Current
1. Log-Polar Boundary Extraction (L-PoBE)
A biologically-inspired model for human visual perception of object boundaries. The biological visual system adopts a seemingly peculiar coordinate system, the complex-log polar coordinate system. We justify nature’s choice by showing the computational advantage of log-polar coordinate system in solving the fundamental problem of object boundary extraction.
Doreen Hii, Zygmunt Pizlo. (19-24 May, 2023). Contour Integration Using Boundary and Region Information. Abstract in Journal of Vision, 22(14), 3374. Poster presented at the 2023 Vision Sciences Society (VSS).
Doreen Hii, Zygmunt Pizlo. (1-2 June, 2022). Biologically-Inspired Log-Polar Boundary Extraction Model for Noisy
Colored Images. Abstract in Journal of Vision, 22(14), 3374. Poster presented at the 2022 Virtual Meeting of Vision Sciences Society (V-VSS).
2. Bilateral Symmetric 3D Reconstructor (BiS3D)
Humans see the world in 3D given 2D retinal images as input. Is there enough information in the 2D images that allow humans to quickly reconstruct the 3D visual world? In this project we explore different deep neural network architectures to exploit the invariance of 2D curves projected from 3D objects, as a step towards building a machine that sees the world as humans do.
Mark Beers, Doreen Hii, Zygmunt Pizlo. (17-22 May, 2024). The Role of Uncertain Perspective Information in Recovering 3D Symmetrical Shapes. Poster presented at the 2024 Meeting of Vision Sciences Society (VSS).
Doreen Hii. Learning to See 3D Planar Symmetrical Objects Using Invariance in a Single 2D Image. 2022.
3. Identifying the Governing Law of Perception [Project Website]
Humans all arrive at the same perception even though the visual world is being reconstructed separately. Therefore, there must be a conservation law that allows us to arrive at the same percept. We propose that Noether’s theorem is general enough to describe human perception, extending its applications beyond Physics to Cognitive Sciences, and potentially unifying the branches of Natural Sciences.
Past
1. MidGet: Estimating Gait Phase for Pedestrian Navigation Using a Central Pattern Generator (CPG) Model
How could we accurately estimate the position of a moving pedestrian when GPS is not available? This project used a central pattern generator (CPG) model, a neural circuit model controlling the generations of cyclic motions in human walking, to reverse engineer gait phases from noisy Inertial Measurement Units (IMUs) measurements. From the reconstructed gait phase, stance phase (the period of time where a foot contacts the ground) was labeled as having zero velocity; thus improving navigational performance at the same time correcting for sensor drift error.
Doreen Hii. MidGet: Estimating Gait Phase for Pedestrian Navigation Using a Central Pattern Generator
(CPG) Model. 2022.
2. Using Meaning Specificity to Aid Negation Handling in Lexicon-Based Sentiment Analysis
How do humans determine the sentiment (the degree of positively or negativity) of a negated expression (a sentence involving the word “no”)? We propose that negation has a stronger effect (and may even invert the sentiment to the opposite extreme) if the adjective being negated carries a general meaning. On the other hand, negation only shifts the sentiment slightly if the adjective being negated carries specific meaning.
Doreen Hii. Using Meaning Specificity to Aid Negation Handling in Sentiment Analysis. Honors Thesis. 2019.
3. Systematic Compositionality in Recurrent Neural Networks (RNNs)
Humans generate and understand infinitely many utterances because of the systematic compositionality of languages. You may have never heard of the utterance, “The blue unicorn drove a Ford on the freeway.” Nonetheless, you understood the sentence because you know how to combine the individual meanings to form a broader concept. What is needed for deep neural networks to arrive at this sophistication level of language understanding?
Brandon Cabrera Marquez, Doreen Hii. Systematic Compositionality in Recurrent Neural Networks (RNNs). 2019.
4. Modeling of Human Endorsement of Generic Sentences
Some generic statements are endorsed easily even if the statement is False; some True generic statements are harder to accept. Applying Bayesian inference for probabilistic reasoning using the Rational Speech Act framework, we demonstrated in addition to the prior prevalence Tessler & Goodman (2016), a increased sense of urgency may also increase the probability for a generic sentence to be endorsed.
Isaiah Cushman, Doreen Hii. Extension of Generic Model: Turning the knob of threshold prior. 2018.
5. Using Single-Channel ERP Analysis for Efficient, Affordable Neuroimaging
How low can you go on your budget for electroencephalogram (EEG) while still being able to perform neuroscientific research? We collected event related potential (ERP) using a custom-built single-channel EEG and analyzed the data using open source code, bringing the total cost < $100. Research funding may have a limit; love for research does not.
Marianne Chavez, Isaiah Cushman, Doreen Hii, Austin Sheagley, Sherin Stephen. Using Single-Channel ERP Analysis for Efficient, Affordable Neuroimaging. The 17th Annual Honors Transfer Council of California Student Research Conference. 2017.