Active Inference Research Articles

Metastable Oscillatory Modes as a Signature of Entropy Management in the Brain

Entropy management, central to the Free Energy Principle, requires a process that temporarily shifts brain activity toward states of lower or higher entropy. Metastable synchronization is a process by which a system achieves entropy fluctuations by intermittently transitioning between states of collective order and disorder. Previous work has shown that collective oscillations, similar to those recorded from the brain, emerge spontaneously from weakly stable synchronization in critically coupled oscillator systems. However, direct evidence linking the formation of collective oscillations to entropy fluctuations is lacking. In this short communication, we demonstrate how the emergence of Metastable Oscillatory Modes (MOMs) is directly associated with a temporary reduction in entropy in the ongoing dynamics. We apply Shannon entropy to the distribution of eigenvalues of phase covariance over sliding time windows, capturing the temporal evolution of entropy at the level of the entire dynamical system. By demonstrating how the formation of MOMs impacts a system’s entropy levels, we bridge theoretical works on the physics of coupled oscillators with the FEP framework, supporting the hypothesis that brain rhythms recorded experimentally are a signature of entropy management.

A hyperelastic constitutive model for soft elastomers considering the entanglement-dependent finite extensibility

In this paper, a novel hyperelastic constitutive model for soft elastomers is developed based on the concept of the tortuous tube. This model incorporates the finite extensibility of the polymer chain, the entanglement contribution to elasticity and the non-affine micro-to-macro scale transition in a unified way. To reflect the entanglement effect and its influence on the deformation of soft elastomers, the tortuous tube concept is introduced. The finite extensibility and conformational statistics of an entangled polymer chain in such a tortuous tube are clarified. By embedding the tortuous tube into the microsphere and employing the principle of minimum averaged free energy, a new non-affine scale transition rule is proposed to establish the relationship between the local deformation of a polymer chain at the microscopic scale and the overall deformation at the macroscopic scale. Based on the probability density function related to the conformational statistics, the Helmholtz free energy is established and further decoupled into a volumetric part and an isochoric part. The spatial Kirchhoff stress tensor and spatial elasticity tensor are derived from the newly established Helmholtz free energy. The proposed model is further implemented into the finite element program ABAQUS by writing a user-defined material subroutine. The prediction capability of the proposed model is verified by simulating the homogeneous and inhomogeneous deformations of soft elastomers under various loading modes, including uniaxial tension, uniaxial compression, pure shear, equi-biaxial tension, general biaxial tensile loadings, inflation and indentation. Moreover, the influence of entanglement concentration on the stretchability and stiffness of soft elastomers is predicted and discussed using the proposed model.

Affordances, mental health and psychiatry

This paper suggests that new directions in mental health and psychiatry are being driven by anomalies within epidemiology, mechanisms, and integration-based cognition. Three contrasting kinds of new work on mental disorder are critically discussed: enactive, ecological and active inference approaches, drawing from work by Varela, Gibson and Friston. These are detailed together with debates about their similarities and differences. All of these positions in fact make common use of ‘affordance’ as a core perception-action mechanism, suggesting promising options for new or enhanced psychiatric interventions. This leads to a critical review of the current state of knowledge on affordances, and their limitations, together with some suggestions as to how this approach might be applied in mental health and psychiatric practice.

PhosX: data-driven kinase activity inference from phosphoproteomics experiments.

The inference of kinase activity from phosphoproteomics data can point to causal mechanisms driving signalling processes and potential drug targets. Identifying the kinases whose change in activity explains the observed phosphorylation profiles, however, remains challenging, and constrained by the manually curated knowledge of kinase-substrate associations. Recently, experimentally determined substrate sequence specificities of human kinases have become available, but robust methods to exploit this new data for kinase activity inference are still missing. We present PhosX, a method to estimate differential kinase activity from phosphoproteomics data that combines state-of-the art statistics in enrichment analysis with kinases' substrate sequence specificity information. Using a large phosphoproteomics dataset with known differentially regulated kinases we show that our method identifies upregulated and downregulated kinases by only relying on the input phosphopeptides' sequences and intensity changes. We find that PhosX outperforms the currently available approach for the same task, and performs better or similarly to state-of-the-art methods that rely on previously known kinase-substrate associations. We therefore recommend its use for data-driven kinase activity inference. PhosX is implemented in Python, open-source under the Apache-2.0 licence, and distributed on the Python Package Index. The code is available on GitHub (https://github.com/alussana/phosx). Supplementary data are available at Bioinformatics online.

Integration of Sense and Control for Uncertain Systems Based on Delayed Feedback Active Inference.

Asa result of the time lag in transmission, the data obtained by the sensor is delayed and does not reflect the state at the current moment. The effects of input delay are often overlooked in active inference (AIF), which may lead to significant deviations in state estimation and increased prediction errors, particularly when the system is subjected to a sudden external stimulus. In this paper, a theoretical framework of delayed feedback active inference (DAIF) is proposed to enhance the applicability of AIF to real systems. The probability model of DAIF is defined by incorporating a control distribution into that of AIF. The free energy of DAIF is defined as the sum of the quadratic state, sense, and control prediction error. A predicted state derived from previous states is defined and introduced as the expectation of the prior distribution of the real-time state. A proportional-integral (PI)-like control based on the predicted state is taken to be the expectation of DAIF preference control, whose gain coefficient is inversely proportional to the measurement accuracy variance. To adaptively compensate for external disturbances, a second-order inverse variance accuracy replaces the fixed sensory accuracy of preference control. The simulation results of the trajectory tracking control of a quadrotor unmanned aerial vehicle (UAV) show that DAIF performs better than AIF in state estimation and disturbance resistance.

Safeguarding User-Centric Privacy in Smart Homes

Internet of Things (IoT) devices have been increasingly deployed in smart homes to automatically monitor and control their environments. Unfortunately, extensive recent research has shown that on-path external adversaries can infer and further fingerprint people’s sensitive private information by analyzing IoT network traffic traces. In addition, most recent approaches that aim to defend against these malicious IoT traffic analytics cannot adequately protect user privacy with reasonable traffic overhead. In particular, these approaches often did not consider practical traffic reshaping limitations, user daily routine permitting, and user privacy protection preference in their design. To address these issues, we design a new low-cost, open source user-centric defense system—PrivacyGuard—that enables people to regain the privacy leakage control of their IoT devices while still permitting sophisticated IoT data analytics that is necessary for smart home automation. In essence, our approach employs intelligent deep convolutional generative adversarial network assisted IoT device traffic signature learning, long short-term memory based artificial traffic signature injection, and partial traffic reshaping to obfuscate private information that can be observed in IoT device traffic traces. We evaluate PrivacyGuard using IoT network traffic traces of 31 IoT devices from five smart homes and buildings. We find that PrivacyGuard can effectively prevent a wide range of state-of-the-art adversarial machine learning and deep learning based user in-home activity inference and fingerprinting attacks and help users achieve the balance between their IoT data utility and privacy preserving.

Bayesian Mechanics of Synaptic Learning Under the Free-Energy Principle.

The brain is a biological system comprising nerve cells and orchestrates its embodied agent's perception, behavior, and learning in dynamic environments. The free-energy principle (FEP) advocated by Karl Friston explicates the local, recurrent, and self-supervised cognitive dynamics of the brain's higher-order functions. In this study, we continue to refine the FEP through a physics-guided formulation; specifically, we apply our theory to synaptic learning by considering it an inference problem under the FEP and derive the governing equations, called Bayesian mechanics. Our study uncovers how the brain infers weight changes and postsynaptic activity, conditioned on the presynaptic input, by deploying generative models of the likelihood and prior belief. Consequently, we exemplify the synaptic efficacy in the brain with a simple model; in particular, we illustrate that the brain organizes an optimal trajectory in neural phase space during synaptic learning in continuous time, which variationally minimizes synaptic surprisal.

A hierarchical active inference model of spatial alternation tasks and the hippocampal-prefrontal circuit

Cognitive problem-solving benefits from cognitive maps aiding navigation and planning. Physical space navigation involves hippocampal (HC) allocentric codes, while abstract task space engages medial prefrontal cortex (mPFC) task-specific codes. Previous studies show that challenging tasks, like spatial alternation, require integrating these two types of maps. The disruption of the HC-mPFC circuit impairs performance. We propose a hierarchical active inference model clarifying how this circuit solves spatial interaction tasks by bridging physical and task-space maps. Simulations demonstrate that the model’s dual layers develop effective cognitive maps for physical and task space. The model solves spatial alternation tasks through reciprocal interactions between the two layers. Disrupting its communication impairs decision-making, which is consistent with empirical evidence. Additionally, the model adapts to switching between multiple alternation rules, providing a mechanistic explanation of how the HC-mPFC circuit supports spatial alternation tasks and the effects of disruption.

A sparse diverse-branch large kernel convolutional neural network for human activity recognition using wearables

During the past decade, large convolutional kernels have long been under the shadow of small convolutional kernels since the introduction of VGG backbone network. It always remains mysterious whether one can design pure convolutional neural network (CNN) while plugging larger kernels to model long-range dependency for human activity recognition (HAR), which has been rarely explored in previous literatures. In this paper, we revive the usage of larger kernels in the context of HAR and attempt to eliminate the performance gap between large kernels and small kernels by strategically applying a large receptive field, without incurring high memory and computational footprints. Built on two recipes, i.e., Diverse-Branch and Dynamic Sparsity, we design a pure CNN architecture named SLK-Net for activity recognition, which is equipped with sparse diverse-branch larger kernels. To validate the effectiveness of our approach, we perform a series of extensive experiments on four public benchmarks including UCI-HAR, WISDM, UniMiB-SHAR and USC-HAD, which show that large kernels can benefit its ability to capture long-range dependency and consistently beat state-of-the-art small-kernel counterparts across a wide range of activity classification tasks. Real activity inference latency is measured on a mobile device, which reveals that such sparse diverse-branch kernels can lead to inference speedup than vanilla large kernels. We hope this work may further inspire relevant CNN-based studies in the HAR community.

Active Inference Models of AV Takeovers: Relating Model Parameters to Trust, Situation Awareness, and Fatigue.

Our objectives were to assess the efficacy of active inference models for capturing driver takeovers from automated vehicles and to evaluate the links between model parameters and self-reported cognitive fatigue, trust, and situation awareness. Control transitions between human drivers and automation pose a substantial safety and performance risk. Models of driver behavior that predict these transitions from data are a critical tool for designing safer, human-centered, systems but current models do not sufficiently account for human factors. Active inference theory is a promising approach to integrate human factors because of its grounding in cognition and translation to a quantitative modeling framework. We used data from a driving simulation to develop an active inference model of takeover performance. After validating the model's predictions, we used Bayesian regression with a spike and slab prior to assess substantial correlations between model parameters and self-reported trust, situation awareness, fatigue, and demographic factors. The model accurately captured driving takeover times. The regression results showed that increases in cognitive fatigue were associated with increased uncertainty about the need to takeover, attributable to mapping observations to environmental states. Higher situation awareness was correlated with a more precise understanding of the environment and state transitions. Higher trust was associated with increased variance in environmental conditions associated with environmental states. The results align with prior theory on trust and active inference and provide a critical connection between complex driver states and interpretable model parameters. The active inference framework can be used in the testing and validation of automated vehicle technology to calibrate design parameters to ensure safety.

Implementing and Evaluating the Timed Up and Go Test Automation Using Smartphones and Smartwatches.

Physical performance tests aim to assess the physical abilities and mobility skills of individuals for various healthcare purposes. They are often driven by experts and usually performed at their practice, and therefore they are resource-intensive and time-demanding. For tests based on objective measurements (e.g., duration, repetitions), technology can be used to automate them, allowing the patients to perform the test themselves, more frequently and anywhere, while alleviating the expert from supervising the test. The well-known Timed Up and Go (TUG) test, typically used for mobility assessment, is an ideal candidate for automation, as inertial sensors (among others) can be deployed to detect the various movements constituting the test without expert supervision. To move from expert-led testing to self-administered testing, we present a mHealth system capable of automating the TUG test using a pocket-sized smartphone or a wrist smartwatch paired with a smartphone, where data from inertial sensors are used to detect the activities carried out by the patient while performing the test and compute their results in real time. All processing (i.e., data processing, machine learning-based activity inference, results calculation) takes place on the smartphone. The use of both devices to automate the TUG test was evaluated (w.r.t. accuracy, reliability and battery consumption) and mutually compared, and set off with a reference method, obtaining excellent Bland-Altman agreement results and Intraclass Correlation Coefficient reliability. Results also suggest that the smartwatch-based system performs better than the smartphone-based system.

Transcriptomics profiling of the non-small cell lung cancer microenvironment across disease stages reveals dual immune cell-type behaviors.

Lung cancer is the leading cause of cancer death worldwide, with poor survival despite recent therapeutic advances. A better understanding of the complexity of the tumor microenvironment is needed to improve patients' outcome. We applied a computational immunology approach (involving immune cell proportion estimation by deconvolution, transcription factor activity inference, pathways and immune scores estimations) in order to characterize bulk transcriptomics of 62 primary lung adenocarcinoma (LUAD) samples from patients across disease stages. Focusing specifically on early stage samples, we validated our findings using an independent LUAD cohort with 70 bulk RNAseq and 15 scRNAseq datasets and on TCGA datasets. Through our methodology and feature integration pipeline, we identified groups of immune cells related to disease stage as well as potential immune response or evasion and survival. More specifically, we reported a duality in the behavior of immune cells, notably natural killer (NK) cells, which was shown to be associated with survival and could be relevant for immune response or evasion. These distinct NK cell populations were further characterized using scRNAseq data, showing potential differences in their cytotoxic activity. The dual profile of several immune cells, most notably T-cell populations, have been discussed in the context of diseases such as cancer. Here, we report the duality of NK cells which should be taken into account in conjunction with other immune cell populations and behaviors in predicting prognosis, immune response or evasion.

Predictive brain activity related to auditory information is associated with performance in speech comprehension tasks in noisy environments.

Understanding speech in noisy environments is challenging even for individuals with normal hearing, and it poses a significant challenge for those with hearing impairments or listening difficulties. There are limitations associated with the current methods of evaluating speech comprehension in such environments, especially in individuals with peripheral hearing impairments. According to the predictive coding model, speech comprehension is an active inference process that integrates sensory information through the interaction of bottom-up and top-down processing. Therefore, in this study, we aimed to examine the role of prediction in speech comprehension using an electrophysiological marker of anticipation: stimulus-preceding negativity (SPN). We measured SPN amplitude in young adults with normal hearing during a time-estimation task with auditory feedback under both quiet and noisy conditions. The results showed that SPN amplitude significantly increased in noisy environments. Moreover, individual differences in SPN amplitude correlated with performance in a speech-in-noise test. The increase in SPN amplitude was interpreted as reflecting the increased requirement for attentional resources for accurate prediction of speech information. These findings suggest that SPN could serve as a noninvasive neural marker for assessing individual differences in top-down processing involved in speech comprehension in noisy environments.

Exploring action-oriented models via active inference for autonomous vehicles

Being able to robustly interact with and navigate a dynamic environment has been a long-standing challenge in intelligent transportation systems. Autonomous agents can use models that mimic the human brain to learn how to respond to other participants’ actions in the environment and proactively coordinate with the dynamics. Modeling brain learning procedures is challenging for multiple reasons, such as stochasticity, multimodality, and unobservant intents. Active inference may be defined as the Bayesian modeling of a brain with a biologically plausible model of the agent. Its primary idea relies on the free energy principle and the prior preference of the agent. It enables the agent to choose an action that leads to its preferred future observations. An exploring action-oriented model is introduced to address the inference complexity and solve the exploration–exploitation dilemma in unobserved environments. It is conducted by adapting active inference to an imitation learning approach and finding a theoretical connection between them. We present a multimodal self-awareness architecture for autonomous driving systems where the proposed techniques are evaluated on their ability to model proper driving behavior. Experimental results provide the basis for the intelligent driving system to make more human-like decisions and improve agent performance to avoid a collision.

Multimodal and Multifactor Branching Time Active Inference.

Active inference is a state-of-the-art framework for modeling the brain that explains a wide range of mechanisms. Recently, two versions of branching time active inference (BTAI) have been developed to handle the exponential (space and time) complexity class that occurs when computing the prior over all possible policies up to the time horizon. However, those two versions of BTAI still suffer from an exponential complexity class with regard to the number of observed and latent variables being modeled. We resolve this limitation by allowing each observation to have its own likelihood mapping and each latent variable to have its own transition mapping. The implicit mean field approximation was tested in terms of its efficiency and computational cost using a dSprites environment in which the metadata of the dSprites data set was used as input to the model. In this setting, earlier implementations of branching time active inference (namely, BTAIVMP and BTAIBF) underperformed in relation to the mean field approximation (BTAI3MF) in terms of performance and computational efficiency. Specifically, BTAIVMP was able to solve 96.9% of the task in 5.1 seconds, and BTAIBF was able to solve 98.6% of the task in 17.5 seconds. Our new approach outperformed both of its predecessors by solving the task completely (100%) in only 2.559 seconds.

Deconstructing Deep Active Inference: A Contrarian Information Gatherer.

Active inference is a theory of perception, learning, and decision making that can be applied to neuroscience, robotics, psychology, and machine learning. Recently, intensive research has been taking place to scale up this framework using Monte Carlo tree search and deep learning. The goal of this activity is to solve more complicated tasks using deep active inference. First, we review the existing literature and then progressively build a deep active inference agent as follows: we (1) implement a variational autoencoder (VAE), (2) implement a deep hidden Markov model (HMM), and (3) implement a deep critical hidden Markov model (CHMM). For the CHMM, we implemented two versions, one minimizing expected free energy, CHMM[EFE] and one maximizing rewards, CHMM[reward]. Then we experimented with three different action selection strategies: the ε-greedy algorithm as well as softmax and best action selection. According to our experiments, the models able to solve the dSprites environment are the ones that maximize rewards. On further inspection, we found that the CHMM minimizing expected free energy almost always picks the same action, which makes it unable to solve the dSprites environment. In contrast, the CHMM maximizing reward keeps on selecting all the actions, enabling it to successfully solve the task. The only difference between those two CHMMs is the epistemic value, which aims to make the outputs of the transition and encoder networks as close as possible. Thus, the CHMM minimizing expected free energy repeatedly picks a single action and becomes an expert at predicting the future when selecting this action. This effectively makes the KL divergence between the output of the transition and encoder networks small. Additionally, when selecting the action down the average reward is zero, while for all the other actions, the expected reward will be negative. Therefore, if the CHMM has to stick to a single action to keep the KL divergence small, then the action down is the most rewarding. We also show in simulation that the epistemic value used in deep active inference can behave degenerately and in certain circumstances effectively lose, rather than gain, information. As the agent minimizing EFE is not able to explore its environment, the appropriate formulation of the epistemic value in deep active inference remains an open question.

Planning with tensor networks based on active inference

Planning with tensor networks based on active inference

Cortical development in the structural model and free energy minimization.

A model of neocortical development invoking Friston's Free Energy Principle is applied within the Structural Model of Barbas etal. and the associated functional interpretation advanced by Tucker and Luu. Evolution of a neural field with Hebbian and anti-Hebbian plasticity, maximizing synchrony and minimizing axonal length by apoptotic selection, leads to paired connection systems with mirror symmetry, interacting via Markov blankets along their line of reflection. Applied to development along the radial lines of development in the Structural Model, a primary Markov blanket emerges between the centrifugal synaptic flux in layers 2,3 and 5,6, versus the centripetal flow in layer 4, and axonal orientations in layer 4 give rise to the differing shape and movement sensitivities characteristic of neurons of dorsal and ventral neocortex. Prediction error minimization along the primary blanket integrates limbic and subcortical networks with the neocortex. Synaptic flux bypassing the blanket triggers the arousal response to surprising stimuli, enabling subsequent adaptation. As development progresses ubiquitous mirror systems separated by Markov blankets and enclosed blankets-within-blankets arise throughout neocortex, creating the typical order and response characteristics of columnar and noncolumnar cortex.

Blind CT Image Quality Assessment Using DDPM-Derived Content and Transformer-Based Evaluator.

Lowering radiation dose per view and utilizing sparse views per scan are two common CT scan modes, albeit often leading to distorted images characterized by noise and streak artifacts. Blind image quality assessment (BIQA) strives to evaluate perceptual quality in alignment with what radiologists perceive, which plays an important role in advancing low-dose CT reconstruction techniques. An intriguing direction involves developing BIQA methods that mimic the operational characteristic of the human visual system (HVS). The internal generative mechanism (IGM) theory reveals that the HVS actively deduces primary content to enhance comprehension. In this study, we introduce an innovative BIQA metric that emulates the active inference process of IGM. Initially, an active inference module, implemented as a denoising diffusion probabilistic model (DDPM), is constructed to anticipate the primary content. Then, the dissimilarity map is derived by assessing the interrelation between the distorted image and its primary content. Subsequently, the distorted image and dissimilarity map are combined into a multi-channel image, which is inputted into a transformer-based image quality evaluator. By leveraging the DDPM-derived primary content, our approach achieves competitive performance on a low-dose CT dataset.

Slow but flexible or fast but rigid? Discrete and continuous processes compared

A tradeoff exists when dealing with complex tasks composed of multiple steps. High-level cognitive processes can find the best sequence of actions to achieve a goal in uncertain environments, but they are slow and require significant computational demand. In contrast, lower-level processing allows reacting to environmental stimuli rapidly, but with limited capacity to determine optimal actions or to replan when expectations are not met. Through reiteration of the same task, biological organisms find the optimal tradeoff: from action primitives, composite trajectories gradually emerge by creating task-specific neural structures. The two frameworks of active inference – a recent brain paradigm that views action and perception as subject to the same free energy minimization imperative – well capture high-level and low-level processes of human behavior, but how task specialization occurs in these terms is still unclear. In this study, we compare two strategies on a dynamic pick-and-place task: a hybrid (discrete-continuous) model with planning capabilities and a continuous-only model with fixed transitions. Both models rely on a hierarchical (intrinsic and extrinsic) structure, well suited for defining reaching and grasping movements, respectively. Our results show that continuous-only models perform better and with minimal resource expenditure but at the cost of less flexibility. Finally, we propose how discrete actions might lead to continuous attractors and compare the two frameworks with different motor learning phases, laying the foundations for further studies on bio-inspired task adaptation.