Built stuff in search, deep learning, LLMs, and computer vision.
We are creating a unified framework to analyze both post-hoc and learning bias mitigation methods. We specifically investigate Bolukbasi et al., Manzini et al., and Maudslay et al. We also provide several metrics for comparison: Word Embedding Association Test (WEAT) - Caliskan et al. (2017) Relative Norm Distance (RND) - Garg et al. (2018), which has two variations (cosine and euclidean) Mean Average Cosine (MAC) - Manzini et al. (2019). Through this project, I gained familiarity with Word2Vec, GloVe, RNNs, LSTMs, Language Models, BERT, etc.
Currently, in a clinical setting, epilepsy patients are monitored via video electroencephalogram (EEG) tests. A video EEG records what the patient experiences on video tape while an EEG device records his or her brainwaves. While patients are typically monitored at all times, epileptic seizures are unpredictable, and it may take time for nurses to respond to the patient. While epileptic seizures are dangerous, patients may also be harmed through other events like a fall or choking. Moreover, there are currently no non-invasive methods for tracking the patient’s location during a seizure. Being able to track a patient in real-time with video EEG would be a promising innovation towards improving the quality of healthcare. I propose using state-of-the-art object detection models like Detectron2 combined with an out-of-distribution (OOD) approach to track patients in the ICU and find anomalies in their movements. See project blogs for more details.
Many robust artificial neural networks are loosely inspired by neurobiological systems. It is known that a form of learning must take place in the human body, which is a living neural network. In the brain, living neurons use synapses to connect to other neurons. Neurons aim to fire at the homeostatic rate, yet what force governs them to adaptively adjust synaptic weights, even with disturbances? The Hebbian-LMS algorithm pieces together the Least Mean Squares (LMS) algorithm of Widrow and Hoff and the extended Hebbian learning rules of Donald Hebb to achieve an unsupervised learning paradigm that is useful for automatic pattern recognition. Could the Hebbian-LMS algorithm be the mechanism by which natural learning takes place in the human body? In this work, I perform a mathematical derivation of the algorithm and conduct multiple robust simulations of the Hebbian-LMS algorithm for simple clustering tasks. Thus, I seek to enhance the argument for why Hebbian-LMS may be nature’s learning algorithm.
There is an increasingly recognized role of diet in the pathophysiology of inflammatory bowel disease (IBD). Unsurprisingly, patients’ questions regarding diet are the most frequent inquiries to providers. Data from randomized controlled trials, however, are limited. Furthermore, maintaining adherence to diets can be challenging and may not reflect real world scenarios. Social media provides an opportunity to examine patient perceptions of actual dietary interventions.
These are projects that I pursued during high school to explore various fields of interest.