Model-based prioritization for acquiring protection

Protection, or the mitigation of harm, often involves the capacity to prospectively plan the actions needed to combat a threat. The computational architecture of decisions involving protection remains unclear, as well as whether these decisions differ from other positive prospective actions. Here we examine effects of valence and context by comparing protection to reward, which occurs in a different context but is also positively valenced, and punishment, which also occurs in an aversive context but differs in valence. We applied computational modeling across three independent studies (Total N=600) using five iterations of a ‘two-step’ behavioral task to examine model-based reinforcement learning for protection, reward, and punishment in humans. Decisions motivated by acquiring safety via protection evoked a higher degree of model-based control than acquiring reward and avoiding punishment, with no significant differences in learning rate. The context-valence asymmetry characteristic of protection increased deployment of flexible decision strategies, suggesting model-based control depends on the context in which outcomes are encountered as well as the valence of the outcome.

Multiscalar Architectures

Many of the great advances in modern computing are supplied by modeling architectures that practice a crucial division in descriptive labor by asking distinct forms of submodeling to work together in cooperative harmony without engaging in a straightforward amalgamation of conclusions. Commonly these distinct submodels are aligned with characteristic scale lengths within their target systems so that a preliminary modeling (Δ‎H) that calculates how a system normally behaves upon a macroscopic scale becomes subjected to corrective suggestions arising from a lower-scale modeling (Δ‎L) that focuses upon the local factors that occasionally upset the behavioral presumptions codified within the Δ‎H scheme. The syntactic safeguards within this technique that avert inconsistency and an unmanageable explosion in computational complexity keep their various levels of submodeling isolated from one another. They only pass corrective messages of a specialized character (called “homogenizations”) amongst themselves without attempting to fully amalgamate their localized conclusions into a shared narrative. The computational architecture merely demands that the various submodels reach accord with respect to the homogenization messages that they exchange amongst themselves. This book argues that unnoticed reasoning arrangements of this kind provide the proper diagnosis of the “Mystery of Physics 101” tensions that troubled Hertz (the distinct usages of “force” he noticed operate upon distinct size scales in the manner of a modern multiscalar scheme). It is then suggested that the natural development of many forms of linguistic attainment lead to reasoning architectures of this general character, although we often fail to recognize the subtle strategies that undergird their operations.

Context-Sensitive Control of Adaptation: Self-Modeling Networks for Human Mental Processes Using Mental Models

Within their mental and social processes, humans often learn, adapt and apply specific mental models of processes in the world or other persons, as a kind of blueprints. In this paper, it is discussed how analysis of this provides useful inspiration for the development of new computational approaches from a Machine Learning and Network-Oriented Modeling perspective. Three main elements are: applying the mental model by internal simulation, developing and revising a mental model by some form of adaptation, and exerting control over this adaptation in a context-sensitive manner. This concept of controlled adaptation relates to the Plasticity Versus Stability Conundrum from neuroscience. The presented analysis has led to a three-level computational architecture for controlled adaptation. It is discussed and illustrated by examples of applications how this three-level computational architecture can be specified based on a self-modeling network and used to model controlled learning and adaptation processes based on mental models in a context-sensitive manner.

A Canonical Scheme of Bottom-Up and Top-Down Information Flows in the Frontoparietal Network

Goal-directed behavior often involves temporal separation and flexible context-dependent association between sensory input and motor output. The control of goal-directed behavior is proposed to lie in the frontoparietal network, but the computational architecture of this network remains elusive. Based on recent rodent studies that measured and manipulated projection neurons in the frontoparietal network together with findings from earlier primate studies, we propose a canonical scheme of information flows in this network. The parietofrontal pathway transmits the spatial information of a sensory stimulus or internal motor bias to drive motor programs in the frontal areas. This pathway might consist of multiple parallel connections, each controlling distinct motor effectors. The frontoparietal pathway sends the spatial information of cognitively processed motor plans through multiple parallel connections. Each of these connections could support distinct spatial functions that use the motor target information, including attention allocation, multi-body part coordination, and forward estimation of movement state (i.e., forward models). The parallel pathways in the frontoparietal network enable dynamic interactions between regions that are tuned for specific goal-directed behaviors. This scheme offers a promising framework within which the computational architecture of the frontoparietal network and the underlying circuit mechanisms can be delineated in a systematic way, providing a holistic understanding of information processing in this network. Clarifying this network may also improve the diagnosis and treatment of behavioral deficits associated with dysfunctional frontoparietal connectivity in various neurological disorders including Alzheimer’s disease.

Computational Architecture Mediating Inhibitory Control of Coordinated Eye-Hand Movements

Significant progress has been made in understanding the computational and neural architecture that mediates eye and hand movements made in isolation. However, less is known about the mechanisms that control these movements when they are coordinated. Here, we outline our computational approaches using accumulation-to-threshold and race-to-threshold models to elucidate the mechanisms that initiate and inhibit these movements. We suggest that, depending on the behavioral context, the initiation and inhibition of coordinated eye-hand movements can operate in two modes- coupled and decoupled. The coupled-mode operates when the task context requires a tight coupling between the effectors; a common command initiates both effectors, and a unitary inhibitory process is responsible for stopping them. Conversely, the decoupled mode operates when the task context demands weaker coupling between the effectors; separate commands initiate the eye and hand, and separate inhibitory processes are responsible for stopping them. We hypothesize that higher-order control processes assess the behavioral context and choose the most appropriate mode. This computational architecture can explain heterogeneous results observed across many studies that have investigated the control of coordinated eye-hand movements and may also serve as a general framework to understand the control of complex multi-effector movements.

Steady state evoked potential (SSEP) responses in the primary and secondary somatosensory cortices of anesthetized cats: Nonlinearity characterized by harmonic and intermodulation frequencies

When presented with an oscillatory sensory input at a particular frequency, F [Hz], neural systems respond with the corresponding frequency, f [Hz], and its multiples. When the input includes two frequencies (F1 and F2) and they are nonlinearly integrated in the system, responses at intermodulation frequencies (i.e., n1*f1+n2*f2 [Hz], where n1 and n2 are non-zero integers) emerge. Utilizing these properties, the steady state evoked potential (SSEP) paradigm allows us to characterize linear and nonlinear neural computation performed in cortical neurocircuitry. Here, we analyzed the steady state evoked local field potentials (LFPs) recorded from the primary (S1) and secondary (S2) somatosensory cortex of anesthetized cats (maintained with alfaxalone) while we presented slow (F1 = 23Hz) and fast (F2 = 200Hz) somatosensory vibration to the contralateral paw pads and digits. Over 9 experimental sessions, we recorded LFPs from N = 1620 and N = 1008 bipolar-referenced sites in S1 and S2 using electrode arrays. Power spectral analyses revealed strong responses at 1) the fundamental (f1, f2), 2) its harmonic, 3) the intermodulation frequencies, and 4) broadband frequencies (50-150Hz). To compare the computational architecture in S1 and S2, we employed simple computational modeling. Our modeling results necessitate nonlinear computation to explain SSEP in S2 more than S1. Combined with our current analysis of LFPs, our paradigm offers a rare opportunity to constrain the computational architecture of hierarchical organization of S1 and S2 and to reveal how a large-scale SSEP can emerge from local neural population activities.

Prior expectations evoke stimulus-specific activity in the deep layers of the primary visual cortex

The way we perceive the world is strongly influenced by our expectations. In line with this, much recent research has revealed that prior expectations strongly modulate sensory processing. However, the neural circuitry through which the brain integrates external sensory inputs with internal expectation signals remains unknown. In order to understand the computational architecture of the cortex, we need to investigate the way these signals flow through the cortical layers. This is crucial because the different cortical layers have distinct intra- and interregional connectivity patterns, and therefore determining which layers are involved in a cortical computation can inform us on the sources and targets of these signals. Here, we used ultra-high field (7T) functional magnetic resonance imaging (fMRI) to reveal that prior expectations evoke stimulus-specific activity selectively in the deep layers of the primary visual cortex (V1). These findings are in line with predictive processing theories proposing that neurons in the deep cortical layers represent perceptual hypotheses and thereby shed light on the computational architecture of cortex.

A Computational Architecture for Inference of a Quantized-CNN for Detecting Atrial Fibrillation

Atrial Fibrillation is a common cardiac arrhythmia, which is characterized by an abnormal heartbeat rhythm that can be life-threatening. Recently, researchers have proposed several Convolutional Neural Networks (CNNs) to detect Atrial Fibrillation. CNNs have high requirements on computing and memory resources, which usually demand the use of High Performance Computing (eg, GPUs). This high energy demand is a challenge for portable devices. Therefore, efficient hardware implementations are required. We propose a computational architecture for the inference of a Quantized Convolutional Neural Network (Q-CNN) that allows the detection of the Atrial Fibrillation (AF). The architecture exploits data-level parallelism by incorporating SIMD-based vector units, which is optimized in terms of computation and storage and also optimized to perform both the convolutional and fully connected layers. The computational architecture was implemented and tested in a Xilinx Artix-7 FPGA. We present the experimental results regarding the quantization process in a different number of bits, hardware resources, and precision. The results show an accuracy of 94% accuracy for 22-bits. This work aims to be the basis for the future implementation of a portable, low-cost, and high-reliability device for the diagnosis of Atrial Fibrillation.