Mapping similarity to false memory
A computational cognitive-science study testing which representation of similarity best predicts when people remember words they never saw.
View projectDiya Gupta.
About
About me
Hello! I am Diya Gupta, a student at Los Altos High School. I am interested in how artificial intelligence represents meaning, how symbolic systems encode ideas, and how mathematics can make complex patterns precise. Outside of these interests, I enjoy baking, reading, hiking, and thrifting.
Selected work
Research projects
Projects in cognitive science, mathematics, accessibility, and computational memory.
PROMYS 2026
Six weeks of number theory, proof writing, and collaborative research on Diophantine approximations.
View projectSensory Memory Lab
An independent computational-memory project exploring how sensory cues, emotion, and learned personal associations shape what we remember.
View projectSecond Look — Read + Count
An accessibility app with writing support, number-sense checks, adaptive practice, and progress tracking for learners with dyslexia or dyscalculia-related difficulties.
View projectFlagship research · ASI / ISEF
Research in progress
What makes a memory feel true?
In the Deese–Roediger–McDermott paradigm, people study a list of related words—such as bed, rest, and dream—then often remember a closely related word like sleep that was never shown. My research asks what kind of similarity best explains that effect.
Research question
Which representation of similarity best predicts published human false-recall and false-recognition rates across 55 DRM word lists when the memory equation is held fixed?
The approach
One memory model.
Six ways to define similarity.
Each DRM list contains 15 studied words and one unpresented “critical lure.” I calculate the similarity between every studied word and its lure, then compare six representational spaces under the same global-matching framework.
False recall and false recognition are treated as separate continuous outcomes. The final comparison will use held-out predictions and prespecified metrics—RMSE, MAE, Pearson correlation, and Spearman correlation—so each representation is evaluated on equal terms.
- 01Association strengthHow often one word brings another to mind
- 02Latent semantic analysisMeaning learned from patterns across documents
- 03Static word embeddingsGloVe geometry from large-scale language use
- 04Transformer embeddingsDistilBERT representations of the experimental words
- 05Phonological similarityHow much the words sound alike
- 06Orthographic similarityHow much their written forms overlap
55published DRM lists
825studied-word / lure pairs
2human-memory outcomes
6similarity representations
Research timeline
Project timeline
This is a living record of the project. Each phase holds its planned goal and a results log that will grow as analyses are completed.
Phase 0Frame the question and build the extraction systemAugust 2026complete
Read into the DRM literature, define what a finished comparison must contain, and turn the published appendices into a reproducible data pipeline.
Results log
- Research question and confirmatory boundary documented.
- Repository, schema, validators, and independent data-entry workflow established.
- All 55 DRM lists and 825 studied-word rows extracted for verification.
Phase 1Freeze a trustworthy data foundationSeptember 2026complete
Reconcile independent entries, validate every list and outcome, preserve source provenance, and release an immutable analysis dataset with checksums.
Results log
- Verified Appendix A and Appendix B data released as a checksum-protected version.
- Additional published recall and recognition criterion datasets acquired and documented.
- Quality-control evidence and source provenance preserved alongside the data.
Phase 2Build the similarity spacesOctober 2026In progress
Give every representation the same interface, audit missing words, and compare the geometry of each space before any memory outcome is modeled.
Results log
- USF association strength, pretrained TASA LSA, and GloVe 6B 300d spaces implemented.
- Bare-word and neutral-carrier DistilBERT representations implemented with complete 825/825 pair coverage.
- Phonological and orthographic representations remain in development.
Phase 3Freeze the model and run the main comparisonNovember - early December 2026upcoming
Specify one global-matching equation before analysis, then compare held-out false-recall and false-recognition predictions across every completed space.
Results log
Results will be added here as this phase develops.
Phase 4Test lure-level predictions and dissociationsDecember 2026upcoming
Ask which specific critical lures are predicted to intrude, inspect residuals, and test whether recall and recognition favor different similarity spaces.
Results log
Results will be added here as this phase develops.
Phase 5Write the paper and build the posterJanuary 2027upcoming
Turn committed figures and tables into a full methods-and-results narrative, abstract, poster, and science-fair application.
Results log
Results will be added here as this phase develops.
Phase 6Present, revise, and publishFebruary - May 2027upcoming
Use judging questions to improve the explanation, carry the strongest result through fair season, and prepare the finished work for journal submission.
Results log
Results will be added here as this phase develops.
ASI · False memory
Research journal
Notes from our work: what we tried, what we learned, and what comes next.
Loading journal…
Independent research
Human Compression
Research journal
Notes from our work: what we tried, what we learned, and what comes next.
Loading journal…
PROMYS · Summer 2026
PROMYS
At the Program in Mathematics for Young Scientists, I spent six weeks at Boston University immersed in problems that did not come with a formula or an obvious first step. The experience taught me to treat uncertainty as an invitation: test small cases, look for structure, make a conjecture, and keep refining it until the proof holds.
01ExperimentBuild examples and test edge cases
02ConjectureTurn patterns into precise claims
03ChallengeSearch for counterexamples
04ProveExplain why the pattern must hold
Discovering structure in the integers
Daily Number Theory lectures and problem sets formed the center of the summer. Instead of being handed finished theorems, I used numerical experiments to find patterns, formulate claims, and develop proofs from first principles. Feedback from counselors and conversations with other students made proof-writing feel less like recording an answer and more like communicating an idea with precision.
Finding the best rational approximations
In my Expo Lab, my team studied Diophantine approximation: how fractions can come as close as possible to an irrational number. We combined the recursive structure of continued fractions with the geometry of Farey tessellations, tracing rational approximations as geodesics in the upper half-plane.
Using determinant-one matrices, we developed a geometric way to peel back layers of the tessellation and move toward increasingly accurate convergents. We then connected that picture to a proof of optimality: a fraction that improves on a continued-fraction convergent must have a larger denominator.
Read our final paperPROMYS changed my relationship with mathematics. I became more patient with problems that resisted an immediate solution, more willing to revise a promising idea, and more attentive to the difference between evidence and proof. Just as importantly, I learned that deep mathematical work is collaborative: the strongest ideas often emerged while comparing examples, questioning assumptions, and explaining a proof to someone else.
Explore mathematics at PROMYSIndependent research · In development
Sensory Memory LabSensory and emotional memory research
A familiar song, the smell of a family meal, or a glimpse of a place can bring an experience back with surprising force. I am building a transparent computational model to study how sensory overlap, emotion, time, and personal context work together during retrieval.
The goal is not to imitate an individual brain. It is to turn specific ideas about remembering into measurable, falsifiable predictions.
Research question
How much do learned cue associations and personal context improve predictions of remembering beyond sensory modality alone?
Encode
Represent an event through sensory features, attention, emotional conditions, and context.
Retain
Model how accessibility changes with elapsed time, rehearsal, and competing experiences.
Retrieve
Let a partial cue activate several candidate event traces, including a no-retrieval outcome.
Evaluate
Test the model with synthetic recovery, stress tests, and held-out human observations where suitable data exist.
The model
How the model works
Each event is stored as a trace with modality-specific features—such as sound, odor, visual content, and taste—plus its context, time, and observed encoding conditions. When a later cue appears, the model scores overlap only across genuinely shared information and combines it with contextual matching and a transparent retention rule.
A probabilistic competition then produces one of three outcomes: retrieve the intended event, retrieve a competing event, or retrieve nothing. Every term remains visible so I can test what changes when a cue is missing, conflicting, unfamiliar, or shared by several memories.
Measured separately
- AccessibilityWhether a cue retrieves the target event
- Retrieval timeHow long remembering takes
- Emotional responseIntensity and valence during retrieval
- VividnessThe subjective clarity of the recollection
- Detail accuracyAgreement with recorded truth, when truth exists
First empirical direction
Does personal meaning add predictive value?
The first planned analysis asks whether recorded personal-association variables improve held-out predictions of memory vividness beyond cue type, familiarity, rehearsal, and age group. Music- and food-evoked memory diaries offer a useful starting point because they separate cue familiarity from a person’s learned relationship to that cue.
Scientific boundary
Simulation is not validation.
Synthetic examples can show that the equations behave as intended, but they cannot establish how human autobiographical memory works. Empirical claims will use untouched participants or stimuli, appropriate baselines, uncertainty estimates, and failure cases. Retrieval probability, vividness, emotion, nostalgia, and accuracy will never be collapsed into one “memory strength” score.
Current stage
Research scope and candidate architecture defined. Next: choose the primary empirical emphasis, inspect the selected dataset and codebook, then build the first transparent simulator and baseline comparison.
Sensory Memory Lab
Research journal
Notes from our work: what we tried, what we learned, and what comes next.
Loading journal…
Accessibility app
Read + CountSecond Look
I built Second Look to support children who process words or numbers differently. The app pairs a dyslexia-friendly writing checker with a number-sense experience designed around skills that can be difficult for learners with dyscalculia.
California Senate Bill 1067
Early screening should lead to useful support.
Signed into law on September 27, 2026, SB 1067 requires California schools to screen kindergarten through second-grade students annually for early math difficulties beginning in the 2028–29 school year and to provide evidence-based support when students need it. The law complements California's existing K–2 reading-difficulties screening requirement. Second Look responds to the same need for earlier support, but it is a practice tool—not a diagnostic instrument or an official school screener.
Words
Flags misspellings and harder real-word mix-ups, then gives graduated hints so a student can identify and correct the error instead of receiving an automatic answer.
Numbers
Checks dot sense, symbolic magnitude, quantity–numeral mapping, and number-line estimation, then recommends practice based on the learner's weakest relative skill.
Practice and progress
Includes three number-sense activities and an adult dashboard showing accuracy, response time, recurring patterns, session history, and change over time.
Beyond research
Activities and volunteering
Robotics, swimming, and math tutoring.
Robotics · EagleStrike Team 114
I am a member of my school's FRC robotics team, EagleStrike Team 114, where I serve as a lead of code. I work on coding the various subsystems of the robot. I also have worked to teach code to new members of the team.
Theseus was the team's second robot for the 2025 REEFSCAPE season. Its software used Java and WPILib to control the intake, elevator, and end effector, run autonomous paths, and align the robot to the reef with front and rear cameras.
Swimming
I have been swimming since 4th grade. I began swimming for Palo Alto Stanford Aquatics. I now swim on my school team as well. My favorite event is the 100 breaststroke.
Buddies4Math volunteering
I volunteer with Buddies4Math to teach math to underprivileged elementary-school students in local schools. Through weekly after-school tutoring, I help students work through problems, build confidence, and see mathematics as something they can explore rather than something to fear.
Teaching has made me more attentive to how different students think. I have learned to explain the same idea in several ways, listen for where confusion begins, and celebrate the moment a student finds a path that makes sense to them.
Learn about Buddies4MathSelected writing