About Me

I am a data scientist at Apple on the FaceID team. While at Apple, my work has focused on understanding model limitations and developing policies based on machine learning models using large datasets. During graduate school, my research focused on deep learning classification algorithms for temporal data. In particular, I worked on developing extensions of neural network architectures to complex temporal datasets. Currently, I am interested in developing/applying model architectures for real world problems, explainable machine learning models, and working with large and messy datasets.

Research Projects

Imbalanced Multivariate Sequence Classification

It is known that standard machine learning methods perform poorly when the dataset is imbalanced. For example, if we’re predicting if an image is a dog or a cat, and 99% of the training set is composed of cat images, a model is unlikely to learn what a dog looks like. Techniques that are used for imbalanced datasets generally involve oversampling the minority data, ensembling the data, or generating synthetic minority data to supplement the training set. While methods exist for oversampling timeseries data, these methods are not suited for multivariate temporal data as these methods cannot account for correlations between features at different timesteps. We developed a Generative Adversarial Network based method for generating synthetic minority data for multivariate temporal data. This method significantly improve model classification accuracy.

Explainability of Recurrent Neural Networks (RNNs)

Neural network models have been shown to outperform standard machine learning methods on many classification tasks. However, they lack the transparency of standard classification methods such as decision trees. I am interested in comparing general explainability methods to RNN specific methods in order to determine which features that are important for model classification. This project is supported by an industry partner.

Set Based Models

Neural network architectures have been developed for timeseries classification. These include LSTMs, GRUs, and transformers. I am interested in understanding how these models can be used when considering a list of input data with no inherent order. This project is supported by an industry partner.