Computational Explanation Research Articles

A neural network model of differentiation and integration of competing memories.

What determines when neural representations of memories move together (integrate) or apart (differentiate)? Classic supervised learning models posit that, when two stimuli predict similar outcomes, their representations should integrate. However, these models have recently been challenged by studies showing that pairing two stimuli with a shared associate can sometimes cause differentiation, depending on the parameters of the study and the brain region being examined. Here, we provide a purely unsupervised neural network model that can explain these and other related findings. The model can exhibit integration or differentiation depending on the amount of activity allowed to spread to competitors - inactive memories are not modified, connections to moderately active competitors are weakened (leading to differentiation), and connections to highly active competitors are strengthened (leading to integration). The model also makes several novel predictions - most importantly, that when differentiation occurs as a result of this unsupervised learning mechanism, it will be rapid and asymmetric, and it will give rise to anticorrelated representations in the region of the brain that is the source of the differentiation. Overall, these modeling results provide a computational explanation for a diverse set of seemingly contradictory empirical findings in the memory literature, as well as new insights into the dynamics at play during learning.

A neural network model of differentiation and integration of competing memories

What determines when neural representations of memories move together (integrate) or apart (differentiate)? Classic supervised learning models posit that, when two stimuli predict similar outcomes, their representations should integrate. However, these models have recently been challenged by studies showing that pairing two stimuli with a shared associate can sometimes cause differentiation, depending on the parameters of the study and the brain region being examined. Here, we provide a purely unsupervised neural network model that can explain these and other related findings. The model can exhibit integration or differentiation depending on the amount of activity allowed to spread to competitors — inactive memories are not modified, connections to moderately active competitors are weakened (leading to differentiation), and connections to highly active competitors are strengthened (leading to integration). The model also makes several novel predictions — most importantly, that when differentiation occurs as a result of this unsupervised learning mechanism, it will be rapid and asymmetric, and it will give rise to anticorrelated representations in the region of the brain that is the source of the differentiation. Overall, these modeling results provide a computational explanation for a diverse set of seemingly contradictory empirical findings in the memory literature, as well as new insights into the dynamics at play during learning.

Differentiation and Integration of Competing Memories: A Neural Network Model.

What determines when neural representations of memories move together (integrate) or apart (differentiate)? Classic supervised learning models posit that, when two stimuli predict similar outcomes, their representations should integrate. However, these models have recently been challenged by studies showing that pairing two stimuli with a shared associate can sometimes cause differentiation, depending on the parameters of the study and the brain region being examined. Here, we provide a purely unsupervised neural network model that can explain these and other related findings. The model can exhibit integration or differentiation depending on the amount of activity allowed to spread to competitors - inactive memories are not modified, connections to moderately active competitors are weakened (leading to differentiation), and connections to highly active competitors are strengthened (leading to integration). The model also makes several novel predictions - most importantly, that when differentiation occurs as a result of this unsupervised learning mechanism, it will be rapid and asymmetric, and it will give rise to anticorrelated representations in the region of the brain that is the source of the differentiation. Overall, these modeling results provide a computational explanation for a diverse set of seemingly contradictory empirical findings in the memory literature, as well as new insights into the dynamics at play during learning.

How computation explains

Cognitive science gives computational explanations of the brain. Philosophers have treated these explanations as if they simply claim that the brain computes. We have therefore assumed that to understand how and why computational explanation works, we must understand what it is to compute. In contrast, I argue that we can understand computational explanation by describing the resources it brings to bear on the study of the brain. Specifically, I argue that it introduces concepts and formalisms that complement cognitive science's modeling goals. This allows us to understand computational explanation without having to debate what it is to compute.

Computational Explanation of the Photovoltaic Cells Properties of the PCBM and PC71BM Derivatives using the Density Functional Theory

Computational Explanation of the Photovoltaic Cells Properties of the PCBM and PC71BM Derivatives using the Density Functional Theory

Brain-inspired bodily self-perception model for robot rubber hand illusion

The core of bodily self-consciousness involves perceiving ownership of one's body. A central question is how body illusions like the rubber hand illusion (RHI) occur. Existing theoretical models still lack satisfying computational explanations from connectionist perspectives, especially for how the brain encodes body perception and generates illusions from neuronal interactions. Moreover, the integration of disability experiments is also neglected. Here, we integrate biological findings of bodily self-consciousness to propose a brain-inspired bodily self-perception model by which perceptions of bodily self are autonomously constructed without any supervision signals. We successfully validated the model with six RHI experiments and a disability experiment on an iCub humanoid robot and simulated environments. The results show that our model can not only well-replicate the behavioral and neural data of monkeys in biological experiments but also reasonably explain the causes and results of RHI at the neuronal level, thus contributing to the revelation of mechanisms underlying RHI.

Limitative computational explanations

Limitative computational explanations

Toward computational neuroconstructivism: a framework for developmental systems neuroscience.

Brain development is underpinned by complex interactions between neural assemblies, driving structural and functional change. This neuroconstructivism (the notion that neural functions are shaped by these interactions) is core to some developmental theories. However, due to their complexity, understanding underlying developmental mechanisms is challenging. Elsewhere in neurobiology, a computational revolution has shown that mathematical models of hidden biological mechanisms can bridge observations with theory building. Can we build a similar computational framework yielding mechanistic insights for brain development? Here, we outline the conceptual and technical challenges of addressing this theory gap, and demonstrate that there is great potential in specifying brain development as mathematically defined processes operating within physical constraints. We provide examples, alongside broader ingredients needed, as the field explores computational explanations of system-wide development.

Novel Methods of AI-Based Gait Analysis in Post-Stroke Patients

Research on gait function assessment is important not only in terms of the patient’s mobility, but also in terms of the patient’s current and future quality of life, ability to achieve health goals, family life, study and/or work, and participation in society. The main methods used herein include a literature review and an analysis of our own original research and concepts. This study used the historical data of 92 ischemic stroke patients (convenience trial) undergoing two kinds of rehabilitation. An artificial neural network, fractal analysis, and fuzzy analysis were used to analyze the results. Our findings suggest that artificial neural networks, fuzzy logic, and multifractal analysis are useful for building simple, low-cost, and efficient computational tools for gait analysis, especially in post-stroke patients. The novelty lies in the simultaneous application of the three aforementioned technologies to develop a computational model for the analysis of a patient’s post-stroke gait. The contribution of this work consists not only in its proposal of a new and useful clinical tool for gait assessment, even in the most severe post-stroke cases, but also in its attempt to offer a comprehensive computational explanation of observed gait phenomena and mechanisms. We conclude by anticipating more advanced and broader future applications of artificial intelligence (AI) in gait analysis, especially in post-stroke patients.

Computing in the nick of time

AbstractThe medium‐independence of computational descriptions has shaped common conceptions of computational explanation. So long as our goal is to explain how a system successfully carries out its computations, then we only need to describe the abstract series of operations that achieve the desired input–output mapping, however they may be implemented. It is argued that this abstract conception of computational explanation cannot be applied to so‐called real‐time computing systems, in which meeting temporal deadlines imposed by the systems with which a device interfaces are constitutive of the computing tasks that a device performs. Instead, real‐time computing reveals the need for alternative conceptions of computational explanation, as well as computational implementation, that eschew medium‐independence.

Macaques recognize features in synthetic images derived from ventral stream neurons

Primates can recognize features in virtually all types of images, an ability that still requires a comprehensive computational explanation. One hypothesis is that visual cortex neurons learn patterns from scenes, objects, and textures, and use these patterns to interpolate incoming visual information. We have used machine learning algorithms to instantiate visual patterns stored by neurons-we call these highly activating images prototypes. Prototypes from inferotemporal (IT) neurons often resemble parts of real-world objects, such as monkey faces and body parts, a similarity established via pretrained neural networks [C. R. Ponce et al.,Cell 177, 999-1009.e10 (2019)]and naïve human participants [A. Bardon, W. Xiao, C. R. Ponce, M. S. Livingstone, G. Kreiman, Proc. Natl. Acad. Sci. U.S.A. 119, e2118705119 (2022)]. However, it is not known whether monkeys themselves perceive similarities between neuronal prototypes and real-world objects. Here, we investigated whether monkeys reported similarities between prototypes and real-world objects using a two-alternative forced choice task. We trained the animals to saccade to synthetic images of monkeys, and subsequently tested how they classified prototypes synthesized from IT and primary visual cortex (V1). We found monkeys classified IT prototypes as conspecifics more often than they did random generator images and V1 prototypes, and their choices were partially predicted by convolutional neural networks. Further, we confirmed that monkeys could abstract general shape information from images of real-world objects. Finally, we verified these results with human participants. Our results provide further evidence that prototypes from cortical neurons represent interpretable abstractions from the visual world.

Mindful attention promotes control of brain network dynamics for self-regulation and discontinues the past from the present

Mindful attention is characterized by acknowledging the present experience as a transient mental event. Early stages of mindfulness practice may require greater neural effort for later efficiency. Early effort may self-regulate behavior and focalize the present, but this understanding lacks a computational explanation. Here we used network control theory as a model of how external control inputs-operationalizing effort-distribute changes in neural activity evoked during mindful attention across the white matter network. We hypothesized that individuals with greater network controllability, thereby efficiently distributing control inputs, effectively self-regulate behavior. We further hypothesized that brain regions that utilize greater control input exhibit shorter intrinsic timescales of neural activity. Shorter timescales characterize quickly discontinuing past processing to focalize the present. We tested these hypotheses in a randomized controlled study that primed participants to either mindfully respond or naturally react to alcohol cues during fMRI and administered text reminders and measurements of alcohol consumption during 4 wk postscan. We found that participants with greater network controllability moderated alcohol consumption. Mindful regulation of alcohol cues, compared to one's own natural reactions, reduced craving, but craving did not differ from the baseline group. Mindful regulation of alcohol cues, compared to the natural reactions of the baseline group, involved more-effortful control of neural dynamics across cognitive control and attention subnetworks. This effort persisted in the natural reactions of the mindful group compared to the baseline group. More-effortful neural states had shorter timescales than less effortful states, offering an explanation for how mindful attention promotes being present.

No computation without implementation? A potential problem for the single hierarchy view of physical computation

The so-called integration problem concerning mechanistic and computational explanation asks how they are related to each other. One approach is that a computational explanation is a species of mechanistic explanation. According to this view, computational or mathematical descriptions are mechanism sketches or macroscopic descriptions that include computationally relevant and exclude computationally irrelevant physical properties. Some suggest that this results in a so-called single hierarchy view of physical computation, where computational or mathematical properties sit together in the same mechanistic hierarchy with the implementational properties. This view can be contrasted with a separate hierarchy view, according to which computational and physical descriptions have their own hierarchies which are related to each other via a bridging implementation relation. The single hierarchy view has been criticized for downplaying the explanatory value of computational explanations and not being hospitable to multiple realization of cognitive processes. In this paper, I argue that (1) the aforementioned criticisms fail, and (2) there might be a deeper problem with the single hierarchy view, which is that the single hierarchy view might collapse into a separate hierarchy view. The kind of abstraction used by the single hierarchy view does not seem to grant mathematical or computational descriptions but only more stripped physical or implementational descriptions.

Computational Modelling for Alcohol Use Disorder

In this paper, I examine Reinforcement Learning (RL) modelling practice in psychiatry, in the context of alcohol use disorders. I argue that the epistemic roles RL currently plays in the development of psychiatric classification and search for explanations of clinically relevant phenomena are best appreciated in terms of Chang’s (2004) account of epistemic iteration, and by distinguishing mechanistic and aetiological modes of computational explanation.

Evaluation of the Adsorptive Process on Adsorbent Surfaces as a Function of Pressure in an Isosteric System Compared with Adsorption Isotherm

The developing significance of adsorption has brought about a steadily expanding quantity of logical and innovative writing on different adsorbents. This paper intends to propose experimental and computational methods for measuring the strength of adsorbate–solid interactions. It primarily focuses on the use of graphs to measure the effectiveness of an adsorbate’s bonds with the solid adsorbent by determining the isosteric heat. The Clausius–Clapeyron model equation is used to determine the isosteric enthalpy of adsorption from two adsorption isotherms at various but close temperatures, with ΔT of 10 °C. A full computational explanation of the Clausius–Clapeyron model equation for determining ΔHads is provided using experimental data. Logarithmic plots of uptakes vs. p in the low-pressure zone for the Freundlich–Langmuir graph are used to assess and confirm the quality of the crucial underlying isotherms. The isosteric heat was found to be between 13.5 kJ/mol and 16 kJ/mol.

The problem with appealing to history in defining neural representations

Representations seem to play a major role in many neuroscientific explanations. Philosophers have long attempted to properly define what it means for a neural state to be a representation of a specific content. Teleosemantic theories of content which characterize representations, in part, by appealing to a historical notion of function, are often regarded as our best path towards an account of neural representations. This paper points to the anti-representationalist consequences of these accounts. I argue that assuming such teleosemantic views will deprive representations of their explanatory role in computational explanations. My argument rests on the claim that many explanations in cognitive neuroscience are entirely independent of any historical considerations. In making this claim, I will also offer an adapted version of the famous Swampperson thought experiment, which is better suited to discussions of subpersonal neural representations.

Microwave assisted post-synthetic modification of IRMOF-3 and MIL-68-NH2 onto cotton for Fuel purification with computational explanation

Removal of phenol (as oxygenated compound) and indole (as nitrogen-based compound) from liquid fuel was currently studied using modified cotton fabric-based metal organic frameworks (MOFs). Functionalized cotton (PO@Cotton) was firstly synthesized by interaction with epichlorohydrin (ECH). Two MOFs based on Zn (IRMOF-3) and In (MIL-68-NH2) were applied to PO@Cottton via the microwave assisted post-synthetic modification technique. The obtained MOF@PO@Cotton was employed in removal of phenol and indole from n-heptane. Langmuir maximum capacity (Qm) of phenol and indole was greatly enlarged from 144.2 and 183.7 mg/g for cotton to 403.5 – 416.5 mg/g and 480.7 – 529.7 mg/g for MOF@PO@Cotton, respectively. The adsorption amounts of MOF@PO@Cotton fabric were reduced by 18.8 – 20.0 % for phenol and 8.2 – 17.7 % for indole, after 5 reusing cycles. The adsorption mechanism of pollutants was suggested by using Monte Carlo simulation to be carried out through hydrogen bonding and π-π interactions.

Learning About the Self: Motives for Coherence and Positivity Constrain Learning From Self-Relevant Social Feedback

People learn about themselves from social feedback, but desires for coherence and positivity constrain how feedback is incorporated into the self-concept. We developed a network-based model of the self-concept and embedded it in a reinforcement-learning framework to provide a computational account of how motivations shape self-learning from feedback. Participants (N = 46 adult university students) received feedback while evaluating themselves on traits drawn from a causal network of trait semantics. Network-defined communities were assigned different likelihoods of positive feedback. Participants learned from positive feedback but dismissed negative feedback, as reflected by asymmetries in computational parameters that represent the incorporation of positive versus negative outcomes. Furthermore, participants were constrained in how they incorporated feedback: Self-evaluations changed less for traits that have more implications and are thus more important to the coherence of the network. We provide a computational explanation of how motives for coherence and positivity jointly constrain learning about the self from feedback, an explanation that makes testable predictions for future clinical research.

The determinacy of computation

A skeptical worry known as ‘the indeterminacy of computation’ animates much recent philosophical reflection on the computational identity of physical systems. On the one hand, computational explanation seems to require that physical computing systems fall under a single, unique computational description at a time. On the other, if a physical system falls under any computational description, it seems to fall under many simultaneously. Absent some principled reason to take just one of these descriptions in particular as relevant for computational explanation, widespread failure of computational explanation would appear to follow. This paper advances a new solution to the indeterminacy of computation. Very roughly, I argue that the computational identity of a physical system is determinate relative to a contextually specified way of regarding that system computationally—known as a labelling scheme. When a system simultaneously implements multiple computations, it does so relative to different labelling schemes. But relative to a fixed labelling scheme, a physical system has a unique computational identity. I argue that this relativistic conception of computational identity vindicates computational explanation in the face of simultaneous implementation.

Representational and Computational Changes in Heritage Language Grammars

Abstract The notion of complexity has been applied to descriptions and comparisons of languages and to explanations related to ease and difficulty of various linguistic phenomena in first and second language acquisition. It has been noted that compared to baseline grammars, heritage language grammars are less complex, displaying morphological simplification and structural shrinking, especially among heritage speakers with lower proficiency in the language. On some recent proposals of gender agreement in Spanish and Norwegian (Fuchs et al., 2015; Lohndal & Putnam, 2020), these differences are representational, affecting the projection of functional categories and feature specifications in the syntax. An alternative possibility is that differences between baseline and heritage grammars arise from computational considerations related to bilingualism, affecting speed of lexical access and feature reassembly online in the minority language. We illustrate this proposal with empirical data from gender agreement and differential object marking. Although presented as alternatives, the representational and computational explanations are not incompatible, and may both be adequate to capture varying levels of variability modulated by linguistic proficiency. These proposals formalize bilingual acquisition models of grammar competition and directly relate the availability and type of input (the acquisition evidence) to the locus and nature of the grammatical differences between heritage and baseline grammars.