580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can Learning Vector Quantization algorithms be optimized for computational efficiency in large-scale and high-dimensional data?
This body of research investigates methods to enhance the speed and reduce the computational load of Learning Vector Quantization (LVQ) algorithms when processing large datasets or high-dimensional vectors, which is critical for their practical applicability in real-world scenarios, such as speech compression or image processing.
Key finding: Introduces a modified LVQ variant named PDLVQ, which implements a Partial Distance (PD) computation strategy to avoid unnecessary calculations during classification. Experiments show PDLVQ achieves up to 37% improvement in... Read more
Key finding: Presents an analysis of multiple exact Euclidean nearest neighbor search algorithms applicable to vector quantization and self-organizing maps, proposing a method that reduces computation by approximately 85% across various... Read more
Key finding: Designs and implements a highly parallel and partially dynamically reconfigurable hardware architecture (RP-LVQ) for LVQ networks on FPGA, demonstrating adaptability to different model topologies. Achieves throughput of over... Read more
Key finding: Proposes an efficient algorithm for vector quantizer design applicable to known probabilistic models or data-driven training. By avoiding variational methods and differentiation, the algorithm supports complex distortion... Read more
Key finding: Introduces a new competitive learning algorithm incorporating a 'conscience' mechanism that improves efficiency and convergence in neural network-based vector quantizer codebook design. The algorithm yields near-optimal... Read more
2. What are effective approaches for learning and optimizing quantization parameters within neural networks, and how do these impact LVQ-based models?
This research theme covers methods for learning quantization parameters such as step size, bitwidth, and codebook size directly during training of neural networks, enhancing the efficacy of discretization processes. Since LVQ models operate by quantizing input data to prototypes, the precision and adaptability of quantization directly influence model accuracy and robustness.
Key finding: Develops Learned Step Size Quantization (LSQ), a method that trains quantizer step sizes alongside network weights through refined gradient approximations sensitive to quantized state transitions. LSQ achieves... Read more
Key finding: Proposes Differentiable Quantization (DQ), a framework enabling learning of quantizer step size, dynamic range, and bitwidth by gradient descent with novel parametrizations minimizing training difficulties. The approach... Read more
Key finding: Introduces ZEROQ, a zero-shot quantization method that generates synthetic distilled datasets matching batch normalization statistics without requiring original training data. The framework supports uniform and... Read more
Key finding: Proposes KURE, a kurtosis regularization technique to reshape weight distributions toward uniformity during training, thereby improving model robustness to variations in quantization step sizes and policies. Models trained... Read more
Key finding: Presents Dynamic Vector Quantization (DVQ), which adaptively selects discretization tightness conditioned on input complexity by dynamically choosing the number of codes and codebook size per input. Theoretically shown to... Read more
3. How can Learning Vector Quantization be effectively applied and adapted for practical classification tasks in varied domains?
This theme focuses on empirical applications of LVQ algorithms to domain-specific problems, including feature extraction, data preprocessing, and evaluation techniques relevant to LVQ performance and usability in real-world classification challenges.
Key finding: Applies LVQ to classify two coffee bean types—Arabica and Robusta—using first-order texture features extracted from images. Achieves classification accuracy of 71% on training and 96% on test datasets with a learning rate of... Read more
Key finding: Evaluates LVQ variants (LVQ1, LVQ2.1, LVQ3, LVQX, OLVQ1) under different prototype initialization schemes and normalization methods (e.g., z-score, linear scaling). Finds that initialization and normalization significantly... Read more
Key finding: Implements LVQ for bacilli bacteria classification causing ARI using features such as area, perimeter, and shape factor from microscopic images. Demonstrates high training (97%) and test accuracy (86%) with low learning... Read more
Key finding: Though details are limited due to partial text, the work centers on developing an LVQ-based classification system within an educational game for learning the Balinese script. Early findings suggest LVQ can be adapted for... Read more