Safe-Embed: Unveiling the Safety-Critical Knowledge of Sentence Encoders

Despite the impressive capabilities of Large Language Models (LLMs) in various tasks, their vulnerability to unsafe prompts remains a critical issue. These prompts can lead LLMs to generate responses on illegal or sensitive topics, posing a significant threat to their safe and ethical use. Existing approaches attempt to address this issue using classification models, but they have several drawbacks. With the increasing complexity of unsafe prompts, similarity search-based techniques that identify specific features of unsafe prompts provide a more robust and effective solution to this evolving problem. This paper investigates the potential of sentence encoders to distinguish safe from unsafe prompts, and the ability to classify various unsafe prompts according to a safety taxonomy. We introduce new pairwise datasets and the Categorical Purity (CP) metric to measure this capability. Our findings reveal both the effectiveness and limitations of existing sentence encoders, proposing directions to improve sentence encoders to operate as more robust safety detectors. Our code is available at https://github.com/JwdanielJung/Safe-Embed.

Beyond Binary Gender Labels: Revealing Gender Biases in LLMs through Gender-Neutral Name Predictions

Name-based gender prediction has traditionally categorized individuals as either female or male based on their names, using a binary classification system. That binary approach can be problematic in the cases of gender-neutral names that do not align with any one gender, among other reasons. Relying solely on binary gender categories without recognizing gender-neutral names can reduce the inclusiveness of gender prediction tasks. We introduce an additional gender category, i.e., "neutral", to study and address potential gender biases in Large Language Models (LLMs). We evaluate the performance of several foundational and large language models in predicting gender based on first names only. Additionally, we investigate the impact of adding birth years to enhance the accuracy of gender prediction, accounting for shifting associations between names and genders over time. Our findings indicate that most LLMs identify male and female names with high accuracy (over 80%) but struggle with gender-neutral names (under 40%), and the accuracy of gender prediction is higher for English-based first names than non-English names. The experimental results show that incorporating the birth year does not improve the overall accuracy of gender prediction, especially for names with evolving gender associations. We recommend using caution when applying LLMs for gender identification in downstream tasks, particularly when dealing with non-binary gender labels.

Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder

Super-resolution (SR) and image generation are important tasks in computer vision and are widely adopted in real-world applications. Most existing methods, however, generate images only at fixed-scale magnification and suffer from over-smoothing and artifacts. Additionally, they do not offer enough diversity of output images nor image consistency at different scales. Most relevant work applied Implicit Neural Representation (INR) to the denoising diffusion model to obtain continuous-resolution yet diverse and high-quality SR results. Since this model operates in the image space, the larger the resolution of image is produced, the more memory and inference time is required, and it also does not maintain scale-specific consistency. We propose a novel pipeline that can super-resolve an input image or generate from a random noise a novel image at arbitrary scales. The method consists of a pretrained auto-encoder, a latent diffusion model, and an implicit neural decoder, and their learning strategies. The proposed method adopts diffusion processes in a latent space, thus efficient, yet aligned with output image space decoded by MLPs at arbitrary scales. More specifically, our arbitrary-scale decoder is designed by the symmetric decoder w/o up-scaling from the pretrained auto-encoder, and Local Implicit Image Function (LIIF) in series. The latent diffusion process is learnt by the denoising and the alignment losses jointly. Errors in output images are backpropagated via the fixed decoder, improving the quality of output images. In the extensive experiments using multiple public benchmarks on the two tasks i.e. image super-resolution and novel image generation at arbitrary scales, the proposed method outperforms relevant methods in metrics of image quality, diversity and scale consistency. It is significantly better than the relevant prior-art in the inference speed and memory usage.

Generating automatically labeled data for author name disambiguation: An iterative clustering method

To train algorithms for supervised author name disambiguation, many studies have relied on hand-labeled truth data that are very laborious to generate. This paper shows that labeled training data can be automatically generated using information features such as email address, coauthor names, and cited references that are available from publication records. For this purpose, high-precision rules for matching name instances on each feature are decided using an external-authority database. Then, selected name instances in target ambiguous data go through the process of pairwise matching based on the rules. Next, they are merged into clusters by a generic entity resolution algorithm. The clustering procedure is repeated over other features until further merging is impossible. Tested on 26,566 instances out of the population of 228K author name instances, this iterative clustering produced accurately labeled data with pairwise F1 = 0.99. The labeled data represented the population data in terms of name ethnicity and co-disambiguating name group size distributions. In addition, trained on the labeled data, machine learning algorithms disambiguated 24K names in test data with performance of pairwise F1 = 0.90 ~ 0.92. Several challenges are discussed for applying this method to resolving author name ambiguity in large-scale scholarly data.

A fast and integrative algorithm for clustering performance evaluation in author name disambiguation

Author name disambiguation results are often evaluated by measures such as Cluster-F, K-metric, Pairwise-F, Splitting & Lumping Error, and B-cubed. Although these measures have distinctive evaluation schemes, this paper shows that they can be calculated in a single framework by a set of common steps that compare truth and predicted clusters through two hash tables recording information about name instances with their predicted cluster indices and frequencies of those indices per truth cluster. This integrative calculation reduces greatly calculation runtime, which is scalable to a clustering task involving millions of name instances within a few seconds. During the integration process, B-cubed and K-metric are shown to produce the same precision and recall scores. In this framework, especially, name instance pairs for Pairwise-F are counted using a heuristic, surpassing a state-of-the-art algorithm in speedy calculation. Details of the integrative calculation are described with examples and pseudo-code to assist scholars to implement each measure easily and validate the correctness of implementation. The integrative calculation will help scholars compare similarities and differences of multiple measures before they select ones that characterize best the clustering performances of their disambiguation methods.

Effect of forename string on author name disambiguation

In author name disambiguation, author forenames are used to decide which name instances are disambiguated together and how much they are likely to refer to the same author. Despite such a crucial role of forenames, their effect on the performances of heuristic (string matching) and algorithmic disambiguation is not well understood. This study assesses the contributions of forenames in author name disambiguation using multiple labeled datasets under varying ratios and lengths of full forenames, reflecting real-world scenarios in which an author is represented by forename variants (synonym) and some authors share the same forenames (homonym). Results show that increasing the ratios of full forenames improves substantially the performances of both heuristic and machine-learning-based disambiguation. Performance gains by algorithmic disambiguation are pronounced when many forenames are initialized or homonym is prevalent. As the ratios of full forenames increase, however, they become marginal compared to the performances by string matching. Using a small portion of forename strings does not reduce much the performances of both heuristic and algorithmic disambiguation compared to using full-length strings. These findings provide practical suggestions such as restoring initialized forenames into a full-string format via record linkage for improved disambiguation performances.

ORCID-linked labeled data for evaluating author name disambiguation at scale

How can we evaluate the performance of a disambiguation method implemented on big bibliographic data? This study suggests that the open researcher profile system, ORCID, can be used as an authority source to label name instances at scale. This study demonstrates the potential by evaluating the disambiguation performances of Author-ity2009 (which algorithmically disambiguates author names in MEDLINE) using 3 million name instances that are automatically labeled through linkage to 5 million ORCID researcher profiles. Results show that although ORCID-linked labeled data do not effectively represent the population of name instances in Author-ity2009, they do effectively capture the 'high precision over high recall' performances of Author-ity2009. In addition, ORCID-linked labeled data can provide nuanced details about the Author-ity2009's performance when name instances are evaluated within and across ethnicity categories. As ORCID continues to be expanded to include more researchers, labeled data via ORCID-linkage can be improved in representing the population of a whole disambiguated data and updated on a regular basis. This can benefit author name disambiguation researchers and practitioners who need large-scale labeled data but lack resources for manual labeling or access to other authority sources for linkage-based labeling. The ORCID-linked labeled data for Author-tiy2009 are publicly available for validation and reuse.

Unifying Activation- and Timing-based Learning Rules for Spiking Neural Networks

For the gradient computation across the time domain in Spiking Neural Networks (SNNs) training, two different approaches have been independently studied. The first is to compute the gradients with respect to the change in spike activation (activation-based methods), and the second is to compute the gradients with respect to the change in spike timing (timing-based methods). In this work, we present a comparative study of the two methods and propose a new supervised learning method that combines them. The proposed method utilizes each individual spike more effectively by shifting spike timings as in the timing-based methods as wells as generating and removing spikes as in the activation-based methods. Experimental results showed that the proposed method achieves higher performance in terms of both accuracy and efficiency than the previous approaches.

BinaryDuo: Reducing Gradient Mismatch in Binary Activation Network by Coupling Binary Activations

Binary Neural Networks (BNNs) have been garnering interest thanks to their compute cost reduction and memory savings. However, BNNs suffer from performance degradation mainly due to the gradient mismatch caused by binarizing activations. Previous works tried to address the gradient mismatch problem by reducing the discrepancy between activation functions used at forward pass and its differentiable approximation used at backward pass, which is an indirect measure. In this work, we use the gradient of smoothed loss function to better estimate the gradient mismatch in quantized neural network. Analysis using the gradient mismatch estimator indicates that using higher precision for activation is more effective than modifying the differentiable approximation of activation function. Based on the observation, we propose a new training scheme for binary activation networks called BinaryDuo in which two binary activations are coupled into a ternary activation during training. Experimental results show that BinaryDuo outperforms state-of-the-art BNNs on various benchmarks with the same amount of parameters and computing cost.

The impact of imbalanced training data on machine learning for author name disambiguation

In supervised machine learning for author name disambiguation, negative training data are often dominantly larger than positive training data. This paper examines how the ratios of negative to positive training data can affect the performance of machine learning algorithms to disambiguate author names in bibliographic records. On multiple labeled datasets, three classifiers - Logistic Regression, Na\"ive Bayes, and Random Forest - are trained through representative features such as coauthor names, and title words extracted from the same training data but with various positive-negative training data ratios. Results show that increasing negative training data can improve disambiguation performance but with a few percent of performance gains and sometimes degrade it. Logistic Regression and Na\"ive Bayes learn optimal disambiguation models even with a base ratio (1:1) of positive and negative training data. Also, the performance improvement by Random Forest tends to quickly saturate roughly after 1:10 ~ 1:15. These findings imply that contrary to the common practice using all training data, name disambiguation algorithms can be trained using part of negative training data without degrading much disambiguation performance while increasing computational efficiency. This study calls for more attention from author name disambiguation scholars to methods for machine learning from imbalanced data.