Poster Presenters for the Graduate Center Computer Science Student Workshop 2015
- Exploring Semantic Representations for a Textual Entailment System – Michelle Renee Morales
- Reservoir Computing for Recurrent Neural Networks – Felix Grezes
- Will My Paper Be Influential? Determining and Predicting Citation Patterns – David Guy Brizan, Kevin Gallagher, Arnab Jahanghir, Theodore Brown
- Classification of Vehicle Parts in Unstructured 3D Point Clouds – Allan Zelener, Phillipos Mordohai, Ioannis Stamos
- Finish Your Haribo Before You Speak: Detecting Eating Events To Classify Food Type From Speech – Min Ma, Guozhen An, Rachel Rakov, Michael J Madden, Andrew Rosenberg
- Automatic Recognition of Unified Parkinson’s Disease Rating from Speech with Acoustic, i-Vector and Phonotactic Features – Guozhen An, David Guy Brizan, Hernisa Kacorri, Min Ma, Michelle Morales, Ali Raza Syed, Andrew Rosenberg
Given two sentences (T and H) a Recognizing Textual Entailment (RTE) computational system must determine whether T entails H. Certain linguistic phenomena complicate this task. One property researchers have considered is the effects of monotonicity. Some work has investigated whether monotonicity properties can be learned automatically. In addition, there exists a growing body of research supporting the ability of word vectors to capture semantic properties of words. This work builds upon these ideas and introduces a novel approach to learning downward entailing operators using word vector representations. We propose an unsupervised method for automatically detecting downward entailing operators (DEOs). We compare our approach with two previous approaches: Cheung and Penn (2012) and Danescu et al. (2009). Similar to previous work, we leverage the linguistic knowledge that negative polarity items (NPIs) tend to appear in DE contexts to help automatically detect DEOs from text. Different than previous approaches, our method is the first to use word vectors and a similarity measure to rank possible DEOs. Moreover, we are the first to explore the effects the size of a corpus may have on detection. We find our approach outperforms previous approaches in precision and is the least affected by data size.
Historically, research has focused on Recurrent Neural Networks (RNNs) because of their capacity to model dynamical system and/or their biological plausibility. However traditional neural network approaches, such has gradient descent, are either impossible or too slow to be applied successfully to RNNs. The Reservoir Computing paradigm, by separating the optimization of the network layer from the output layer, attempts to solve these issues. Since 2001, it has proven itself to be fast and effective in a variety of applications.
Based on citations, we define the influence of publications in three ways: the count of citations, the diversity of research areas of the citations and the longevity of citations. We then rank publications within a cohort of year and journal in order to normalize and define the most influential by rank. Using this technique, we take a machine learning approach — extract a number of features to predict the most influential — and we find that we are consistently able to predict the most influential publications above baseline (chance) performance.
Unprecedented amounts of 3D data can be acquired in urban environments, but their use for scene understanding is challenging due to varying data resolution and variability of objects in the same class. An additional challenge is due to the nature of the point clouds themselves, since they lack detailed geometric or semantic information that would aid scene understanding. In this paper we present a general algorithm for segmenting and jointly classifying object parts and the object itself. Our pipeline consists of local feature extraction, robust RANSAC part segmentation, part-level feature extraction, a structured model for parts in objects, and classification using state-of-the-art classifiers. We have tested this pipeline in a very challenging dataset that consists of real world scans of vehicles. Our contributions include the development of a segmentation and classification pipeline for objects and their parts; and a method for segmentation that is robust to the complexity of unstructured 3D point clouds, as well as a part ordering strategy for the sequential structured model and a joint feature representation between object parts.
Although research has been done in classifying common conditions under which speech is produced including speaker-state detection such as intoxication detection and automatic detection of laughter/filled pauses, there has been almost no research differentiating speech produced under multiple eating conditions. In this paper, we take a two-stage framework to deal with a 7-class (6 food types plus “No-Food”) classification problem. We first extend the baseline features to distinguish audio containing speech-while-eating from No-Food. We then sub-categorize speech-while-eating by first detecting eat-ing events, distinguishing “Eating-Only” and “Speaking-While- Eating” segments. We then classify each utterance using a combination of utterance-level acoustic features and segment-level food type predictions, exploring multiple combination strategies.
Automatic Recognition of Unified Parkinson’s Disease Rating from Speech with Acoustic, i-Vector and Phonotactic Features
Guozhen An, David Guy Brizan, Hernisa Kacorri, Min Ma, Michelle Morales, Ali Raza Syed, Andrew Rosenberg
Parkinson’s Disease is a neurodegenerative disease affecting millions of people globally, most of whom present difficulties producing speech sounds. In this paper, we describe a system to identify the degree to which a person suffers from the disease. We use a number of automatic phone recognition-based features and we augment these with i-vector features and utterance- level acoustic aggregations. On the Interspeech 2015 ComParE challenge corpus, we find that these features allow for prediction well above the challenge baseline, particularly under cross- validation evaluation.