Omar Claflin

CV list of pubs/presentations/posters (most are talks) on the bottom. Footnoted into seven descriptive project categories.

There are other projects I’ve worked on without communications not listed here. For instance, I did a private DARPA-funded project from 2020-2021 communicated solely via Army & DARPA reports, and relevant PMs; industry projects not communicated broadly; or me failing to follow-through on communicating an academic project. There are a lot of other projects/research experiences I didn’t lead myself not listed here as well.

Independent

¹ LLM Interpretability and Neural Representations: Independent research attempting to advance current interpretability methods for large language models, demonstrating that neural codes capture more than feature identity through dual encoding mechanisms. After reading Anthropic’s research blog, I wondered if the “neural code” of LLMs were capturing feature relationships alongside feature identity (which is what is focused on currently). I enhance Sparse Autoencoders (SAE, typical method in intrepretability) with a Factorization Model to examine interactions of features, including non-linear interactions. I follow-up with a couple kinds of behaviorial analysis to confirm, and repeat with an integrated joint (SAE-NFM) architecture. [More here: https://omarclaflin.com/llm-interpretability-project-dual-encoding-in-neural-network-representations/ ]

² Real-Time Assessment Using IRT and Signal Detection Theory: I had access to Renaissance’s very large dataset of K-12 student learning on hundreds of thousands of questions over millions of students. I developed (and deployed, & improved further) an improved student performance prediction model that outperforms current psychometric approaches, when tested on hundreds of millions of real datapoints. I first productionalized a comprehensive psychometrics model, tried various variants, but found importing Signal Detection Theory methods vastly improved the model performance (when measured by prediction accuracy, not model fit). I then used a modern ML approach to further improve accuracy from its initial algorithmic. Tons of design and production decisions were made for speed and stateless model execution. [I might write more on that later & add a link here.]

³ Large-Scale Attention Training Analysis: I had access to Lumosity’s massive dataset on their multitasking game (admittedly based on my developed laboratory game below). As is typical with any industry (or even academic lab) dataset, lots of data issues including events you wish were being captured. However, I was able to, through lots of custom data preprocessing, reconstruct interesting and accurate event data. I filtered for participants that had at least 100 sessions, excluded very poorly performing elderly (>70 yo, with ability below the lowest level in the game, which is adaptive), and ended up with close to 50k participants, across 100 sessions. I then created a custom, sensitive measure (but relatively straightforward, hopefully) of attentional deployment between the two tasks, across 100 sessions. I then was able to see different clustered patterns of ‘learning” during exposure to a new multitask paradigm, and draw a very statistically-solid (but only not causally strong) conclusion that those who focused on learning one task first (focused their attention on one task at to the detriment of the other, early on), learned faster in the long run, and hit higher multitasking ability levels in the latest stages. Which, colloquially, sounds right at least.

Post Doc

⁴ Decoding features of human memory neural code: I joined a lab with DARPA funding, studying human memory, with fMRI and VR, with experience using geometric analyses (MVPA). I designed a novel memory recall paradigm with 6 dimensions (potentially a couple more, I can’t remember) balanced with extensive Monte Carlo simulations in MATLAB to generate balanced, counter-balanced participant designs. I was attempting to decode neural activity during recall (fMRI) using a clever comparison paradigm (asking participants to recall one object, then another, preceded by one of six comparison prompts; so that we could get clean neural codes, at least for the first object). Objects were memorized via extensive practice by walking around in an immersive VR environment I designed and administered.. The idea was features of events/objects in human memory including: object identity, comparison goals, absolute distance, relative distance, time differences, different environments, and path directions, even on the same objects, would be decodable in the neural code.

⁵ Neuropharmacology and Brain Imaging: I had access to a very large collected dataset of very controlled, multiple-session, repeated within-subject fMRI scans with a repeated paradigm (which is pretty rare). Used an interactive version of MVPA (to my knowledge, I think I was the first to create this methodological variant) which allowed the application of MVPA to interaction effects (not just main effects) from studied features. In this case, the four features studied were: pharmaceutical dose, task goal, visual inputs, and emotional salience. From memory, the drug even at extremely low doses (nicotine), had, by, far the largest impact on visual cortical patterns, followed by instruction/goals (which makes sense from a top-down attention PoV), then the other two.
Colloquially, the ability of nicotine to help ‘organize’ disordered thinking in schizophrenics is well known, so the cortical pattern result makes sense. Because of the interaction ability of the method, we could see a hierarchy (dendrogram) of specific effects (such as nicotine*attention interactions showing a multiplicatively greater separation of emotional valence or visual stimuli type [in the neural code of the visual cortex], when on nicotine and/or attending, on the downstream features). This sounds technical and irrelevant but interactions are interesting because they point at potentiation type effects rather than simple additive effects. So in this case, the neural activity of the occipital cortex is structured the same [Goals –> Stimuli Type] but nicotine caused much greater separation of the neural code patterns, so basically, increasing the signal-to-noise, in essence.

Grad School

⁶ Multitasking, Aging, and Video Game Training: Comprehensive research program examining age-related cognitive decline in multitasking abilities and remediation through video game interventions. When I joined the lab, they had a proto-project with a sloppy game engine, no ERPs (clean stimuli-elicited EEG waves), no designed conditions, etc. I re-coded a driving game engine from scratch in C++ to better control for laboratory conditions like stimulus presentation timing, eye gaze control, controlling for visual noise, visual motion (and lots of other overlooked cognitive design choices), along with designing extensive but sensible experimental conditions, successfully getting ERPs on the paradigm with EEG collection, and something resembling an ERP in fMRI/EEG combined collection, designing the adaptivity algorithm for single-task and multitask to place participants right around 80% performance, and much more. The Nature paper tries to argue for a generalized cognitive ability that is improved, which I personally found somewhat debatable (I’m not credited with writing contributions on it). However, specific and very related perception & cognitive skills definitely improve, even on older adults, with an adaptive task practice.
Another thing that might be interesting is, on the custom multitask I designed, peak performance is around 24 yo before decline begins. Also, I get asked a lot, no gender effects were seen in the controlled multitask across hundreds of participants across all decades of life.
Also, I designed all the early Akili Interactive product algorithms in their first two years (first FDA approved video game), including a nifty multi-level thresholding system and a quick psychometric validation pipeline for future changes to the system. I mention it here because these were related to my work for this project, and was on the early 5-person team but left when they decided to pursue a therapeutic direction instead of a diagnostic one.

⁷ Attention and Visual Conflict Processing: Investigation of attentional mechanisms on multitasking using the above-described video game paradigm (fMRI, EEG, etc). I also found the Superior Parietal Lobule is a localized area responsible for individual differences in continuous multitasking (using fMRI), if multitasking performance measurement is well-controlled for individual differences on the single tasks. This finding has been duplicated by at least a dozen other studies since.
A variety of poster presentations that showed an appropriately adapted, complex  tracking task with appropriate posthoc measures, can detect multitask switching with millisecond precision. These correlated strongly with behavioral measures such as reaction time to the secondary task and neural measures associated with prefrontal cognitive control, across participant, when appropriately baselined across conditions. In an arxiv paper (gave up on submitting it when I left academics), I found evidence that attention acts as metamonitoring suppressing goal conflict within visual cortex, rather than the simply suppressing irrelevant perception signals, when carefully controlling timing and recording with EEG.

Med School

⁸ Spinal Cord Injury Pain Research: For a summer research project, I found a variety of seemingly disparate metabolic pathways in the literature not shown to interact, but sharing common substrates, indicating to me that maybe inhibiting COX-1 (through something simple like a simple NSAID [ibuprofen, aspirin, etc.]) may, through chaining metabolic pathways, prevent neuropathic receptor expression that is known to be responsible for neuropathic pain symptoms after things like spinal cord injury. I designed the study, executed nearly the entirety of it, on mice and rat models, by giving them NSAIDs and then spinal cord injuries, recording pain sensitivity, recovery, and genetic expression (via mRNA). My naive thought at the time was that this could be an example of “translational research” immediately improving post-spinal cord injury protocols to reduce incidence of neuropathic pain.

Undergrad

⁹ Protein Folding Research: Independent computational project (along with lots of theory help from John Osterhout) to predict protein folding using a novel folding approach beating results at the time (CASP 6??) coded in C++ on my dorm computer, thus justifying the cost of my custom gaming rig. Basically, instead of simulating all the water molecules which has the effect of ‘pushing’ the protein components together, we simulated a ‘hydrophobic collapse’ (Osterhout’s idea) by Monte Carlo-ing attractive forces between hydrophobic elements. I further refined this with a fairly simple secondary structure prediction algorithm (forcing components of the protein into ‘alpha helices’, ‘beta sheets’, etc, based on the amino acid sequence) with a relatively simple moving average window of Ramachandran angles to generate predictions with confidences of secondary structure (again, Osterhout’s idea; although I enhanced the algorithms). Collapsing these seemed to work well on small proteins, which is all I attempted before graduating.

Publications & Presentations

Claflin, O. Feature Integration Spaces: Joint Training Reveals Dual Encoding in Neural Network Representations. Preprint, submitted to AAAI 2026, Jun 30, 2025.¹

Claflin, O. Improving IRT Based Real-Time Assessment Using Signal Detection Theory. NCSA 2024, Jun 1, 2024.²

Claflin, O. Cognitive Performance & Neurotechnology. NeuroTechX, 2020.³,⁴,⁶,⁷

Claflin, O. A large-scale analysis of attentional deployment across one hundred sessions of adaptive multitask training. Cognitive Science Society, 2020.³

Claflin, O. Mechanisms of skill improvement of over 50,000 participants across 100 sessions of an attention training program. Society for Neuroscience, 2018.³

Claflin, O. Brain-training Game Design. Lumos Lab Headquarters, Feb 2018.⁶,⁷

Claflin, O. Attention Acts Non-Selectively to Suppress Goal Conflict in the Visual Cortex. arxiv.org, 2016.⁷

Claflin, O. Task Psychometrics & Multivariate Pattern Analytical fMRI Methods. Stanford Pan Lab, Sept 2016.⁴,⁵,⁶,⁷

Claflin, O. Using Representational Similarity Analysis and Support Vector Machines to Quantify Small Inhaled Nicotine Doses on Brain Activity. Stanford University Department of Psychiatry & Behavioral Psychology, Aug 2016.⁴,⁵

Claflin, O. Dissociating the Organizational Hierarchy of Cognition, Visual Stimuli, and Neuropharmacological Agents Using MVPA Approaches on Brain Activity. UCLA Semel Molecular Imaging Group, Jan 2016.⁵

Claflin, O. Quantifying the Effect of Neuropharmacological Dose on Global Brain Patterns by using Multidimensional Reconstruction from Representational Similarity Analysis Distances. Translational Neuroscience of Drug Abuse Training Group, Feb 2016.⁵

Claflin, O. Dissociating the Organizational Hierarchy of Cognition, Visual Stimuli, and Neuropharmacological Agents Using MVPA Approaches on Brain Activity. UCLA Department of Neurosurgery, Jan 2015.⁵

Claflin, O, Zanto, T., Gazzaley, A. Neural sources of performance decline during continuous multitasking. Cortex, June 2015.⁷

Claflin, O, Anguera, J.A., Zanto, T., & Gazzaley, A. Efficient task switching underlies optimal multitasking performance (March 2015). Cognitive Neuroscience Society meeting, San Francisco, CA.⁷

Claflin, O. Advancing Quantitative Psychology Methods with Interpretable Neural Activity. UCLA Department of Psychology, Oct 2014.⁶,⁷

Anguera, J.A., Boccanfuso, J., Rintoul, J.L., Claflin, O., Faraji, F., Janowich, J., Kong E., Laraburro, Y., Rolle, C., Johnston, E., & Gazzaley, A. (2013) Video game training enhances cognitive control in older adults. Nature (cover). 501: 97-101.⁶

Claflin, O, Anguera, J.A., Rolle, C., & Gazzaley, A. Video game play during fMRI to explore neural basis of speed of processing differences across early adult lifespan (2013). Entertainment Software & Cognitive Neurotherapeutics Society (ESCoNS) Conference, Los Angeles, CA.⁶

Claflin, O, Anguera, J.A., & Gazzaley, A. Perceptual and central interference in dual-task performance (2013). Cognitive Neuroscience Society meeting, San Francisco, CA.⁷

J.A. Anguera, J. Boccanfuso, J.L. Rintoul, O. Claflin, J. Avila, Y. Cristo, F. Faraji, E. Kong, R. Moustafa, C. Rolle, E. Johnston, A. Gazzaley. (2012, October). Multitasking costs across the adult lifespan and their remediation through video game training. Society for Neuroscience; New Orleans, LA.⁶

Anguera, J.A., Boccanfuso, J., Rintoul, J. L., Claflin, O., Kong, E., Cristo, Y., E., Faraji, F., Moustafa, R., Johnston, E., Gazzaley, A. (2011, September). Training age-related multitasking deficits in older adults through an action driving video game. Entertainment Software and Cognitive Neurotherapeutics Society; San Francisco, CA.⁶

Anguera, J.A., Rintoul, J.L., Claflin, O., Johnston, E., Faraji, F., Gazzaley, A. (2010, October). Age-related changes in distraction & multitasking during a driving video game. Society for Neuroscience; San Diego, CA.⁶

Claflin, O., Turner N., Gerovac, T., Miranpuri, G., Vemuganti, R., Resnick, D. (2005, June). Differential expression of nociceptive genes influence the pain behavior following spinal cord injury in adult rats. American Society for Neurochemistry Annual Meeting; Madison, WI.⁸

Claflin, O., Turner N., Gerovac, T., Miranpuri, G., Vemuganti, R., Resnick, D. (2005, November). Differential expression of nociceptive genes influence the pain behavior following spinal cord injury in adult rats. Shapiro Summer Research Symposium; Madison, WI.⁸

Claflin, Omar. (2004, April). Prediction of Protein Folding Using First Principles and a Monte Carlo Hydrophobic Collapse. University of Arizona Department of Biochemistry Research Symposium; Tucson, AZ.⁹

John J. Osterhout, Mary Craven, Omar Claflin, Dawn Barnes and Tony Kanavage. (2006, February). Protein Folding: Design and Disease. BIO5 at University of Arizona Life Sciences; Tucson, AZ.⁹

---