Recent Posts

If you’ve used, or are considering, AWS/Azure/GCloud for Machine Learning …

8 min read

You already know that building your own Deep Learning Computer is 10x cheaper than …

10 min read

Stanford, the #1 computer science school in the nation, is giving away its core computer science …

7 min read

Projects

Recognizing themes in Amazon reviews through Unsupervised Multi-Document Summarization

We propose a system for unsupervised extractive summarization using Deep Learning feature extractors combined with several different models: k-means, affinity propagation, DBSCAN, and PageRank. We also propose a system for unsupervised abstractive summarization using a Deep Learning model. These summarization models capture the major themes in all the reviews of a product and optionally report the number of users who mentioned the same theme, hence giving shoppers a sense of salience.
2019
Poster PDF Code

Automatic Graphic Generation from Text using Generative Adversarial Networks

A computer that can generate graphics in a desired style from text captions is very powerful and its applications in marketing, advertising, and gaming are incredibly useful. We present a Generative Styled Network (GSN), which is capable of taking as input a text caption and some style images, and then proceed to produce an image on a white background that matches both the text description and the style desired.
2018
Poster PDF Code Dataset

High Performance Question Answering using Natural Language Processing with Neural Networks

We trained a question-answering model for Wikipedia, using the Stanford Question Answering Dataset. We achieved state-of-the-art F1 score of 75.8% on the test set. We demonstrate that the model generalizes at 68-90% correct using our hand-crafted datasets in 3 other types of text: Top News, Financial News, and Corporate Annual Report. Code available upon request.
2018
Poster PDF Dataset

Deep Bin Picking using Reinforcement Learning

We use reinforcement learning to solve the challenge of a robot picking items from a deep bin. This task is challenging, as lifting one item may cause other items to be unintentially lifted as well. We show that Pybullet and Dex-Net can be used to create the simulation environment for this task and Q-learning significantly outperforms random and greedy algorithms.
2018
Poster PDF Code

Affordable Self Driving Cars and Robots with Semantic Segmentation

Image segmentation is critical for the camera system in self driving cars as it informs the car of object boundaries for precise navigation. We present image cropping as a method to speed up training in a Fully Convolutional Network and compare against softmax regression and maximum likelihood methods using the Cityscape dataset.
2017
Poster PDF Code Dataset

Contact