Angelica Crown
Undergraduate Institution: Wesleyan University
Mentor: David Creswell and Janine Dutcher
University and Department: Carnegie Mellon University
Project Description:
Under the guidance of Janine Dutcher and David Creswell, Angelica Crown worked to specialize a data processing pipeline for diffusion neuroimaging data. We were provided BASH scripts to convert MRI data into BIDS format, run QSI Prep (MRI data processing taking BIDS input), and run reconstruction. In the given scripts, the code was run in a singularity container. The goal of this summer project was to convert the scripts to run in a docker container and debug the code to work on the data for the Headspace mindfulness app study conducted by the Health and Human Performance Lab at Carnegie Mellon University.
The time allowed this summer only allowed for adjustments to the bids format and QSI Prep scripts. Brain masking and de-noising was performed on 12 participants of the 100 participants. Each session for a participant takes four hours to run. The output from QSI Prep shows promising data quality, as seen through the DWI summary comparing the cross-correlation value and the head-motion correction model.
Future work on this project includes reconstruction and analysis of white matter tracts, as well as changes in gray matter structure, resting state, threat and reward-related task responding, and multi-modal analyses. This study will investigate the brain as a mechanism for intervention effects on stress and workplace outcomes, using MRI data, gene expression, and self-report.
Jessica De La Torre
Undergraduate Institution: University of California San Diego
Mentor: Alison Barth
University and Department: Carnegie Mellon University
Project Description:
In the cortex, somatostatin neuron networks are essential for learning and reward-based behaviors. To study the plasticity of somatostatin neurons, multi whisker sensory association training was used. In this training, an air puff to the whiskers was coupled with a water reward. Occasionally, there were blank trials where an air puff was delivered without a water reward. While in the training cage, the mouse is allowed to roam freely. During the recording sessions, the mouse receives an air puff while head fixed to a microscope. Behaviors of the mouse including whisking, locomotion, and somatostatin neuron activity are recorded in these sessions. In this study, the correlation between locomotion and the activity of the somatostatin neurons is being studied. Locomotion is an important factor to study because it takes sensory inputs to develop an appropriate physical response, like running. This brought us to ask: Does locomotion modulate the activity of somatostatin neurons in mice? The data collected has aligned with current knowledge of somatostatin neurons; the subtypes behave differently. In response to the air puff, cells tend show higher calcium transients after the puff. When correlated to the locomotion, the neurons also tend to respond in a varied manner, with some portraying a higher calcium transient when the mouse is stationary. This is an important observation for the classification of somatostatin neuron subtypes, and can be a possible way to identify certain subtypes that are consistent with their stimulus responses. To accomplish this, Jessica visualized the mouse velocities across time, and identified some epochs that stood out because of their variance.
Abraham Eldo
Undergraduate Institution: University of Pittsburgh
Mentor: Michele Insanally
University and Department: University of Pittsburgh
Project Description:
Neural responses to pure tones are conserved throughout the mammalian auditory pathway. Furthermore, inactivation experiments in the rodent auditory cortex (AC) show that AC is required for various auditory tasks. However, the precise timing of stimulus processing in AC remains unknown. Previous research shows that frontal cortex inactivation during stimulus offset results in a significant drop in performance. Similarly, decoding accuracy of AC neurons is highest at stimulus offset. These suggest that stimulus processing in AC after stimulus presentation is important for sensory guided decision making. In order to explore this proposition, Abe developed a novel two-alternative forced choice auditory categorization task for head-fixed mice, incorporating a delay period between stimulus and response. Mice are trained to categorize pure tones as either high or low and report their decision by licking left or right after a 500 ms delay period signaled by an LED go-cue. High and low categories are separated by a boundary frequency of 14 kHz. Each category consists of 5 pure tones modulated by 0.1 octaves away from the category boundary. Abe outlined a 4 phase shaping paradigm to train animals to both respond after the delay period and categorize tones correctly. To determine if stimulus processing in AC after stimulus presentation is important for sensory guided decision making, he will optogenetically inactivate AC during the stimulus and delay periods while recording single units with high density silicon probes. Inactivations during the delay period will occur in five discrete 100ms windows to determine at what time AC inactivation results in the greatest behavioral deficits. Additionally, he will use various decoding approaches to decode stimulus and choice from AC neurons. He will then compare decoding accuracies across inactivation windows to determine which inactivation window results in greatest loss in decoding accuracy. Together, these experiments will help determine the precise timing of AC stimulus processing for sensory guided behavior.
Sofia Juliani
Undergraduate Institution: Rutgers University--New Brunswick
Mentor: Steve Chase and Matt Smith
University and Department: Carnegie Mellon University
Project Description:
Recordings of neural data are inherently noisy, and removing this noise remains a challenging problem. Removing noise from neural recordings can facilitate post-hoc analysis of neural data, as well as increase the performance of real-time applications such as brain-computer interfaces. An existing method to de-noise neural signals involves hand-labeling waveforms after data collection, but it is time-consuming and cannot be implemented for real-time use of brain-computer interfaces. Previous work has created a neural network classifier that inputs a waveform and outputs the likelihood of it being a spike (Issar et al., 2020). This model preprocesses waveforms effectively compared to other approaches but it does not generalize well to data from different brain regions.
This summer, Sofia implemented two approaches to improve performance of this classifier: balancing the data used to train the model, and testing different architectures. To better balance the training data for the model, Sofia utilized k-means clustering to identify different features of hand-labeled spike and noise waveforms collected from various brain regions including prefrontal cortex, primary motor cortex, frontal eye fields, and visual cortex in macaque monkeys. This method balanced the multiple forms of spikes and noise waves in the dataset. Next, Sofia ran experiments on multiple neural network architectures to further improve classification. Sofia retrained the original model architecture, as well as a model with two hidden layers, and two convolutional neural networks on the improved dataset. After computing the accuracy of these models, she found that they performed significantly better than the original architecture. The best convolutional neural network had a classification accuracy of 88.76% on waveforms from multiple brain regions, compared to the original model’s accuracy of 87.51%. Thus, k-means clustering provides an effective way of curating a balanced dataset to train an effective model for denoising neural recordings. This improvement would thus increase the efficacy of brain-computer interfaces in real-time and facilitate post-hoc analysis of neural activity.
Kelly Kennedy
Undergraduate Institution: Florida State University
Mentor: Jason Bohland
University and Department: University of Pittsburgh
Project Description:
Developmental stuttering is a neurodevelopmental disorder that affects 5-8% of children. Children who stutter (CWS) also often comorbid symptoms of atypical executive function, including attention disorders such as Attention Deficit Hyperactive Disorder (ADHD). It has been hypothesized that stuttering may also involve speech sound processing deficits rather than only speech motor control issues. In a collaboration between the Brain Systems for Language Lab and the Speech Neural Systems Lab, we are investigating the relationship between neural systems for attention, speech sound processing, and speaking, as well as the differences between stuttering and control subjects. Children (ages 7-15) participated in a functional magnetic resonance imaging (fMRI) experiment involving a complex storybook listening task. The fMRI data were preprocessed using fMRIprep software. This summer, Kelly used CONN software to assess task-dependent functional connectivity (FC) between the attention networks (ATTN), Speech Motor Articulatory Networks (SMAN), and the auditory cortex (AUD) in control subjects. In CONN, she estimated condition-specific ROI-to-ROI FC matrices across 28 ROIs and contrasted the listening condition with a resting condition. She found significant increases in FC between ROIs in the dorsal and ventral auditory streams during the listening task. She also discovered an increase in FC between AUD and ATTN during the listening task, though this finding did not reach statistical significance. Data collection is ongoing, which will help to achieve higher statistical power and allow for testing of group differences between CWS and controls.
Emily Kolach
Undergraduate Institution: The University of Texas at Austin
Mentor: Tobias Teichert and Aaron Batista
University and Department: University of Pittsburgh
Project Description:
Internal physiological states such as hunger or thirst are powerful motivators that profoundly alter our thoughts, emotions, and behaviors. While these effects must be mediated by the central nervous system, we are only beginning to understand how brain function of a sated animal differs from one that is highly motivated to seek out food or water. Under the mentorship of Dr. Tobias Teichert and Dr. Aaron Batista, Emily Kolach investigated the specific frequencies and locations that differed among these highly contrasting states. They ran fast fourier transforms on resting state data from before and after a task to determine if there is any variation in power. Upon finding considerable differences between states, they searched for regions and frequencies where the effect was most remarkable. They found that the difference in power was most significant in the parietal and temporal cortices in the delta band. They ran correlation analyses amongst hundreds of channels and frequencies, as well as several dependent variables to identify if neural activity correlated with behavioral. Future plans involve confirming these correlations in a separate dataset, understanding the relation between high delta power and internal states, and studying the delta-power over smaller increments of the resting state to see how long this effect is visible.
Yunshu Li
Undergraduate Institution: Carnegie Mellon University
Mentor: Barbara Shinn-Cunningham
University and Department: Carnegie Mellon University
Project Description:
Auditory selective attention, the ability to focus on specific sounds while ignoring competing sounds, enables communication in complex auditory environments. Previous studies have demonstrated that attention strongly modulates cortical representations of sound, but whether and where this modulation occurs in subcortical structures remains unclear.
In this project, Yunshu Li studied the effects of auditory selective attention on different levels of the auditory processing pathway using a combination of behavioral, computational, signal-processing, and non-invasive neuroimaging methods. Specifically, Yunshu collected and analyzed electroencephalography (EEG) data to evaluate event-related potentials (ERPs), an index of cortical activity, as well as auditory brainstem responses (ABRs, subcortical responses to sound) during a selective listening task. Subjects attend to a 3-note melody presented to one ear in one range of pitches while ignoring an interleaved, competing melody played to the other ear in a different pitch range. Melodies were composed of pitch-evoking pseudo-tones formed by convolving a periodic impulse train with a tone pip. Each tone pip within a pseudo-note elicits one ABR, while the pseudo-note onset elicits a strong cortical response.
Initial results produced clear cortical ERPs to each note and subcortical ABRs to each pip. From the cortical responses, we analyzed how attention modulated both the ERP phase and the inter-trial phase coherence (ITPC) at 1.5 Hz. We observed that attention enhanced the ERP evoked by the note onset when it was the “target” stream (as quantified by the cortical P1-N1 peak difference). Additionally, the ITPC showed peaks at the within-melody repetition rate of 1.5 Hz. Importantly, the best-performing listeners showed nearly a 180-degree phase separation between conditions. We also found robust ABRs evoked by each tone pip, and consistent with previous results in the lab, we also see a post-wave V peak in the ABR that appears to be modulated by attention.
Caleb McKinney
Undergraduate Institution: Rice University
Mentor: Leila Wehbe
University and Department: Carnegie Mellon University
Project Description:
Large neuroimaging datasets and advances in computer vision have enabled researchers to train machine learning models that predict the neural response of a subject to visual stimuli. Recent work has demonstrated that these brain encoders can be used to explore the functional organization of the human visual system by identifying regions of interest within the higher visual cortex and generating synthetic images that maximize expected activity in these voxels (BrainDiVE; Luo et al., 2023). Using a similar pipeline, Caleb explored how brain encoders can be coaxed to reveal subtle differences between subjects. First, he created brain encoding models for 8 subjects using CLIP and a trained linear probe. Then, he compared the learned encoder weights to localize differences in semantic selectivity across subjects. Finally, he discovered “individualized” images that are highly activating within a region of interest for one subject but not others. He ranked real and synthetic images according to this heuristic and discovered that individual preferences emerge. For example, in the food region, both real and synthetic individualized images for subject 1 depicted savory, pungent foods, while subject 3 exhibited an apparent preference for colorful, fruity items. Further research is necessary to validate that these encoding differences align with behavioral preferences. However, this work introduces promising, data-driven methods to understand neurodivergence and neuroplasticity within the visual system.
Prachi Mital
Undergraduate Institution: University of Pittsburgh
Mentor: Gelsy Torres-Oviedo
University and Department: University of Pittsburgh
Dravin Raj
Undergraduate Institution: University of Texas at Dallas
Mentor: Pulkit Grover
University and Department: Carnegie Mellon University
Project Description:
Addiction neuroscience is predominantly focused on the goal of understanding the mechanisms that underlie addictive behaviors. Generally, these behaviors are characterized as euphoric and rewarding sensations elicited through an unrestrained overflow of dopamine release. When neural systems overflow with dopamine, maybe due to genetic, psychological, or exogenous predispositions, an individual can incorporate harmful activities into their daily regimen, such as substance abuse, compulsive gambling, impulsive spending, all centered around this idea of taking excessively dangerous risks to satisfy that heightened appetite for reward. To decipher the brain’s risk and reward mechanisms behind addictive behaviors with the hopes of identifying novel patterns that will propel us in understanding and treating disorders of drug addiction, Dravin played a key role in implementing and utilizing five behavioral experiments that encompass a range of cognitive behaviors categorized under addiction. These five experiments were the Stroop, Balloon Analog Risk, Delay Discounting, Stop Signal Reaction Time, and Rutledge Passive Lottery tasks. Before the experiments would be used on hospitalized, epileptic patients to acquire data, he engaged in interface debugging using a Natus-Arduino to track trigger potentials upon pertinent stimuli. These experiments were then tested on one epileptic patient at Allegheny Hospital Network this summer and yielded new, intriguing data for the lab, specifically for the Delay Discounting, Stop Signal Reaction Time, and Rutledge Passive Lottery tasks.
Additionally, he also performed two types of analysis on Stroop data accumulated by the Grover Lab, consisting of five epileptic patients who performed the Stroop task while connected to stereo electroencephalography (sEEG) electrodes. First, he did a statistical analysis of the behavioral data, acquired by the Arduino. The results he found were consistent with previous literature, indicating that the task was implemented effectively and efficiently. Second, he did a waveform analysis of the neural data, acquired by the sEEG. This included the extrapolation of power spectral density (PSD) plots and phase-amplitude coupling (PAC) comodulograms of the congruent and incongruent trials for four mesolimbic brain regions: anterior cingulate cortex (ACC), orbitofrontal cortex (OFC), head of the hippocampus, and tail of the hippocampus. These preliminary visualizations open the door for deeper analysis into the individual frequency ranges where different patterns emerge for congruent and incongruent PSDs and PACs.
Sebastian Ruiz
Undergraduate Institution: Florida State University
Mentor: Tim Verstynen
University and Department: Carnegie Mellon University
Project Description:
Artificial intelligence agents often do not perform well, or generalize, to environments that slightly differ to the ones they were trained on. Humans are better at adapting to new environments, and it might be because humans use imagination to navigate novel experiences by leveraging past experiences to find invariances.
Sebastian sought to replicate a previously published study which indicated that an imagination-based artificial intelligence agent could offer better generalization to novel games and was unable to. Sebastian played an instrumental role by creating analyses that ruled out other possible causes for the incongruence in test results, when an error was found in the originally published tests.
Eric Tao
Undergraduate Institution: University of Pennsylvania
Mentor: Xaq Pitkow
University and Department: Carnegie Mellon University
Project Description:
Past research has identified repeated motifs in how neurons connect to one another in numerous model organisms, including in the C. elegans nervous system and the rat neocortex. These motifs include reciprocal pairs, pairs of neurons that connect to each other in both directions, and transitive triangles, triplets of neurons connected in a specific triangular pattern. Despite these findings, little is known about why these motifs recur and what purpose they may serve for brain function.
Eric Tao's summer project explored the hypothesis that these motifs help the brain integrate information over time to create an accurate picture of the outside world. To test this, they simulated noisy data representing changes in an external environment using a hidden Markov model. Then, they evaluated recurrent neural networks with different connectivity structures to see which ones were best at inferring the true state of the environment given the noisy data.
Their findings revealed that the two motifs had very different properties. While neural networks with reciprocal-pair patterns were highly effective at tracking variables in the external world which were related in a reciprocal-pair pattern, transitive triangles were much less useful for the analogous task. This suggests that for some types of patterns, this information-integration task requires a strong correspondence between the structure of the external world and the structure of the brain while for other types of patterns, the correspondence is much looser.
Future directions include investigating why these two motifs behave differently and how factors such as the sparsity of neural connections and the signal-to-noise ratio of the task impact this behavior
Zoe Wu
Undergraduate Institution: Harvard University
Mentor: Byron Yu
University and Department: Carnegie Mellon University
Project Description:
This project investigates how the primary motor cortex (M1) and prefrontal cortex (PFC) contribute to visually guided versus memory-guided reaching tasks. While visually guided reaches show higher success rates, the neural basis for this difference is not fully understood. We recorded neural activity from M1 and PFC during tasks where subjects reached for targets based on either visual cues or memory. Our findings show that M1 maintains consistent directional tuning across both task types, indicating its role in movement execution. In contrast, PFC activity changes significantly during memory-guided tasks, especially in the delay period, suggesting its involvement in cognitive processing and memory retention. This study sheds light on the distinct functions of M1 and PFC in integrating sensory information and memory to guide motor actions.