machine learning | John Thorp

Reflectify

Tue, 21 Jan 2025 00:00:00 +0000

Reflectify is an AI-powered research tool that analyzes students’ written reflections on their learning experiences to predict exam performance and provide personalized improvement strategies.

Built using natural language processing, machine learning, and SHAP explainable AI, Reflectify helps students understand which reflection strategies actually improve their academic outcomes.

🔬 Research-Based Features

Performance Prediction: Probabilistic exam improvement forecasting
Strategy Detection: Automatic identification of learning and reflection strategies
SHAP Explainability: Clear explanations of which features matter most
Personalized Feedback: GPT-4 powered recommendations tailored to individual reflections
Evidence-Based Guidance: Specific action steps based on educational research

📊 Educational Impact

This tool bridges the gap between metacognitive theory and practical application, helping students develop more effective reflection practices through data-driven insights.

contextual stability as a continuous moderator of event segmentation

Sat, 31 Dec 2022 00:00:00 +0000

Our senses receive a constant stream of information, with one moment leading continuously to the next. Afterwards, however, we remember our experiences as discrete events. These events are thought to be typically organized around stable contexts, with context often consisting of an unchanging goal or physical location. Studies of the individual differences in this segmentation process have contributed insights into basic cognition and mapped novel clinical markers. For instance, the ability to normatively detect the boundaries between these events is weakened across such wide-reaching disorders as Alzheimer’s disorder, schizophrenia, autism-spectrum disorder, and even during healthy aging. While the hippocampus is known to hold a critical role in event segmentation broadly, current functional models have yet to be extended to its role in event segmentation. A major limitation of the existing literature to parse these individual differences and neural functional models, however, is that it has treated event boundaries as binary occurrences. Theoretical accounts hold that more stable contexts should lead to stronger event boundaries, and that the ability to gradate these event boundaries ought to rely on mechanisms of cognitive control. Critically, these claims have yet be tested. This ommission leaves sources of variance that, in tandem with cognitive control functions, may provide meaningful clinical markers as well as provide theoretical insight.

To test this, I built a paradigm that modulates the stability of a context as a continuous function rather than a binary occurence. This allowed me to test how individuals’ event segmentation behaviors evolved over a continuous range of contextual stability, and particularly how these developed non-linearly and differently from each other. I found that the difference in participants’ memory for boundary and non-boundary item-color memory (a standard measure of event segmentation) increased non-linearly as contexts became more stable and that participants did this very differently, with some increasing very smoothly and others increasing suddenly at an inflection point. Future work will tie these behaviors to existing clinical markers and cognitive control measures known to covary with clinical outcomes. fMRI studies can then examine how univariate signals evolve across the body of the hippocampus as well as how multivariate signals in the lateral entorhinal cortex process temporal context at different levels of contextual stability.

representational granularity across the hippocampus

Tue, 31 May 2022 00:00:00 +0000

A particularly elusive puzzle concerning the hippocampus is how the structural differences along its long anteroposterior axis might beget meaningful functional differences, particularly in terms of the granularity of information processing. One measure posits to quantify this granularity by calculating the average statistical independence of the BOLD signal across neighboring voxels, or intervoxel similarity (IVS), and has shown the anterior hippocampus to process coarser-grained information than the posterior hippocampus. This measure, however, has yielded opposing results in studies of developmental and healthy aging samples, which also varied in fMRI acquisition parameters and hippocampal parcellation methods. To reconcile these findings, we measured IVS across two separate resting-state fMRI acquisitions and compared the results across many of the most widely used parcellation methods in a large young-adult sample of male and female humans (Acquisition 1, N = 233; Acquisition 2, N = 176). Finding conflicting results across acquisitions and parcellations, we reasoned that a data-driven approach to hippocampal parcellation is necessary. To this end, we implemented a group masked independent components analysis to identify functional subunits of the hippocampus, most notably separating the anterior hippocampus into separate anterior-medial, anterior-lateral, and posteroanterior-lateral components. Measuring IVS across these components revealed a decrease in IVS along the medial-lateral axis of the anterior hippocampus but an increase from anterior to posterior. We conclude that intervoxel similarity is deeply affected by parcellation and that grounding one’s parcellation in a functionally informed approach might allow for a more complex and reliable characterization of the hippocampus.

Data-driven clustering of functional signals reveals gradients in processing both within the anterior hippocampus and across its long axis

Sat, 01 Jan 2022 00:00:00 +0000