typically developing children. We perform topological data analysis (TDA) to assess the alteration in the global topology of the brain white matter structural covariance network among children. We use persistent homology, an algebraic technique in TDA, to analyze topological features in the brain covariance networks constructed from structural magnetic resonance imaging and diffusion tensor imaging. We develop a novel framework for statistical inference based on the Wasserstein distance to assess the significance of the observed topological differences. Using these methods in comparing maltreated children with a typically developing control group, we find that maltreatment may increase homogeneity in white matter structures and thus induce higher correlations in the structural covariance; this is reflected in the topological profile. Our findings strongly suggest that TDA can be a valuable framework to model altered topological structures of the brain. The MATLAB codes and processed data used in this study can be found at https://github.com/laplcebeltrami/maltreated.

Modelling transmission dynamics and genomic diversity in a recombining parasite population

The genomic diversity of a parasite population is shaped by its transmission dynamics but superinfection, cotranmission and recombination make this relationship complex and hard to analyse. This paper aims to simplify the problem by introducing the concept of a genomic transmission graph with three basic parameters: the effective number of hosts, the quantum of transmission and the crossing rate of transmission chains. This enables rapid simulation of coalescence times in a recombining parasite population with superinfection and cotransmission, and it also provides a mathematical framework for analysis of within-host variation. Taking malaria as an example, we use this theoretical model to examine how transmission dynamics and migration affect parasite genomic diversity, including the effective recombination rate and haplotypic metrics of recent common ancestry. We show how key transmission parameters can be inferred from deep sequencing data and as a proof of concept we estimate the Plasmodium falciparum transmission bottleneck. Finally we discuss the potential applications of this novel inferential framework in genomic surveillance for malaria control and elimination. Online tools for exploring the genomic transmission graph are available at d-kwiat.github.io/gtg.

An Overview of the Free Energy Principle and Related Research.

The free energy principle and its corollary, the active inference framework, serve as theoretical foundations in the domain of neuroscience, explaining the genesis of intelligent behavior. This principle states that the processes of perception, learning, and decision making-within an agent-are all driven by the objective of "minimizing free energy," evincing the following behaviors: learning and employing a generative model of the environment to interpret observations, thereby achieving perception, and selecting actions to maintain a stable preferred state and minimize the uncertainty about the environment, thereby achieving decision making. This fundamental principle can be used to explain how the brain processes perceptual information, learns about the environment, and selects actions. Two pivotal tenets are that the agent employs a generative model for perception and planning and that interaction with the world (and other agents) enhances the performance of the generative model and augments perception. With the evolution of control theory and deep learning tools, agents based on the FEP have been instantiated in various ways across different domains, guiding the design of a multitude of generative models and decision-making algorithms. This letter first introduces the basic concepts of the FEP, followed by its historical development and connections with other theories of intelligence, and then delves into the specific application of the FEP to perception and decision making, encompassing both low-dimensional simple situations and high-dimensional complex situations. It compares the FEP with model-based reinforcement learning to show that the FEP provides a better objective function. We illustrate this using numerical studies of Dreamer3 by adding expected information gain into the standard objective function. In a complementary fashion, existing reinforcement learning, and deep learning algorithms can also help implement the FEP-based agents. Finally, we discuss the various capabilities that agents need to possess in complex environments and state that the FEP can aid agents in acquiring these capabilities.

Classification and estimation of case-mix adjusted performance indices for binary outcomes

AbstractIn this paper, we propose a general class of indices that can be used for comparing the performances of organizations providing a given public service to citizens, such as universities, hospitals, nursing homes, employment agencies or other institutions. In particular, we handle the case where evaluation is performed by assessing the probability that a given event has happened as a result of the service provided to users requiring it. Indices are designed for settings where users can be divided into groups with similar characteristics in order to account for case-mix, that is, for the different composition of users within each organization with respect to personal features influencing the probability of the event at hand. For the proposed class, we build a taxonomy leading to nine index types. These different types constitute a useful toolbox to satisfy specific needs and/or criteria set by the evaluator in applied contexts. A general inferential framework is also discussed to deal with settings where, whatever the index chosen, its value has to be estimated from sample data. A simulation study based on a real-world dataset is presented to assess the behavior of indices’ estimators.

Design of an Iterative Method for Enhanced Multimodal Time Series Analysis Using Graph Attention Networks, Variational Graph Autoencoders, and Transfer Learning

In the ever-evolving landscape of data analysis, the need to efficiently and accurately interpret multimodal time series data has become paramount. Traditional methods often fall short in addressing the complex dependencies and dynamics inherent in such data, limiting their effectiveness in real-world applications. This work introduces a comprehensive approach that leverages Graph Attention Networks (GATs), Variational Graph Autoencoders (VGAEs), transfer learning with pretrained transformers, and Bayesian state-space models to overcome these limitations. GATs are selected for their ability to dynamically focus on relevant modalities through attention mechanisms, thereby capturing the intricate relationships between different data modalities. This method significantly enhances the model's ability to integrate multimodal information, leading to notable improvements in classification, prediction, and anomaly detection tasks. VGAEs are utilized to learn latent representations within a graph-based framework, promoting unsupervised learning while unveiling the underlying data structure. The resultant embeddings are pivotal for downstream tasks like clustering and visualization, encapsulating the interactions within multimodal time series data effectively. Furthermore, this work incorporates transfer learning with pretrained transformers to harness extensive knowledge from large datasets, adapting it to multimodal time series analysis. This strategy excels in capturing long-range dependencies, thereby augmenting generalization and performance in data-scarce scenarios. Bayesian state-space models are employed to elucidate the temporal dynamics and uncertainties of time series data, offering a robust framework for probabilistic inference and enhancing the interpretability and reliability of forecasting and anomaly detection. The efficacy of the proposed model is rigorously evaluated using diverse datasets, including the Yahoo! Stock Dataset, Forest Cover Dataset, and an empirical collection of 100k time series data samples. The results demonstrate a significant leap in performance metrics, including a 9.5% increase in precision, 8.5% boost in accuracy, 8.3% rise in recall, 10.4% reduction in delay, 9.4% enhancement in AUC, and a 5.9% improvement in specificity, alongside superior pre-emption capabilities compared to existing methods. This work not only addresses the pressing need for advanced multimodal time series analysis techniques but also sets a new benchmark for efficiency and accuracy. The integration of GATs, VGAEs, transfer learning with pretrained transformers, and Bayesian state-space models presents a formidable approach that significantly advances the field, offering profound impacts on a wide array of applications.

Green’s matching: an efficient approach to parameter estimation in complex dynamic systems

Abstract Parameters of differential equations are essential to characterize intrinsic behaviours of dynamic systems. Numerous methods for estimating parameters in dynamic systems are computationally and/or statistically inadequate, especially for complex systems with general-order differential operators, such as motion dynamics. This article presents Green’s matching, a computationally tractable and statistically efficient two-step method, which only needs to approximate trajectories in dynamic systems but not their derivatives due to the inverse of differential operators by Green’s function. This yields a statistically optimal guarantee for parameter estimation in general-order equations, a feature not shared by existing methods, and provides an efficient framework for broad statistical inferences in complex dynamic systems.

Application of the target trial emulation framework to external comparator studies

External comparator (EC) studies are increasingly being used to generate evidence that supports the evaluation of emerging pharmacological treatments for regulatory and health technology assessment (HTA) purposes. However, the reliability of evidence generated from EC studies can vary. In this paper, we outline how an existing framework for causal inference, the target trial emulation (TTE) framework, can be appropriately applied to improve the design and analysis of EC studies. Applying the TTE framework involves specifying the protocol of an ideal target trial which would answer the causal question of interest, then emulating its key elements under real-world (RW) settings. We describe each component of the original TTE framework and explain how it can be applied to EC studies, supplemented with practical recommendations. We also highlight special considerations and limitations in applying the TTE framework to EC studies. We describe how the TTE framework can be applied to improve the clarity, transparency, and reliability of evidence generated from EC studies.

ARM-CO-UP: ARM CO operative U tilization of P rocessors

HMPSoCs combine different processors on a single chip. They enable powerful embedded devices, which increasingly perform ML inference tasks at the edge. State-of-the-art HMPSoCs can perform on-chip embedded inference using different processors, such as CPUs, GPUs, and NPUs. HMPSoCs can potentially overcome the limitation of low single-processor CNN inference performance and efficiency by cooperative use of multiple processors. However, standard inference frameworks for edge devices typically utilize only a single processor. We present the ARM-CO-UP framework built on the ARM-CL library. The ARM-CO-UP framework supports two modes of operation – Pipeline and Switch. It optimizes inference throughput using pipelined execution of network partitions for consecutive input frames in the Pipeline mode. It improves inference latency through layer-switched inference for a single input frame in the Switch mode. Furthermore, it supports layer-wise CPU/GPU DVFS in both modes for improving power efficiency and energy consumption. ARM-CO-UP is a comprehensive framework for multi-processor CNN inference that automates CNN partitioning and mapping, pipeline synchronization, processor type switching, layer-wise DVFS , and closed-source NPU integration.