Simultaneous localization and mapping (SLAM) has been an emerging research topic in the fields of robotics, autonomous driving, and unmanned aerial vehicles over the past thirty years. State of the art SLAM research is often inaccessible for undergraduate student researchers due to expensive hardware and difficult software setup. We present a cost-friendly vehicle research platform and a robust implementation of SLAM. Our SLAM algorithm fuses visual stereo image and 2D light detection and ranging (Lidar) data and uses loop closure for accurate odometry estimation. Our algorithm is benchmarked against other popular SLAM algorithms using the publicly available KITTI dataset and shown to be very accurate. For educational purposes, we publicly share the models and code presented in this work*.

With the increasing trend of neural network models towards larger structures with more layers, we expect a corresponding exponential increase in the number of possible architectures. In this paper, we apply a hybrid evolutionary search procedure to define the initialization and architectural parameters of convolutional networks, one of the first successful deep network models. We make use of stochastic diagonal Levenberg-Marquardt to accelerate the convergence of training, lowering the time cost of fitness evaluation. Using parameters found from the evolutionary search together with absolute value and local contrast normalization preprocessing between layers, we achieve the best known performance on several of the MNIST Variations, rectangles-image and convex image datasets.

A 4-bit adaptive differential pulse-code modulation (ADPCM) scheme applied to the sensor data of a Zigbee based wireless sensor network node is shown to decrease the energy consumption of the analog front-end of the node by 58%. Simulation results from an energy model of an 802.15.4 based analog front-end show that the energy consumed by the network node is inversely proportional to the number of bits used to encode the digital data. The quantization error of a 4-bit ADPCM scheme is on average -14 dB for low frequency data and only 3 dB higher than a traditional 8-bit PCM scheme. By modifying the modulation scheme in the software with no modification of the hardware, the lifetime of the node can be increased significantly with minimal modifications.

The emergence of the Internet as today’s primary medium of music distribution has brought about demands for fast and reliable ways to organize, access, and discover music online. To date, many applications designed to perform such tasks have risen to popularity; each relies on a specific form of music metadata to help consumers discover songs and artists that appeal to their tastes. Very few of these applications, however, analyze the signal waveforms of songs directly. This low-level representation can provide dimensions of information that are inaccessible by metadata alone. To address this issue, we have implemented signalbased measures of musical similarity that have been optimized based on their correlations with human judgments. Furthermore, multiple recommendation engines relying on these measures have been implemented. These systems recommend songs to volunteers based on other songs they find appealing. Blind experiments have been conducted in which volunteers rate the systems’ recommendations along with recommendations of leading online music discovery tools (Allmusic which uses genre labels, Pandora which uses musicological analysis, and Last.fm which uses collaborative filtering), random baseline recommendations, and personal recommendations by the first author. This paper shows that the signal-based engines perform about as well as popular, commercial, state-of-the-art systems.

Physicians have many questions when caring for patients, and frequently need to seek answers for their questions. Information retrieval systems (e.g., PubMed) typically return a list of documents in response to a user's query. Frequently the number of returned documents is large and makes physicians' information seeking "practical only 'after hours' and not in the clinical settings". Question answering techniques are based on automatically analyzing thousands of electronic documents to generate short-text answers in response to clinical questions that are posed by physicians. The authors address physicians' information needs and described the design, implementation, and evaluation of the medical question answering system (MedQA). Although our long term goal is to enable MedQA to answer all types of medical questions, currently, we implemented MedQA to integrate information retrieval, extraction, and summarization techniques to automatically generate paragraph-level text for definitional questions (i.e., "What is X?"). MedQA can be accessed at http://www.dbmi.columbia.edu/~yuh9001/research/MedQA.html.

Research has shown that answers do not exist in biomedical corpora for many questions posed by physicians. We have therefore developed a question filtering component that determines whether or not a posed question is answerable. Using 200 clinical questions that have been annotated by physicians to be answerable or unanswerable, we have explored the use of supervised machine-learning algorithms to automatically classify questions into one of these two categories. We also have incorporated semantic features from a large biomedical knowledge terminology. Our results show that incorporating semantic features in general enhances the performance of question classification and the best system is a probabilistic indexing system that achieves an 80.5% accuracy. Our analysis also shows that stop words may play an important role for separating Answerable from Unanswerable .

We are developing a biomedical question answering system. This paper describes our system's architecture and our question analysis component. Specifically, we have explored the use of various supervised machine learning approaches to filter out unanswerable questions based on physicians' annotations.

We present supervised machine-learning approaches to automatically classify medical questions based on a hierarchical evidence taxonomy created by physicians. We show that SVMs is the best classifier for this task and that a ladder approach, which incorporates the knowledge representation of the hierarchical evidence taxonomy, leads to the highest performance. We have explored the use of features from a large, robust biomedical knowledge resource, namely, the Unified Medical Language System (UMLS), and we have found that performance is generally enhanced by including these features in addition to bag-ofwords.

Attempts to use natural language processing (NLP) for text categorization and information retrieval (IR) have had mixed results. Nevertheless, there is a strong intuition that NLP is important, at least for some tasks. In this paper, we discuss a task involving captioned images for which the subject and the predicate are critical. The usefulness of NLP for this task is established in two ways. In addition to the standard method of introducing a new system and comparing its performance with others in the literature, we also present evidence from experiments with human subjects showing that NLP generally improves speed and accuracy.

This paper explores the use of a statistical technique known as density estimation to potentially improve the results of text categorization systems which label documents by computing similarities between documents and categories. In addition to potentially improving a system's overall accuracy, density estimation converts similarity scores to probabilities. These probabilities provide con dence measures for a system's predictions which are easily interpretable and could potentially help to combine results of various systems. We discuss the results of three complete experiments on three separate data sets applying density estimation to the results of a TF*IDF/Rocchio system, and we compare these results to those of many competing approaches.

Recently, there have been significant advances in several areas of language technology, including clustering, text categorization, and summarization. However, efforts to combine technology from these areas in a practical system for information access have been limited. In this paper, we present Columbia's Newsblaster system for online news summarization. Many of the tools developed at Columbia over the years are combined together to produce a system that crawls the web for news articles, clusters them on specific topics and produces multidocument summaries for each cluster.

This paper introduces a term weighting method for text categorization based on smoothing ideas borrowed from speech recognition. Empirical estimates of weights (likelihood ratios) become unstable when counts are small. Instead of estimating weights for individual words, as Naive Bayes does, words with similar features are grouped into bins, and a single weight is estimated for each bin. This weight is then assigned to all of the words in the bin. The bin-based method is intended for tasks where there is insufficient training data to estimate a separate weight for each word. Experiments show the bin-based method is highly competitive with other current methods. In particular, this method is most similar to Naive Bayes; it generally performs at least as well as Naive Bayes, and sometimes better.

Annotating photographs automatically with content descriptions facilitates organization, storage, and search over visual information. We present an integrated approach for scene classi cation that combines image-based and text-based approaches. On the text side, we use the text accompanying an image in a novel TF*IDF vector-based approach to classi cation. On the image side, we present a novel OF*IIF (object frequency) vector-based approach to classi cation. Objects are de ned by clustering of segmented regions of training images. The image based OF*IIF approach is synergistic with the text based TF*IDF approach. By integrating the TF*IDF approach and the OF*IIF approach, we achieved a classi cation accuracy of 86%. This is an improvement of approximately 12% over existing image classi ers, an improvement of approximately 3% over the TF*IDF image classi er based on textual information, and an improvement of approximately 4% over the OF*IIF image classi er based on visual information.

The rapid expansion of multimedia digital collections brings to the fore the need for classifying not only text documents but their embedded non-textual parts as well. We propose a model for basing classification of multimedia on broad, non-topical features, and show how information on targeted nearby pieces of text can be used to effectively classify photographs on a first such feature, distinguishing between indoor and outdoor images. We examine several variations to a TF*IDF-based approach for this task, empirically analyze their effects, and evaluate our system on a large collection of images from current news newsgroups. In addition, we investigate alternative classification and evaluation methods, and the effect that a secondary feature can have on indoor/outdoor classification. We obtain a classification accuracy of 82%, a number that clearly outperforms baseline estimates and competing image-based approaches and nears the accuracy of humans who perform the same task with access to comparable information.