Learning Task Research Articles

InvPT++: Inverted Pyramid Multi-Task Transformer for Visual Scene Understanding.

Multi-task scene understanding aims to design models that can simultaneously predict several scene understanding tasks with one versatile model. Previous studies typically process multi-task features in a more local way, and thus cannot effectively learn spatially global and cross-task interactions, which hampers the models' ability to fully leverage the consistency of various tasks in multi-task learning. To tackle this problem, we propose an Inverted Pyramid multi-task Transformer, capable of modeling cross-task interaction among spatial features of different tasks in a global context. Specifically, we first utilize a transformer encoder to capture task-generic features for all tasks. And then, we design a transformer decoder to establish spatial and cross-task interaction globally, and a novel UP-Transformer block is devised to increase the resolutions of multi-task features gradually and establish cross-task interaction at different scales. Furthermore, two types of Cross-Scale Self-Attention modules, i.e., Fusion Attention and Selective Attention, are proposed to efficiently facilitate cross-task interaction across different feature scales. An Encoder Feature Aggregation strategy is further introduced to better model multi-scale information in the decoder. Comprehensive experiments on several 2D/3D multi-task benchmarks clearly demonstrate our proposal's effectiveness, establishing significant state-of-the-art performances. The codes are publicly available https://github.com/prismformore/Multi-Task-Transformer/tree/main/InvPT.

Temporal validity reassessment: commonsense reasoning about information obsoleteness

It is useful for machines to know whether text information remains valid or not for various applications including text comprehension, story understanding, temporal information retrieval, and user state tracking on microblogs as well as via chatbot conversations. This kind of inference is still difficult for current models, including also large language models, as it requires temporal commonsense knowledge and reasoning. We approach in this paper the task of Temporal Validity Reassessment, inspired by traditional natural language reasoning to determine the updates of the temporal validity of text content. The task requires judgment whether actions expressed in a sentence are still ongoing or rather completed, hence, whether the sentence still remains valid or has become obsolete, given the presence of context in the form of a supplementary content such as a follow-up sentence. We first construct our own dataset for this task and train several machine learning models. Then we propose an effective method for learning information from an external knowledge base that gives information regarding temporal commonsense knowledge. Using our prepared dataset, we introduce a machine learning model that incorporates the information from the knowledge base and demonstrate that incorporating external knowledge generally improves the results. We also experiment with different embedding types to represent temporal commonsense knowledge as well as with data augmentation methods to increase the size of our dataset.

GPT-4 in Education: Evaluating Aptness, Reliability, and Loss of Coherence in Solving Calculus Problems and Grading Submissions

AbstractIn this paper, we initially investigate the capabilities of GPT-3 5 and GPT-4 in solving college-level calculus problems, an essential segment of mathematics that remains under-explored so far. Although improving upon earlier versions, GPT-4 attains approximately 65% accuracy for standard problems and decreases to 20% for competition-like scenarios. Overall, the models prove to be unreliable due to common arithmetic errors.Our primary contribution lies then in examining the use of ChatGPT for grading solutions to calculus exercises. Our objectives are to probe an in-context learning task with less emphasis over direct calculations; recognize positive applications of ChatGPT in educational contexts; highlight a potentially emerging facet of AI that could necessitate oversight; and introduce unconventional AI benchmarks, for which models like GPT are untrained. Pertaining to the latter, we uncover a tendency for loss of coherence in extended contexts. Our findings suggest that while the current ChatGPT exhibits comprehension of the grading task and often provides relevant outputs, the consistency of grading is marred by occasional loss of coherence and hallucinations. Intriguingly, GPT-4's overall scores, delivered in mere moments, align closely with human graders, although its detailed accuracy remains suboptimal.This work suggests that, when appropriately orchestrated, collaboration between human graders and LLMs like GPT-4 might combine their unique strengths while mitigating their respective shortcomings In this direction, it is imperative to consider implementing transparency, fairness, and appropriate regulations in the near future.

Self-initiated self-repairs of connected speech and novel vocabulary learning during the first year of recovery from aphasia: four longitudinal case studies

ABSTRACT Background Self-initiated self-repair of problems in speech production has been suggested to be associated with aphasia treatment gains and recovery, but the topic has been little studied. Similarly, the ability to learn novel words has been associated with anomia treatment benefits, although longitudinal data on such learning is lacking. Aims The present study’s aim was to examine whether and how self-initiated self-repairs in connected speech and the ability to learn novel vocabulary reflected recovery from aphasia as measured with cognitive-linguistic tests. Additionally, our aim was to explore whether self-initiated repair and novel word learning changed in parallel during recovery. Methods & Procedures Four people with aphasia (PWA) (2 anomic, 1 Wernicke, and 1 global aphasia) were tested in the subacute (11 to 68 days post stroke) and chronic (12 months post stroke) phases of recovery with cognitive-linguistic tests, connected speech tasks, and a computerized learning task. In the learning task, PWA associated 6 pseudowords with 6 unfamiliar referents. The associations were learned from an ambiguous setup with feedback. At both time points, learning was measured as correct immediate recognition of the items and recognition maintenance one week post-training. The proportion of self-initiated self-repairs and the proportion of unrepaired problems in connected speech were used as measures of self-repair behavior. Outcomes & Results All participants initiated self-repairs both in the subacute and chronic phases, and their self-repairs were most often related to aphasic difficulties in speech production. Between the subacute and chronic phases, self-repairs changed both in quality and quantity. The proportion of self-repair elements decreased in 4/4 participants and the proportion of unrepaired problem segments decreased in 3/4 participants. Of the participants, 3/4 showed significant novel word learning and/or maintenance already in the subacute phase and 4 /4 in the chronic phase. Conclusion Reduction in self-initiated self-repairs of speech reflected recovery from aphasia. Novel word learning improved during aphasia recovery from the subacute to the chronic phase. Novel word learning was possible even in subacute, severe aphasia and even in an ambiguous task set-up with no explicitly provided word-referent connections. Novel word learning and self-initiated self-repairs did not change in a clearly parallel manner; however, these findings need to be tested with a larger sample.

A Natural Sonification Mapping for Handwriting

The sonification of handwriting has been shown effective in various learning tasks. In this paper, the authors investigate the sound design used for handwriting interaction based on a simple and cost-efficient prototype. The authentic interaction sound is compared with physically informed sonification designs that employ either natural or inverted mapping. In an experiment, participants copied text and drawings. The authors found simple measures of the structure-borne audio signal that showed how participants were affected in their movements, but only when drawing. In contrast, participants rated the sound features differently only for writing. The authentic interaction sound generally scored best, followed by a natural sonification mapping.

Co‐designing enduring learning analytics prediction and support tools in undergraduate biology courses

AbstractEven highly motivated undergraduates drift off their STEM career pathways. In large introductory STEM classes, instructors struggle to identify and support these students. To address these issues, we developed co‐redesign methods in partnership with disciplinary experts to create high‐structure STEM courses that better support students and produce informative digital event data. To those data, we applied theory‐ and context‐relevant labels to reflect active and self‐regulated learning processes involving LMS‐hosted course materials, formative assessments, and help‐seeking tools. We illustrate the predictive benefits of this process across two cycles of model creation and reapplication. In cycle 1, we used theory‐relevant features from 3 weeks of data to inform a prediction model that accurately identified struggling students and sustained its accuracy when reapplied in future semesters. In cycle 2, we refit a model with temporally contextualized features that achieved superior accuracy using data from just two class meetings. This modelling approach can produce durable learning analytics solutions that afford scaled and sustained prediction and intervention opportunities that involve explainable artificial intelligence products. Those same products that inform prediction can also guide intervention approaches and inform future instructional design and delivery.Practitioner notesWhat is already known about this topic Learning analytics includes an evolving collection of methods for tracing and understanding student learning through their engagements with learning technologies. Prediction models based on demographic data can perpetuate systemic biases. Prediction models based on behavioural event data can produce accurate predictions of academic success, and validation efforts can enrich those data to reflect students' self‐regulated learning processes within learning tasks. What this paper adds Learning analytics can be successfully applied to predict performance in an authentic postsecondary STEM context, and the use of context and theory as guides for feature engineering can ensure sustained predictive accuracy upon reapplication. The consistent types of learning resources and cyclical nature of their provisioning from lesson to lesson are hallmarks of high‐structure active learning designs that are known to benefit learners. These designs also provide opportunities for observing and modelling contextually grounded, theory‐aligned and temporally positioned learning events that informed prediction models that accurately classified students upon initial and later reapplications in subsequent semesters. Co‐design relationships where researchers and instructors work together toward pedagogical implementation and course instrumentation are essential to developing unique insights for feature engineering and producing explainable artificial intelligence approaches to predictive modelling. Implications for practice and/or policy High‐structure course designs can scaffold student engagement with course materials to make learning more effective and products of feature engineering more explainable. Learning analytics initiatives can avoid perpetuation of systemic biases when methods prioritize theory‐informed behavioural data that reflect learning processes, sensitivity to instructional context and development of explainable predictors of success rather than relying on students' demographic characteristics as predictors. Prioritizing behaviours as predictors improves explainability in ways that can inform the redesign of courses and design of learning supports, which further informs the refinement of learning theories and their applications.

Belief Updating, Childhood Maltreatment, and Paranoia in Schizophrenia-Spectrum Disorders.

Exposure to childhood maltreatment-a risk factor for psychosis is associated with paranoia-may impact one's beliefs about the world and how beliefs are updated. We hypothesized that increased exposure to childhood maltreatment is related to volatility-related belief updating, specifically higher expectations of volatility, and that these relationships are strongest for threat-related maltreatment. Additionally, we tested whether belief updating mediates the relationship between maltreatment and paranoia. Belief updating was measured in 75 patients with schizophrenia-spectrum disorders and 76 nonpsychiatric controls using a 3-option probabilistic reversal learning (3PRL) task. A Hierarchical Gaussian Filter (HGF) was used to estimate computational parameters of belief updating, including prior expectations of volatility (μ03). The Childhood Trauma Questionnaire (CTQ) was used to assess cumulative maltreatment, threat, and deprivation exposure. Paranoia was measured using the Positive and Negative Syndrome Scale (PANSS) and the revised Green et al. Paranoid Thoughts Scale (R-GPTS). Greater exposure to childhood maltreatment is associated with higher prior expectations of volatility in the whole sample and in individuals with schizophrenia-spectrum disorders. This was specific to threat-related maltreatment, rather than deprivation, in schizophrenia-spectrum disorders. Paranoia was associated with both exposure to childhood maltreatment and volatility priors, but we did not observe a significant indirect effect of volatility priors on the relationship between maltreatment and paranoia. Our study suggests that individuals with schizophrenia-spectrum disorders who were exposed to threatening experiences during childhood expect their environment to be more volatile, potentially facilitating aberrant belief updating and conferring risk for paranoia.

Behavioral Interventions Can Improve Brain Injury-Induced Deficits in Behavioral Flexibility and Impulsivity Linked to Impaired Reward-Feedback Beta Oscillations.

Traumatic brain injury (TBI) affects a large population, resulting in severe cognitive impairments. Although cognitive rehabilitation is an accepted treatment for some deficits, studies in patients are limited in ability to probe physiological and behavioral mechanisms. Therefore, animal models are needed to optimize strategies. Frontal TBI in a rat model results in robust and replicable cognitive deficits, making this an ideal candidate for investigating various behavioral interventions. In this study, we report three distinct frontal TBI experiments assessing behavior well into the chronic post-injury period using male Long-Evans rats. First, we evaluated the impact of frontal injury on local field potentials recorded simultaneously from 12 brain regions during a probabilistic reversal learning (PbR) task. Next, a set of rats were tested on a similar PbR task or an impulsivity task (differential reinforcement of low-rate behavior [DRL]) and half received salient cues associated with reinforcement contingencies to encourage engagement in the target behavior. After intervention on the PbR task, brains were stained for markers of activity. On the DRL task, cue relevance was decoupled from outcomes to determine if beneficial effects persisted on impulsive behavior. TBI decreased the ability to detect reinforced outcomes; this was evident in task performance and reward-feedback signals occurring at beta frequencies in lateral orbitofrontal cortex (OFC) and associated frontostriatal regions. The behavioral intervention improved flexibility and increased OFC activity. Intervention also reduced impulsivity, even after cues were decoupled, which was partially mediated by improvements in timing behavior. The current study established a platform to begin investigating cognitive rehabilitation in rats and identified a strong role for dysfunctional OFC signaling in probabilistic learning after frontal TBI.

Enhancing academic performance and student engagement in health education: insights from Work Station Learning Activities (WSLA)

Making health science students aware of the importance of basic science knowledge for professional practice is a major educational challenge, especially during the early years of preclinical courses. Here, using an integrated curricular approach, we analyze whether Work Station Learning Activities (WSLA), which combine active learning methodologies for teaching basic science in clinical scenarios, can help to develop deeper learning and student engagement. In order to increase student motivation, we evaluated the effectiveness of WSLA using statistical analyses and an observation tool based on the ICAP (Interactive, Constructive, Active, and Passive) framework, which categorizes learning tasks based on the nature of student engagement. Statistical analyses revealed positive correlations between the different summative evaluations along the development of the activities, indicating the learning process inherent to WSLA progression and affirming the positive influence of WSLA on academic outcomes. Comparing the pre- and post-tests, students scored significantly higher on the post-test (statistically significant p < 0.001). WSLA promotes both constructivist and interactive learning, as validated by its alignment with the ICAP model. The study examines student engagement through systematic observation, revealing a relationship between student engagement and final grades. Students who exhibit constructive learning consistently earn higher grades, emphasizing the positive impact of active engagement. Thus, passive behavior profiles show a significant proportion of fails (40%), while constructive profiles stand out as the sole recipients of the coveted excellent rating. Ultimately, this study contributes to our understanding of the effectiveness of WSLA in promoting active learning and enhancing student engagement within integrated health education curricula. It highlights the importance of active learning behaviors for academic success and suggests avenues for further research to optimize integrated teaching methodologies in medical education.

Morphophonemic analysis boosts orthographic and semantic learning of academic words for Spanish–English bilinguals

BackgroundUpper elementary grade students encounter increasingly complex texts with abundant morphologically complex words. Despite the positive effects of morphology‐based vocabulary instruction, emergent bilinguals with limited word reading skills may need additional support.MethodsThis study investigated the effects of morphological analysis and morphophonemic analysis instruction on the orthographic and semantic learning of morphologically complex academic words. Fourth‐ and fifth‐grade Spanish–English bilingual students (N = 30) in the United States participated in two learning tasks. In learning task one, participants learned two sets of carefully matched derivational words under morphological and whole‐word learning conditions. In learning task two, the same participants learned two other sets of words under morphophonemic and whole‐word learning conditions. Each learning task included two learning sessions interweaved by meaning recall and spelling production assessments. Cross‐classified multilevel regression was used to assess the effects of intervention conditions, child and item predictors, as well as cross‐level interactions.ResultsFindings showed no significant difference in meaning recall and spelling measures between morphological and whole‐word conditions. Students performed significantly better in the morphophonemic condition than in the whole‐word condition for both meaning and affix spelling outcomes. The effect of morphophonemic intervention over whole word condition was stronger for younger students and words of lower base frequencies. Student meaning and spelling performance was significantly predicted by their word reading skills.ConclusionsMorphophonemic analysis instruction enhances word meaning recall and spelling of complex derivative words.

Creating realistic anterior segment optical coherence tomography images using generative adversarial networks

AimsTo develop a generative adversarial network (GAN) capable of generating realistic high-resolution anterior segment optical coherence tomography (AS-OCT) images.MethodsThis study included 142 628 AS-OCT B-scans from the American University of...

Learning vision-based robotic manipulation tasks sequentially in offline reinforcement learning settings

Abstract With the rise of deep reinforcement learning (RL) methods, many complex robotic manipulation tasks are being solved. However, harnessing the full power of deep learning requires large datasets. Online RL does not suit itself readily into this paradigm due to costly and time-consuming agent-environment interaction. Therefore, many offline RL algorithms have recently been proposed to learn robotic tasks. But mainly, all such methods focus on a single-task or multitask learning, which requires retraining whenever we need to learn a new task. Continuously learning tasks without forgetting previous knowledge combined with the power of offline deep RL would allow us to scale the number of tasks by adding them one after another. This paper investigates the effectiveness of regularisation-based methods like synaptic intelligence for sequentially learning image-based robotic manipulation tasks in an offline-RL setup. We evaluate the performance of this combined framework against common challenges of sequential learning: catastrophic forgetting and forward knowledge transfer. We performed experiments with different task combinations to analyse the effect of task ordering. We also investigated the effect of the number of object configurations and the density of robot trajectories. We found that learning tasks sequentially helps in the retention of knowledge from previous tasks, thereby reducing the time required to learn a new task. Regularisation-based approaches for continuous learning, like the synaptic intelligence method, help mitigate catastrophic forgetting but have shown only limited transfer of knowledge from previous tasks.

Multiplexed representation of others in the hippocampal CA1 subfield of female mice

Hippocampal place cells represent the position of a rodent within an environment. In addition, recent experiments show that the CA1 subfield of a passive observer also represents the position of a conspecific performing a spatial task. However, whether this representation is allocentric, egocentric or mixed is less clear. In this study we investigated the representation of others during free behavior and in a task where female mice learned to follow a conspecific for a reward. We found that most cells represent the position of others relative to self-position (social-vector cells) rather than to the environment, with a prevalence of purely egocentric coding modulated by context and mouse identity. Learning of a pursuit task improved the tuning of social-vector cells, but their number remained invariant. Collectively, our results suggest that the hippocampus flexibly codes the position of others in multiple coordinate systems, albeit favoring the self as a reference point.

Latent Representation Learning for Geospatial Entities

Representation learning has been instrumental in the success of machine learning, offering compact and performant data representations for diverse downstream tasks. In the spatial domain, it has been pivotal in extracting latent patterns from various data types, including points, polylines, polygons, and networked structures. However, existing approaches often fall short of explicitly capturing both semantic and spatial information, relying on proxies and synthetic features. This paper presents GeoNN, a novel graph neural network-based model designed to learn spatially-aware embeddings for geospatial entities. GeoNN leverages edge features generated from geodesic functions, dynamically selecting relevant features based on relative locations. It introduces both transductive (GeoNN-T) and inductive (GeoNN-I) models, ensuring effective encoding of geospatial features and scalability with entity changes. Extensive experiments demonstrate GeoNN’s effectiveness in location-sensitive superpixel-based graphs and real-world points of interest, outperforming baselines across various evaluation measures.

Microglial TNFα controls daily changes in synaptic GABAARs and sleep slow waves.

Microglia sense the changes in their environment. How microglia actively translate these changes into suitable cues to adapt brain physiology is unknown. We reveal an activity-dependent regulation of cortical inhibitory synapses by microglia, driven by purinergic signaling acting on P2RX7 and mediated by microglia-derived TNFα. We demonstrate that sleep induces microglia-dependent synaptic enrichment of GABAARs in a manner dependent on microglial TNFα and P2RX7. We further show that microglia-specific depletion of TNFα alters slow waves during NREM sleep and blunt memory consolidation in sleep-dependent learning tasks. Together, our results reveal that microglia orchestrate sleep-intrinsic plasticity of synaptic GABAARs, sculpt sleep slow waves, and support memory consolidation.

Enhanced cognitive flexibility and phasic striatal dopamine dynamics in a mouse model of low striatal tonic dopamine.

The catecholamine neuromodulators dopamine and norepinephrine are implicated in motor function, motivation, and cognition. Although roles for striatal dopamine in these aspects of behavior are well established, the specific roles for cortical catecholamines in regulating striatal dopamine dynamics and behavior are less clear. We recently showed that elevating cortical dopamine but not norepinephrine suppresses hyperactivity in dopamine transporter knockout (DAT-KO) mice, which have elevated striatal dopamine levels. In contrast, norepinephrine transporter knockout (NET-KO) mice have a phenotype distinct from DAT-KO mice, as they show elevated extracellular cortical catecholamines but reduced baseline striatal dopamine levels. Here we evaluated the consequences of altered catecholamine levels in NET-KO mice on cognitive flexibility and striatal dopamine dynamics. In a probabilistic reversal learning task, NET-KO mice showed enhanced reversal learning, which was consistent with larger phasic dopamine transients (dLight) in the dorsomedial striatum (DMS) during reward delivery and reward omission, compared to WT controls. Selective depletion of dorsal medial prefrontal cortex (mPFC) norepinephrine in WT mice did not alter performance on the reversal learning task but reduced nestlet shredding. Surprisingly, NET-KO mice did not show altered breakpoints in a progressive ratio task, suggesting intact food motivation. Collectively, these studies show novel roles of cortical catecholamines in the regulation of tonic and phasic striatal dopamine dynamics and cognitive flexibility, updating our current views on dopamine regulation and informing future therapeutic strategies to counter multiple psychiatric disorders.

Influence of Error-Augmentation on the Dynamics of Visuomotor Skill Acquisition: Insights from Proxy-Process Models.

Our study addresses the critical question of how learners acquire skills without the constant crutch of feedback, utilizing a specialized training approach with intermittent feedback. Despite recognized benefits in skill retention, the underlying mechanisms of intermittent feedback in motor control neuroscience remain elusive. Leveraging a previously published dataset from visuomotor learning experiments with intermittent feedback, we tested a wide range of proxy-process models that posit the presence of an inferred error signal even when an explicit sensory performance is not present. Furthermore, these proxy-process models investigated the impact of error-augmentation (EA) training on visuomotor learning dynamics. Rigorous cross-validation consistently identified a second-order proxy-process model structure accurately predicting motor learning across subjects and learning tasks. Model parameters elucidated the varying influences of EA settings on the rates of change in error, inter-trial variability, and steady-state performance. We then introduced a dynamic-Proxy support Multi-Rate Motor Learning (dPxMRML) model, which shed light on EA's effects on the fast and slow learning dynamics. The dPxMRML model accurately predicted subjects' performance during and beyond training phases, highlighting EA settings conducive to long-term retention. This research yields crucial insights for personalized training program design, applicable in neuro-rehabilitation, sports, and performance training.

A Theoretical View of Linear Backpropagation and its Convergence.

Backpropagation (BP) is widely used for calculating gradients in deep neural networks (DNNs). Applied often along with stochastic gradient descent (SGD) or its variants, BP is considered as a de-facto choice in a variety of machine learning tasks including DNN training and adversarial attack/defense. Recently, a linear variant of BP named LinBP was introduced for generating more transferable adversarial examples for performing black-box attacks, by (Guo et al. 2020). Although it has been shown empirically effective in black-box attacks, theoretical studies and convergence analyses of such a method is lacking. This paper serves as a complement and somewhat an extension to Guo et al. (2020) paper, by providing theoretical analyses on LinBP in neural-network-involved learning tasks, including adversarial attack and model training. We demonstrate that, somewhat surprisingly, LinBP can lead to faster convergence in these tasks in the same hyper-parameter settings, compared to BP. We confirm our theoretical results with extensive experiments.

Enhancing Visual Grounding in Vision-Language Pre-Training With Position-Guided Text Prompts.

Vision-Language Pre-Training (VLP) has demonstrated remarkable potential in aligning image and text pairs, paving the way for a wide range of cross-modal learning tasks. Nevertheless, we have observed that VLP models often fall short in terms of visual grounding and localization capabilities, which are crucial for many downstream tasks, such as visual reasoning. In response, we introduce a novel Position-guided Text Prompt (PTP) paradigm to bolster the visual grounding abilities of cross-modal models trained with VLP. In the VLP phase, PTP divides an image into N x N blocks and employs a widely-used object detector to identify objects within each block. PTP then reframes the visual grounding task as a fill-in-the-blank problem, encouraging the model to predict objects in given blocks or regress the blocks of a given object, exemplified by filling "[P]" or "[O]" in a PTP sentence such as "The block [P] has a [O]." This strategy enhances the visual grounding capabilities of VLP models, enabling them to better tackle various downstream tasks. Additionally, we integrate the seconda-order relationships between objects to further enhance the visual grounding capabilities of our proposed PTP paradigm. Incorporating PTP into several state-of-the-art VLP frameworks leads to consistently significant improvements across representative cross-modal learning model architectures and multiple benchmarks, such as zero-shot Flickr30 k Retrieval (+5.6 in average recall@1) for ViLT baseline, and COCO Captioning (+5.5 in CIDEr) for the state-of-the-art BLIP baseline. Furthermore, PTP attains comparable results with object-detector-based methods and a faster inference speed, as it discards its object detector during inference, unlike other approaches.

42. Posttraumatic Stress Disorder (PTSD) is Associated With Abnormal Amygdala and Striatal Signaling of Temporal Prediction Errors During a Passive Reward Learning Task

42. Posttraumatic Stress Disorder (PTSD) is Associated With Abnormal Amygdala and Striatal Signaling of Temporal Prediction Errors During a Passive Reward Learning Task