Urwa Muaz

This paper proposes a novel method to address gender bias in the neural language models. We introduce a penalty term to the objective function of the language model to penalize the discrimination against gender. This method benefits from being simple, intuitive and can be easily incorporated into any text generation model. The proposed model’s performance was evaluated using multiple fairness metrics as well as perplexity showing that this method, when trained with counterfactual data augmentation, outperforms all existing techniques in the literature.

I attended Yann Lecun's Deep Learning class at NYU last year, and this was my term project from it.

I was awarded the academic award for the class of 2019 at NYU CUSP. This also gave me the honor to address the esteemed audience at the convocation ceremony as a valedictorian speaker.

I proposed and presented a cost-effective automated system to measure (rather than estimate) the fare evasion in New York subways using security camera footage to MTA.

This was a global high school competition where about 90 countries participated.

We wanted to observe if the steering behaviors described are observed in urban settings. We collected video data from a few urban spaces around metro tech to observe the pedestrian mobility tracks in those spaces and used Computer Vision to perform the analysis.

We collected geo-tagged data from Flickr and Twitter and performed textual and visual sentiment analysis of different locations in New York.

We used health survey data to identify socioeconomic health and lifestyle conditions that influence smoking behavior. Additionally we investigated to what extent can smoking behavior be predicted on individual level based on these factors. We had some interesting findings which are listed in the provided link.

● Named entity recognition and a convolutional neural network
classifier pipeline was used to identify the sentences that mention datasets.
● Ngram and cosine similarity based pipeline to match mentions to known datasets.
● Document similarity approaches were used for research field identification.