Human Feedback Research Articles

Toward Unbiased High-Quality Portraits through Latent-Space Evaluation.

Images, texts, voices, and signals can be synthesized by latent spaces in a multidimensional vector, which can be explored without the hurdles of noise or other interfering factors. In this paper, we present a practical use case that demonstrates the power of latent space in exploring complex realities such as image space. We focus on DaVinciFace, an AI-based system that explores the StyleGAN2 space to create a high-quality portrait for anyone in the style of the Renaissance genius Leonardo da Vinci. The user enters one of their portraits and receives the corresponding Da Vinci-style portrait as an output. Since most of Da Vinci's artworks depict young and beautiful women (e.g., "La Belle Ferroniere", "Beatrice de' Benci"), we investigate the ability of DaVinciFace to account for other social categorizations, including gender, race, and age. The experimental results evaluate the effectiveness of our methodology on 1158 portraits acting on the vector representations of the latent space to produce high-quality portraits that retain the facial features of the subject's social categories, and conclude that sparser vectors have a greater effect on these features. To objectively evaluate and quantify our results, we solicited human feedback via a crowd-sourcing campaign. Analysis of the human feedback showed a high tolerance for the loss of important identity features in the resulting portraits when the Da Vinci style is more pronounced, with some exceptions, including Africanized individuals.

Assimilating human feedback from autonomous vehicle interaction in reinforcement learning models

A significant challenge for real-world automated vehicles (AVs) is their interaction with human pedestrians. This paper develops a methodology to directly elicit the AV behaviour pedestrians find suitable by collecting quantitative data that can be used to measure and improve an algorithm's performance. Starting with a Deep Q Network (DQN) trained on a simple Pygame/Python-based pedestrian crossing environment, the reward structure was adapted to allow adjustment by human feedback. Feedback was collected by eliciting behavioural judgements collected from people in a controlled environment. The reward was shaped by the inter-action vector, decomposed into feature aspects for relevant behaviours, thereby facilitating both implicit preference selection and explicit task discovery in tandem. Using computational RL and behavioural-science techniques, we harness a formal iterative feedback loop where the rewards were repeatedly adapted based on human behavioural judgments. Experiments were conducted with 124 participants that showed strong initial improvement in the judgement of AV behaviours with the adaptive reward structure. The results indicate that the primary avenue for enhancing vehicle behaviour lies in the predictability of its movements when introduced. More broadly, recognising AV behaviours that receive favourable human judgments can pave the way for enhanced performance.

ChatGPT: perspectives from human-computer interaction and psychology.

The release of GPT-4 has garnered widespread attention across various fields, signaling the impending widespread adoption and application of Large Language Models (LLMs). However, previous research has predominantly focused on the technical principles of ChatGPT and its social impact, overlooking its effects on human-computer interaction and user psychology. This paper explores the multifaceted impacts of ChatGPT on human-computer interaction, psychology, and society through a literature review. The author investigates ChatGPT's technical foundation, including its Transformer architecture and RLHF (Reinforcement Learning from Human Feedback) process, enabling it to generate human-like responses. In terms of human-computer interaction, the author studies the significant improvements GPT models bring to conversational interfaces. The analysis extends to psychological impacts, weighing the potential of ChatGPT to mimic human empathy and support learning against the risks of reduced interpersonal connections. In the commercial and social domains, the paper discusses the applications of ChatGPT in customer service and social services, highlighting the improvements in efficiency and challenges such as privacy issues. Finally, the author offers predictions and recommendations for ChatGPT's future development directions and its impact on social relationships.

Human feedback enhanced autonomous intelligent systems: a perspective from intelligent driving

Artificial intelligence empowers the rapid development of autonomous intelligent systems (AISs), but it still struggles to cope with open, complex, dynamic, and uncertain environments, limiting its large-scale industrial application. Reliable human feedback provides a mechanism for aligning machine behavior with human values and holds promise as a new paradigm for the evolution and enhancement of machine intelligence. This paper analyzes the engineering insights from ChatGPT and elaborates on the evolution from traditional feedback to human feedback. Then, a unified framework for self-evolving intelligent driving (ID) based on human feedback is proposed. Finally, an application in the congested ramp scenario illustrates the effectiveness of the proposed framework.

Active convolutional neural networks sign language (ActiveCNN-SL) framework: a paradigm shift in deaf-mute communication

Real-time speech-to-text and text-to-speech technologies have significantly influenced the accessibility of communication for individuals who are deaf or mute. This research aims to assess the efficacy of these technologies in facilitating communication between deaf or mute individuals and those who are neither deaf nor mute. A mixed-method approach will incorporate qualitative and quantitative data collection and analysis techniques. The study will involve participants from deaf or mute and non-deaf or non-mute communities. The research will scrutinize the precision and efficiency of communication using these technologies and evaluate user experience and satisfaction. Furthermore, the study intends to pinpoint potential obstacles and limitations of these technologies and offer suggestions for enhancing their effectiveness in fostering inclusivity. The study proposes an active learning framework for sign language gesture recognition, termed Active Convolutional Neural Networks—Sign Language (ActiveCNN-SL). ActiveCNN-SL aims to minimize the labeled data required for training and augment the accuracy of sign language gesture recognition through iterative human feedback. This proposed framework holds the potential to enhance communication accessibility for deaf and mute individuals and encourage inclusivity across various environments. The proposed framework is trained using two primary datasets: (i) the Sign Language Gesture Images Dataset and (ii) the American Sign Language Letters (ASL)—v1. The framework employs Resnet50 and YoloV.8 to train the datasets. It has demonstrated high performance in terms of precision and accuracy. The ResNet model achieved a remarkable accuracy rate of 99.98% during training, and it also exhibited a validation accuracy of 100%, surpassing the baseline CNN and RNN models. The YOLOv8 model outperformed previous methods on the ASL alphabet dataset, achieving an overall mean average accuracy for all classes of 97.8%.

Optimizing large language models in digestive disease: strategies and challenges to improve clinical outcomes.

Large Language Models (LLMs) are transformer-based neural networks with billions of parameters trained on very large text corpora from diverse sources. LLMs have the potential to improve healthcare due to their capability to parse complex concepts and generate context-based responses. The interest in LLMs has not spared digestive disease academics, who have mainly investigated foundational LLM accuracy, which ranges from 25% to 90% and is influenced by the lack of standardized rules to report methodologies and results for LLM-oriented research. In addition, a critical issue is the absence of a universally accepted definition of accuracy, varying from binary to scalar interpretations, often tied to grader expertise without reference to clinical guidelines. We address strategies and challenges to increase accuracy. In particular, LLMs can be infused with domain knowledge using Retrieval Augmented Generation (RAG) or Supervised Fine-Tuning (SFT) with reinforcement learning from human feedback (RLHF). RAG faces challenges with in-context window limits and accurate information retrieval from the provided context. SFT, a deeper adaptation method, is computationally demanding and requires specialized knowledge. LLMs may increase patient quality of care across the field of digestive diseases,where physicians are often engaged in screening, treatment and surveillance for a broad range of pathologies for which in-context learning or SFT with RLHF could improve clinical decision-making and patient outcomes. However, despite their potential, the safe deployment of LLMs in healthcare still needs to overcome hurdles in accuracy, suggesting a need for strategies that integrate human feedback with advanced model training.

Bias-aware ranking from pairwise comparisons

Human feedback is often used, either directly or indirectly, as input to algorithmic decision making. However, humans are biased: if the algorithm that takes as input the human feedback does not control for potential biases, this might result in biased algorithmic decision making, which can have a tangible impact on people’s lives. In this paper, we study how to detect and correct for evaluators’ bias in the task of ranking people (or items) from pairwise comparisons. Specifically, we assume we are given pairwise comparisons of the items to be ranked produced by a set of evaluators. While the pairwise assessments of the evaluators should reflect to a certain extent the latent (unobservable) true quality scores of the items, they might be affected by each evaluator’s own bias against, or in favor, of some groups of items. By detecting and amending evaluators’ biases, we aim to produce a ranking of the items that is, as much as possible, in accordance with the ranking one would produce by having access to the latent quality scores. Our proposal is a novel method that extends the classic Bradley-Terry model by having a bias parameter for each evaluator which distorts the true quality score of each item, depending on the group the item belongs to. Thanks to the simplicity of the model, we are able to write explicitly its log-likelihood w.r.t. the parameters (i.e., items’ latent scores and evaluators’ bias) and optimize by means of the alternating approach. Our experiments on synthetic and real-world data confirm that our method is able to reconstruct the bias of each single evaluator extremely well and thus to outperform several non-trivial competitors in the task of producing a ranking which is as much as possible close to the unbiased ranking.

The Critical Robot: Impact of Performance Feedback on Intrinsic Motivation, Self-Esteem and Psychophysiology in Human–Robot Interaction

Social, anthropomorphic robots are increasingly used in professional work environments to collaborate with humans. However, little is known about how these robots affect human workers in performance-critical aspects, such as feedback. The present study investigates differences between the effects of a robot and a human feedback giver on self-esteem, intrinsic motivation, and psychophysiological reactions. Using a mixed model design for subjective data and a between-subject design for psychophysiological data, we tested 72 participants who performed a cognitive task on working memory, namely the 3-back task. The results indicate that people are more motivated to perform the task when receiving feedback from a robot, but their electrodermal activity and heart rate are higher after receiving positive feedback from a human. There is no difference in electrodermal activity following negative feedback from a human or a robot. Additional analyses show that individuals report feeling less comfortable and perceiving less social warmth when receiving feedback from a robot compared to a human. Furthermore, individuals exhibit higher skin conductance responses when perceiving greater social warmth in their interactions, regardless of whether their interaction partner is a human or a robot. The results suggest that social robots may serve as surrogates for social interaction. However, they seem to have less social presence, which leads to reduced psychophysiological reactions. This knowledge may be used to calibrate arousal in feedback situations.

Research on a Personalized Decision Control Algorithm for Autonomous Vehicles Based on the Reinforcement Learning from Human Feedback Strategy

To address the shortcomings of previous autonomous decision models, which often overlook the personalized features of users, this paper proposes a personalized decision control algorithm for autonomous vehicles based on RLHF (reinforcement learning from human feedback). The algorithm combines two reinforcement learning approaches, DDPG (Deep Deterministic Policy Gradient) and PPO (proximal policy optimization), and divides the control scheme into three phases including pre-training, human evaluation, and parameter optimization. During the pre-training phase, an agent is trained using the DDPG algorithm. In the human evaluation phase, different trajectories generated by the DDPG-trained agent are scored by individuals with different styles, and the respective reward models are trained based on the trajectories. In the parameter optimization phase, the network parameters are updated using the PPO algorithm and the reward values given by the reward model to achieve personalized autonomous vehicle control. To validate the control algorithm designed in this paper, a simulation scenario was built using CARLA_0.9.13 software. The results demonstrate that the proposed algorithm can provide personalized decision control solutions for different styles of people, satisfying human needs while ensuring safety.

Generative AI for scalable feedback to multimodal exercises

Detailed feedback on exercises helps learners become proficient but is time-consuming for educators and, thus, hardly scalable. This manuscript evaluates how well Generative Artificial Intelligence (AI) provides automated feedback on complex multimodal exercises requiring coding, statistics, and economic reasoning. Besides providing this technology through an easily accessible web application, this article evaluates the technology’s performance by comparing the quantitative feedback (i.e., points achieved) from Generative AI models with human expert feedback for 4,349 solutions to marketing analytics exercises. The results show that automated feedback produced by Generative AI (GPT-4) provides almost unbiased evaluations while correlating highly with (r = 0.94) and deviating only 6 % from human evaluations. GPT-4 performs best among seven Generative AI models, albeit at the highest cost. Comparing the models’ performance with costs shows that GPT-4, Mistral Large, Claude 3 Opus, and Gemini 1.0 Pro dominate three other Generative AI models (Claude 3 Sonnet, GPT-3.5, and Gemini 1.5 Pro). Expert assessment of the qualitative feedback (i.e., the AI’s textual response) indicates that it is mostly correct, sufficient, and appropriate for learners. A survey of marketing analytics learners shows that they highly recommend the app and its Generative AI feedback. An advantage of the app is its subject-agnosticism—it does not require any subject- or exercise-specific training. Thus, it is immediately usable for new exercises in marketing analytics and other subjects.

Using ChatGPT in Software Requirements Engineering: A Comprehensive Review

Large language models (LLMs) have had a significant impact on several domains, including software engineering. However, a comprehensive understanding of LLMs’ use, impact, and potential limitations in software engineering is still emerging and remains in its early stages. This paper analyzes the role of large language models (LLMs), such as ChatGPT-3.5, in software requirements engineering, a critical area in software engineering experiencing rapid advances due to artificial intelligence (AI). By analyzing several studies, we systematically evaluate the integration of ChatGPT into software requirements engineering, focusing on its benefits, challenges, and ethical considerations. This evaluation is based on a comparative analysis that highlights ChatGPT’s efficiency in eliciting requirements, accuracy in capturing user needs, potential to improve communication among stakeholders, and impact on the responsibilities of requirements engineers. The selected studies were analyzed for their insights into the effectiveness of ChatGPT, the importance of human feedback, prompt engineering techniques, technological limitations, and future research directions in using LLMs in software requirements engineering. This comprehensive analysis aims to provide a differentiated perspective on how ChatGPT can reshape software requirements engineering practices and provides strategic recommendations for leveraging ChatGPT to effectively improve the software requirements engineering process.

A Framework for Enhancing Behavioral Science Research with Human-Guided Language Models

Many behavioral science studies result in large amounts of unstructured data sets that are costly to code and analyze, requiring multiple reviewers to agree on systematically chosen concepts and themes to categorize responses. Large language models (LLMs) have potential to support this work, demonstrating capabilities for categorizing, summarizing, and otherwise organizing unstructured data. In this paper, we consider that although LLMs have the potential to save time and resources performing coding on qualitative data, the implications for behavioral science research are not yet well understood. Model bias and inaccuracies, reliability, and lack of domain knowledge all necessitate continued human guidance. New methods and interfaces must be developed to enable behavioral science researchers to efficiently and systematically categorize unstructured data together with LLMs. We propose a framework for incorporating human feedback into an annotation workflow, leveraging interactive machine learning to provide oversight while improving a language model's predictions over time.

Web-based diagnostic platform for microorganism-induced deterioration on paper-based cultural relics with iterative training from human feedback

PurposePaper-based artifacts hold significant cultural and social values. However, paper is intrinsically fragile to microorganisms, such as mold, due to its cellulose composition, which can serve as a microorganisms’ nutrient source. Mold not only can damage papers’ structural integrity and pose significant challenges to conservation works but also may subject individuals attending the contaminated artifacts to health risks. Current approaches for strain identification usually require extensive training, prolonged time for analysis, expensive operation costs, and higher risks of secondary damage due to sampling. Thus, in current conservation practices with mold-contaminated artifacts, little pre-screening or strain identification was performed before mold removal, and the cleaning techniques are usually broad-spectrum rather than strain-specific. With deep learning showing promising applications across various domains, this study investigated the feasibility of using a convolutional neural network (CNN) for fast in-situ recognition and classification of mold on paper.MethodsMolds were first non-invasively sampled from ancient Xuan Paper-based Chinese books from the Qing and Ming dynasties. Strains were identified using molecular biology methods and the four most prevalent strains were inoculated on Xuan paper to create mockups for image collection. Microscopic images of the molds as well as their stains situated on paper were collected using a compound microscope and commercial microscope lens for cell phone cameras, which were then used for training CNN models with a transfer learning scheme to perform the classification of mold. To enable involvement and contribution from the research community, a web interface that actuates the process while providing interactive features for users to learn about the information of the classified strain was constructed. Moreover, a feedback functionality in the web interface was embedded for catching potential classification errors, adding additional training images, or introducing new strains, all to refine the generalizability and robustness of the model.Results & ConclusionIn the study, we have constructed a suite of high-confidence classification CNN models for the diagnostic process for mold contamination in conservation. At the same time, a web interface was constructed that allows recurrently refining the model with human feedback through engaging the research community. Overall, the proposed framework opens new avenues for effective and timely identification of mold, thus enabling proactive and targeted mold remediation strategies in conservation.

Minimizing Negative Side Effects in Cooperative Multi-Agent Systems using Distributed Coordination

Autonomous agents operating in real-world environments frequently encounter undesirable outcomes or negative side effects (NSEs) when working collaboratively alongside other agents. Even when agents can execute their primary task optimally when operating in isolation, their training may not account for potential negative interactions that arise in the presence of other agents. We frame the challenge of minimizing NSEs as a lexicographic decentralized Markov decision process in which we assume independence of rewards and transitions with respect to the primary assigned tasks, but recognize that addressing negative side effects creates a form of dependence among the agents. We present a lexicographic Q-learning approach to mitigate the NSEs using human feedback models while maintaining near-optimality with respect to the assigned tasks---up to some given slack. Our empirical evaluation across two domains demonstrates that our collaborative approach effectively mitigates NSEs, outperforming non-collaborative methods.

A fully autonomous robotic ultrasound system for thyroid scanning

The current thyroid ultrasound relies heavily on the experience and skills of the sonographer and the expertise of the radiologist, and the process is physically and cognitively exhausting. In this paper, we report a fully autonomous robotic ultrasound system, which is able to scan thyroid regions without human assistance and identify malignant nod- ules. In this system, human skeleton point recognition, reinforcement learning, and force feedback are used to deal with the difficulties in locating thyroid targets. The orientation of the ultrasound probe is adjusted dynamically via Bayesian optimization. Experimental results on human participants demonstrated that this system can perform high-quality ultrasound scans, close to manual scans obtained by clinicians. Additionally, it has the potential to detect thyroid nodules and provide data on nodule characteristics for American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) calculation.

InstructPatentGPT: training patent language models to follow instructions with human feedback

AbstractIn this research, patent prosecution is conceptualized as a system of reinforcement learning from human feedback. The objective of the system is to increase the likelihood for a language model to generate patent claims that have a higher chance of being granted. To showcase the controllability of the language model, the system learns from granted patents and pre-grant applications with different rewards. The status of “granted” and “pre-grant” are perceived as labeled human feedback implicitly. In addition, specific to patent drafting, the experiments in this research demonstrate the model’s capability to learn from adjusting claim length and inclusion of limiting terms for narrowing claim scope. As proof of concept, the experiments focus on claim ones only and the training data originates from a patent dataset tailored specifically for artificial intelligence. Although the available human feedback in patent prosecution are limited and the quality of generated patent text requires improvement, the experiments following the 3-stage reinforcement learning from human feedback have demonstrated that generative language models are capable of reflecting the human feedback or intent in patent prosecution. To enhance the usability of language models, the implementation in this research utilizes modern techniques that enable execution on a single consumer-grade GPU. The demonstrated proof of concept, which reduces hardware requirements, will prove valuable in the future as more human feedback in patent prosecution become available for broader use, either within patent offices or in the public domain.

ConceptGlassbox: Guided Concept-Based Explanation for Deep Neural Networks

Various industries and fields have utilized machine learning models, particularly those that demand a significant degree of accountability and transparency. With the introduction of the General Data Protection Regulation (GDPR), it has become imperative for machine learning model predictions to be both plausible and verifiable. One approach to explaining these predictions involves assigning an importance score to each input element. Another category aims to quantify the importance of human-understandable concepts to explain global and local model behaviours. The way concepts are constructed in such concept-based explanation techniques lacks inherent interpretability. Additionally, the magnitude and diversity of the discovered concepts make it difficult for machine learning practitioners to comprehend and make sense of the concept space. To this end, we introduce ConceptGlassbox, a novel local explanation framework that seeks to learn high-level transparent concept definitions. Our approach leverages human knowledge and feedback to facilitate the acquisition of concepts with minimal human labelling effort. The ConceptGlassbox learns concepts consistent with the user’s understanding of a concept’s meaning. It then dissects the evidence for the prediction by identifying the key concepts the black-box model uses to arrive at its decision regarding the instance being explained. Additionally, ConceptGlassbox produces counterfactual explanations, proposing the smallest changes to the instance’s concept-based explanation that would result in a counterfactual decision as specified by the user. Our systematic experiments confirm that ConceptGlassbox successfully discovers relevant and comprehensible concepts that are important for neural network predictions.

Bibliometrically and systematically analyzing automated writing evaluation for English learning

Automated writing evaluation is highly discussed in artificial intelligence for English learning. It is necessary to explore the effect of automated writing evaluation on learning English as a second language. This study combined bibliometric analysis and systematic review to explore the use of automated writing evaluation for learning English as a second language. VOSviewer was used to identify the highly discussed topics, the top ten cited authors, organizations, countries, references, and sources in the studies on automated writing evaluation. Fifty-six peer-reviewed articles were selected according to the Preferred Reporting Items for Systematic Review and Meta-analysis Protocols. The analysis revealed that automated writing evaluation is helpful, but its effectiveness varies according to the types of feedback, and it cannot replace the role of human feedback yet. Teachers' roles are significant in integrating automated writing evaluation into the classroom. Future research could focus on the specific ways to integrate automated writing evaluation into the classroom better.

Reinforcement Learning from AI Feedback A Review

Reinforcement Learning from AI Feedback (RLAIF) is a big step forward compared to Reinforcement Learning from Human Feedback (RLHF). It's especially useful for large language models like GPT-4. RLAIF is better because it can handle more data at scale and is more efficient. It uses AI-generated feedback instead of human feedback. This shift to AI-generated feedback enhances the efficiency and speed of training AI systems. Additionally, RLAIF optimizes the AI's ability to align with desired outcomes, although it may not directly improve understanding human preferences. RLAIF uses a Preference Model (PM) that follows constitutional principles. This ensures that AI responses are ethical, safe, and high -quality. The constitution sets rules for AI decision-making. It makes sure AI follows ethical and social standards. This is important as AI keeps evolving. RLAIF is moving towards an automated, moral feedback system focusing on responsible AI governance and ethical guidelines.

Large language models for biomedicine: foundations, opportunities, challenges, and best practices.

Generative large language models (LLMs) are a subset of transformers-based neural network architecture models. LLMs have successfully leveraged a combination of an increased number of parameters, improvements in computational efficiency, and large pre-training datasets to perform a wide spectrum of natural language processing (NLP) tasks. Using a few examples (few-shot) or no examples (zero-shot) for prompt-tuning has enabled LLMs to achieve state-of-the-art performance in a broad range of NLP applications. This article by the American Medical Informatics Association (AMIA) NLP Working Group characterizes the opportunities, challenges, and best practices for our community to leverage and advance the integration of LLMs in downstream NLP applications effectively. This can be accomplished through a variety of approaches, including augmented prompting, instruction prompt tuning, and reinforcement learning from human feedback (RLHF). Our focus is on making LLMs accessible to the broader biomedical informatics community, including clinicians and researchers who may be unfamiliar with NLP. Additionally, NLP practitioners may gain insight from the described best practices. We focus on 3 broad categories of NLP tasks, namely natural language understanding, natural language inferencing, and natural language generation. We review the emerging trends in prompt tuning, instruction fine-tuning, and evaluation metrics used for LLMs while drawing attention to several issues that impact biomedical NLP applications, including falsehoods in generated text (confabulation/hallucinations), toxicity, and dataset contamination leading to overfitting. We also review potential approaches to address some of these current challenges in LLMs, such as chain of thought prompting, and the phenomena of emergent capabilities observed in LLMs that can be leveraged to address complex NLP challenge in biomedical applications.