June Zhang

Honggen Zhang, Xufeng Zhao, Igor Molybog et al.

Aligning large language models (LLMs) to human preferences is a crucial step in building helpful and safe AI tools, which usually involve training on supervised datasets. Popular algorithms such as Direct Preference Optimization (DPO) rely on pairs of AI-generated responses ranked according to human annotation. The response pair annotation process might bring human bias. Building a correct preference dataset is the costly part of the alignment pipeline. To improve annotation efficiency and quality in the LLMs alignment, we propose REAL: Response Embedding-based Alignment for LLMs, a strategy for constructing a high-quality training dataset that focuses on acquiring the less ambiguous preference pairs for labeling out of a set of response candidates. Our selection process is based on the similarity of embedding responses independently of prompts, which guarantees the selection process in an off-policy setting, avoiding adaptively measuring the similarity during the training. Experimental results on real-world dataset SHP2 and synthetic HH-RLHF benchmarks indicate that choosing dissimilar response pairs enhances the direct alignment of LLMs while reducing inherited labeling errors. The model aligned with dissimilar response pairs obtained a better margin and win rate on the dialogue task. Our findings suggest that focusing on distinct pairs can reduce the label error and improve LLM alignment efficiency, saving up to $65\%$ of annotators' work.

Mojtaba Abolfazli, Mohammad Zaeri Arimani, Anders Host-Madsen et al.

We consider the following problem: we have a large dataset of normal data available. We are now given a new, possibly quite small, set of data, and we are to decide if these are normal data, or if they are indicating a new phenomenon. This is a novelty detection or out-of-distribution detection problem. An example is in medicine, where the normal data is for people with no known disease, and the new dataset people with symptoms. Other examples could be in security. We solve this problem by training a bidirectional generative adversarial network (BiGAN) on the normal data and using a Gaussian graphical model to model the output. We then use universal source coding, or minimum description length (MDL) on the output to decide if it is a new distribution, in an implementation of Kolmogorov and Martin-Löf randomness. We apply the methodology to both MNIST data and a real-world electrocardiogram (ECG) dataset of healthy and patients with Kawasaki disease, and show better performance in terms of the ROC curve than similar methods.

Honggen Zhang, Xiangrui Gao, June Zhang et al.

Messenger RNA (mRNA)-based vaccines are accelerating the discovery of new drugs and revolutionizing the pharmaceutical industry. However, selecting particular mRNA sequences for vaccines and therapeutics from extensive mRNA libraries is costly. Effective mRNA therapeutics require carefully designed sequences with optimized expression levels and stability. This paper proposes a novel contextual language model (LM)-based embedding method: mRNA2vec. In contrast to existing mRNA embedding approaches, our method is based on the self-supervised teacher-student learning framework of data2vec. We jointly use the 5' untranslated region (UTR) and coding sequence (CDS) region as the input sequences. We adapt our LM-based approach specifically to mRNA by 1) considering the importance of location on the mRNA sequence with probabilistic masking, 2) using Minimum Free Energy (MFE) prediction and Secondary Structure (SS) classification as additional pretext tasks. mRNA2vec demonstrates significant improvements in translation efficiency (TE) and expression level (EL) prediction tasks in UTR compared to SOTA methods such as UTR-LM. It also gives a competitive performance in mRNA stability and protein production level tasks in CDS such as CodonBERT.

Mojtaba Abolfazli, Mohammad Zaeri Amirani, Anders Høst-Madsen et al.

Given a default distribution $P$ and a set of test data $x^M=\{x_1,x_2,\ldots,x_M\}$ this paper seeks to answer the question if it was likely that $x^M$ was generated by $P$. For discrete distributions, the definitive answer is in principle given by Kolmogorov-Martin-Löf randomness. In this paper we seek to generalize this to continuous distributions. We consider a set of statistics $T_1(x^M),T_2(x^M),\ldots$. To each statistic we associate its maximum entropy distribution and with this a universal source coder. The maximum entropy distributions are subsequently combined to give a total codelength, which is compared with $-\log P(x^M)$. We show that this approach satisfied a number of theoretical properties. For real world data $P$ usually is unknown. We transform data into a standard distribution in the latent space using a bidirectional generate network and use maximum entropy coding there. We compare the resulting method to other methods that also used generative neural networks to detect anomalies. In most cases, our results show better performance.

Honggen Zhang, June Zhang, Igor Molybog

We consider a contrastive learning approach to knowledge graph embedding (KGE) via InfoNCE. For KGE, efficient learning relies on augmenting the training data with negative triples. However, most KGE works overlook the bias from generating the negative triples-false negative triples (factual triples missing from the knowledge graph). We argue that the generation of high-quality (i.e., hard) negative triples might lead to an increase in false negative triples. To mitigate the impact of false negative triples during the generation of hard negative triples, we propose the Hardness and Structure-aware (\textbf{HaSa}) contrastive KGE method, which alleviates the effect of false negative triples while generating the hard negative triples. Experiments show that HaSa improves the performance of InfoNCE-based KGE approaches and achieves state-of-the-art results in several metrics for WN18RR datasets and competitive results for FB15k-237 datasets compared to both classic and pre-trained LM-based KGE methods.

Honggen Zhang, June Zhang

Monitoring news content automatically is an important problem. The news content, unlike traditional text, has a temporal component. However, few works have explored the combination of natural language processing and dynamic system models. One reason is that it is challenging to mathematically model the nuances of natural language. In this paper, we discuss how we built a novel dataset of news articles collected over time. Then, we present a method of converting news text collected over time to a sequence of directed multi-graphs, which represent semantic triples (Subject -> Predicate} ->Object). We model the dynamics of specific topological changes in these graphs using a set of multivariate count series, which we fit the discrete-time Hawkes process. With our real-world data, we show that the multivariate time series contain both dynamic information of how many articles/words were published each day and semantic information of the content of the articles. This yields novel insights into how news events are covered. We show with the experiment that our approach can be used to infer from a sequence of news articles if the articles were published by major or entertainment news outlets.

Mojtaba Abolfazli, Anders Host-Madsen, June Zhang et al.

A classic application of description length is for model selection with the minimum description length (MDL) principle. The focus of this paper is to extend description length for data analysis beyond simple model selection and sequences of scalars. More specifically, we extend the description length for data analysis in Gaussian graphical models. These are powerful tools to model interactions among variables in a sequence of i.i.d Gaussian data in the form of a graph. Our method uses universal graph coding methods to accurately account for model complexity, and therefore provide a more rigorous approach for graph model selection. The developed method is tested with synthetic and electrocardiogram (ECG) data to find the graph model and anomaly in Gaussian graphical models. The experiments show that our method gives better performance compared to commonly used methods.

Mojtaba Abolfazli, Anders Host-Madsen, June Zhang

This paper introduces a new method for model selection and more generally hyperparameter selection in machine learning. Minimum description length (MDL) is an established method for model selection, which is however not directly aimed at minimizing generalization error, which is often the primary goal in machine learning. The paper demonstrates a relationship between generalization error and a difference of description lengths of the training data; we call this difference differential description length (DDL). This allows prediction of generalization error from the training data alone by performing encoding of the training data. DDL can then be used for model selection by choosing the model with the smallest predicted generalization error. We show how this method can be used for linear regression and neural networks and deep learning. Experimental results show that DDL leads to smaller generalization error than cross-validation and traditional MDL and Bayes methods.