Ezgi Can Ozan

Visual saliency by extended quantum cuts

In this study, we propose an unsupervised, state-of-the-art saliency map generation algorithm which is based on a recently proposed link between quantum mechanics and spectral graph clustering, Quantum Cuts. The proposed algorithm forms a graph among superpixels extracted from an image and optimizes a criterion related to the image boundary, local contrast and area information. Furthermore, the effects of the graph connectivity, superpixel shape irregularity, superpixel size and how to determine the affinity between superpixels are analyzed in detail. Furthermore, we introduce a novel approach to propose several saliency maps. Resulting saliency maps consistently achieves a state-of-the-art performance in a large number of publicly available benchmark datasets in this domain, containing around 18k images in total.

Filterbank learning for deep neural network based polyphonic sound event detection.

Deep learning techniques such as deep feedforward neural networks and deep convolutional neural networks have recently been shown to improve the performance in sound event detection compared to traditional methods such as Gaussian mixture models. One of the key factors of this improvement is the capability of deep architectures to automatically learn higher levels of acoustic features in each layer. In this work, we aim to combine the feature learning capabilities of deep architectures with the empirical knowledge of human perception. We use the first layer of a deep neural network to learn a mapping from a high-resolution magnitude spectrum to smaller amount of frequency bands, which effectively learns a filterbank for the sound event detection task. We initialize the first hidden layer weights to match with the perceptually motivated mel filterbank magnitude response. We also integrate this initialization scheme with context windowing by using an appropriately constrained deep convolutional neural network. The proposed method does not only result with better detection accuracy, but also provides insight on the frequencies deemed essential for better discrimination of given sound events.

K-Subspaces Quantization for Approximate Nearest Neighbor Search

Approximate Nearest Neighbor (ANN) search has become a popular approach for performing fast and efficient retrieval on very large-scale datasets in recent years, as the size and dimension of data grow continuously. In this paper, we propose a novel vector quantization method for ANN search which enables faster and more accurate retrieval on publicly available datasets. We define vector quantization as a multiple affine subspace learning problem and explore the quantization centroids on multiple affine subspaces. We propose an iterative approach to minimize the quantization error in order to create a novel quantization scheme, which outperforms the state-of-the-art algorithms. The computational cost of our method is also comparable to that of the competing methods.

Competitive Quantization for Approximate Nearest Neighbor Search

In this study, we propose a novel vector quantization algorithm for Approximate Nearest Neighbor (ANN) search, based on a joint competitive learning strategy and hence called as competitive quantization (CompQ). CompQ is a hierarchical algorithm, which iteratively minimizes the quantization error by jointly optimizing the codebooks in each layer, using a gradient decent approach. An extensive set of experimental results and comparative evaluations show that CompQ outperforms the-state-of-the-art while retaining a comparable computational complexity.

M-PCA Binary Embedding for Approximate Nearest Neighbor Search

Principal Component Analysis (PCA) is widely used within binary embedding methods for approximate nearest neighbor search and has proven to have a significant effect on the performance. Current methods aim to represent the whole data using a single PCA however, considering the Gaussian distribution requirements of PCA, this representation is not appropriate. In this study we propose using Multiple PCA (M-PCA) transformations to represent the whole data and show that it increases the performance significantly compared to methods using a single PCA.

Tut MUVIS image retrieval system proposal for MSR-Bing challenge 2014

This paper presents our system designed for MSR-Bing Image Retrieval Challenge @ ICME 2014. The core of our system is formed by a text processing module combined with a module performing PCA-assisted perceptron regression with random sub-space selection (P2R2S2). P2R2S2 uses Over-Feat features as a starting point and transforms them into more descriptive features via unsupervised training. The relevance score for each query-image pair is obtained by comparing the transformed features of the query image and the relevant training images. We also use a face bank, duplicate image detection, and optical character recognition to boost our evaluation accuracy. Our system achieves 0.5099 in terms of DCG25 on the development set and 0.5116 on the test set.

Pyramid encoding for fast additive quantization

The problem of approximate nearest neighbor (ANN) search in Big Data has been tackled with a variety of recent methods. Vector quantization based solutions have been maintaining the dominant position, as they operate in the original data space, better preserving inter-point distances. Additive quantization (AQ) in particular has pushed the state-of-the-art in search accuracy, but high computational costs of encoding discourage the practical application of the method. This paper proposes pyramid encoding, a novel technique, which can replace the original beam search to provide a significant complexity reduction at the cost of a slight decrease in retrieval performance. AQ with pyramid encoding is experimentally shown to obtain results comparable with the baseline method in accuracy, while offering significant computational benefits.

Joint K-Means quantization for Approximate Nearest Neighbor Search

Recently, Approximate Nearest Neighbor (ANN) Search has become a very popular approach for similarity search on large-scale datasets. In this paper, we propose a novel vector quantization method for ANN, which introduces a joint multi-layer K-Means clustering solution for determination of the codebooks. The performance of the proposed method is improved further by a joint encoding scheme. Experimental results verify the success of the proposed algorithm as it outperforms the state-of-the-art methods.

A vector quantization based k-NN approach for large-scale image classification

The k-nearest-neighbour classifiers (k-NN) have been one of the simplest yet most effective approaches to instance based learning problem for image classification. However, with the growth of the size of image datasets and the number of dimensions of image descriptors, popularity of k-NNs has decreased due to their significant storage requirements and computational costs. In this paper we propose a vector quantization (VQ) based k-NN classifier, which has improved efficiency for both storage requirements and computational costs. We test the proposed method on publicly available large scale image datasets and show that the proposed method performs comparable to traditional k-NN with significantly better complexity and storage requirements.

An Optimized k-NN Approach for Classification on Imbalanced Datasets with Missing Data

In this paper, we describe our solution for the machine learning prediction challenge in IDA 2016. For the given problem of 2-class classification on an imbalanced dataset with missing data, we first develop an imputation method based on k-NN to estimate the missing values. Then we define a tailored representation for the given problem as an optimization scheme, which consists of learned distance and voting weights for k-NN classification. The proposed solution performs better in terms of the given challenge metric compared to the traditional classification methods such as SVM, AdaBoost or Random Forests.