In this paper, we propose a multi-agent reinforcement learning approach, POCA-Mix, to achieve collaborative multi-target search with a visual drone swarm. The proposed approach leverages the benefits of curriculum learning and mixed credit assignment to guide the drone swarm in per-forming the search task with only local visual perception in a constrained 3D environment. To validate the performance of the proposed approach, we conducted simulation experiments with various combinations regarding the number of drones and targets. The results demonstrate that the proposed approach outperforms other baseline methods such as PPO and MA-POCA with a higher success rate in different scenarios.
We study Federated Learning (FL) as the promising technology to design the computer vision Machine Learning-based applications, which are robust to possible Data Quality (DQ) degradation in real Intelligent Transportation Systems (ITSs). As ITSs are composed of diverse data generation and communication units, DQ of images they produce and use may vary dramatically. We study the images of various quality distributed over the local clients and perform iterative FL training with two distinct aggregation strategies: Federated Averaging (FedAvg) and Geometric Median (GM). We consequently evaluate the models cross-trained over a single DQ cohort against other DQ categories. Then we analyze the image (traffic signs) recognition performance and its robustness toward the DQ degradation for each training cohort and aggregation strategy. Based on the results, we provide our recommendations on how to train more robust FL-based computer vision applications for ITSs.
Music generation has attracted growing interest with the advancement of deep generative models. However, generating music conditioned on textual descriptions, known as text-to-music, remains challenging due to the complexity of musical structures and high sampling rate requirements. Despite the task’s significance, prevailing generative models exhibit limitations in music quality, computational efficiency, and generalization ability. This paper introduces JEN-1, a universal high-fidelity model for text-to-music generation. JEN-1 is a diffusion model incorporating both autoregressive and non-autoregressive training in an end-to-end manner, enabling up to 48kHz high-fidelity stereo music generation. Through multi-task in-context learning, JEN-1 performs various generation tasks including text-guided music generation, music inpainting, and continuation. Evaluations demonstrate JEN-1’s superior performance over state-of-the-art methods in text-music alignment and music quality while maintaining computational efficiency. Our demo pages are available at https://jenmusic.ai/audio-demos
Current genomics interventions have limitations in accounting for cell stimuli and the dynamic response to intervention. Although genomic sequencing and analysis have led to significant advances in personalized medicine, the complexity of cellular interactions and the dynamic nature of the cellular response to stimuli pose significant challenges. These limitations can lead to chronic disease recurrence and inefficient genomic interventions. Therefore, it is necessary to capture the full range of cellular responses to develop effective interventions. This paper presents a game-theoretic model of the fight between the cell and intervention, demonstrating analytically and numerically why current interventions become ineffective over time. The performance is analyzed using melanoma regulatory networks, and the role of artificial intelligence in deriving effective solutions is described.
Advanced machine learning techniques are very powerful in predictive tasks. However, they are mostly weak in explaining the inference process and they are mostly treated as black-box models. Fuzzy Network (FN) is powerful white-box technique which is capable of dealing with complexity and linguistic uncertainty. In this paper, a method is introduced to optimise Rule Based Networks using Fuzzy C-Means (FCM) for rule reduction, Genetic Algorithms to tune the membership functions and Backward Selection to reduce the inputs and network branches. A case study in transport and telecommuting is used to illustrate the performance of the proposed method. The results show the FN ability to explain the internal process of decision making and its capabilities in transparency and interpretability as an Explainable AI method.
We appreciate well-functioning technology being able to also personalize its services. However, to protect privacy and avoid a potential misuse of personal data, we are encouraged to limit the amount of personal data we share through apps and Internet services. While some services do not really need all the data they ask us to provide, others depend on it to provide the best possible performance of its service. That regards systems that apply data in machine learning for tasks like medical diagnostics. Especially deep learning algorithms perform better by using a large amount of data and are now able to benefit from the large amount as well with limited training time given access to high-performance computing resources. This paper address and discuss the tradeoffs like the one we have between data sharing minimalization for increased privacy and data maximization for machine learning systems. Perspectives related to ethics, legal, and social issues are considered in the paper. There is no single conclusion on the challenge, but attention to it can increase the awareness that the best balance differs depending on the application addressed.
Federated Learning (FL) provides an approach for performing the collaborative training of AI models without compromising data privacy. However, traditional implementations of FL require complex deployment strategies, making it challenging to perform training across multiple data centers. To address this issue, this article presents a platform designated as Federated Learning Agent (FLAg) which allows users to delegate their federated learning tasks to an automated process. FLAg features digital pathology databases as well as multiple pretrained models. The security of FLAg is ensured through its deployment on an HPC system and bastion host. While FLAg itself only offers FL training services, users may perform data annotation and inference through its integration with a proprietary ALOVAS AI pathology platform for a complete end-to-end process which includes annotation, FL training, and inference.
Artificial intelligence (AI) is increasingly used in healthcare systems and applications (apps) with questions and debates on ethical issues and privacy risks. This research study explores and discusses the ethical challenges, privacy risks, and possible solutions related to protecting user data privacy in AI-enabled healthcare apps. The study is based on the healthcare app named Charlie in one of the fictional case studies designed by Princeton University to elucidate critical thinking and discussions on emerging ethical issues embracing AI.
Robustness and safety constraints are key requirements for AI systems. The robust constrained Markov decision process is a recent task-modelling framework that incorporates behavioural constraints and robustness to reinforcement learning systems. Earlier work proposed the robust constrained policy gradient (RCPG) algorithm, which robustifies either the value or the constraint and updates the worst-case distribution through constrained optimisation on a sorted value list. Highlighting potential downsides of RCPG such as not robustifying the full constrained objective and the lack of incremental learning, this paper introduces algorithms to robustify the Lagrangian and to learn incrementally using gradient descent over an adversarial policy. A theoretical analysis derives the Lagrangian policy gradient for the policy optimisation and the Lagrangian adversarial policy gradient for the adversary optimisation. Empirical experiments injecting perturbations in inventory management and safe navigation tasks demonstrate the benefit of these modifications, and combining both modifications yields the best overall performance.
Gene regulatory networks (GRNs) play crucial roles in various cellular processes, including stress response, DNA repair, and the mechanisms involved in complex diseases such as cancer. Biologists are involved in most biological analyses. Thus, quantifying their policies reflected in available biological data can significantly help us to better understand these complex systems. The primary challenges preventing the utilization of existing machine learning, particularly inverse reinforcement learning techniques, to quantify biologists' knowledge are the limitations and huge amount of uncertainty in biological data. This paper leverages the networklike structure of GRNs to define expert reward functions that contain exponentially fewer parameters than regular reward models. Numerical experiments using mammalian cell cycle and synthetic gene-expression data demonstrate the superior performance of the proposed method in quantifying biologists' policies.
Oral health is crucial to overall good health, and yet because dental services are expensive, time consuming and stressful for patients, dental problems are often left unaddressed, leading to the need for far more expensive and painful interventions down the road. The main goal of this project is using AI and deep learning to develop software to detect cavities and other dental problems and providing the corresponding app to improve dental health. We applied convolutional neural network (CNN) to find the teeth caries. Canny Edge CNN model performed the best training.
This paper describes The People’s Panel for AI – a mechanism to build public trust in AI products and services from conceptualization to deployment. To increase public awareness of how AI and data-driven systems are affecting the lives of ordinary people, a series of Artificial Intelligence Roadshows were delivered in community centers. Community members were recruited to the People’s Panel and completed two days of training about key aspects of data, AI and ethics, including learning a technique for exploring ethical aspects of new technologies (consequence scanning). As part of a pilot study, four People’s Panel sessions were held where tech businesses and researchers pitched their ideas and discussed questions and concerns of the panel members. Through participating in the panel, panel members reported an increase in confidence in being able to question businesses and businesses heard a diverse stakeholder voice on the ethical impacts of their products / services, leading to change.
In this paper, we evaluated several large language models (including ChatGPT, GPT3 and LLAMA) by running standardized personality tests on their results. Generally, we found that each large language model has an internal consistent personality. We further found that LLama tends to score more highly on Neuroticism than other models, whereas ChatGPT/GPT3 tends to score more highly on Conscientiousness and Agreeableness.
Establishing generalizable models for locomotion mode recognition (LMR) of prosthetic gait can be challenging due to limited access of sufficient labelled datasets. Hence, subject-specific models continue to be primarily used. However, there are no studies that investigated the effect of reducing the amount of training data that is presented to the machine learning model during training. Additionally, previously validated LMR models for prosthetic gait primarily used LDA classifiers. However, literature suggests that RF models may improve overall accuracy based on able-body validation. Therefore, to address those gaps, this study compared the performance of LDA and RF models for prosthetic gait and classifiers to LDA. Varied test size ratios data were evaluated to assess the trade-off between performance and amounts of training data.
In today’s industry, old machines, that were not manufactured according to Industry 4.0 standards, may not be equipped with sophisticated sensors for monitoring critical values and ensuring the machine's proper health and operation. As a result, third-party sensors, such as thermometers and vibration sensors, are often integrated into these machines. Unfortunately, despite being able to obtain effective measurements, such sensors lack relativization of these values to the contextual values of each machine. This paper proposes a risk assessment model that digitally mimics a real-life centrifugal governor's operation. The system combines machine learning and data analysis and uses a context-aware algorithm that can work with single or multiple sensors to output aggregated information on a machine’s health.
The energy resource management problem is regarded with great importance in the energy domain due to the current transformation of the electrical grid as a result of the growth of smart grid technologies. In this situation, conventional formulations created for an entirely different scenario occasionally fail to address the issue effectively. Modern metaheuristic optimizers are a powerful tool for handling such issues when old techniques fail. This work proposes a user-friendly web Meta-ERM platform for metaheuristic optimization when solving a given case study’s energy resource management problem and allows the visualization of performance analysis.
Interest in inverse design for the efficient and accurate design of optical devices has increased in recent years. In the case of complex optical problems which span several orders of magnitude, inverse design is an especially difficult problem. In this paper we propose a multi-scale inverse design process which leverages machine learning tools to encode the numerical simulation of optical wave propagation and material wave modulation directly as layers of a neural network. This requires consideration of both the near field electromagnetic response with respect to metasurface (material) devices, as well as far field effects as the wave propagates through space. The end result is the efficient modeling and optimization spanning several orders of magnitude.
Deep learning based methods have demonstrated great success in network intrusion detection. However, the use of Deep Neural Networks (DNNs) makes it difficult to support real-time, packet-level detections in communication networks that handle high-speed traffic with low latency and energy. To this end, this paper proposes a novel approach to efficiently realize a DNN-based classifier by converting it into a pruned, explainable decision tree and evaluating its hardware implementation using an emerging architecture based on memristor devices, in order to support network intrusion detections on the fly. Preliminary experiments on real-world datasets show that the proposed method achieves nearly four orders of magnitude speed up while retaining the desired accuracy.
Accurately predicting delays is crucial for improving passenger service quality and railway traffic management. The import of artificial neural networks into train delay prediction helps to improve prediction accuracy. In this paper, we propose a novel stacking ensemble regression model (ST-NN) that uses MultiLayer Perceptron’s (MLP) neural networks as single learners and MLP as the meta-learner, which enhances the accuracy of Passenger Train arrival delay prediction time in minutes. We evaluated the model performance using Amtrak data to compare with other combinations of stacking ensembles and single learners including, Decision Trees (DT), Random Forest (RF), Gradient Boosting Machines (GBM), XGBoost (XGB), LightGBM (LGBM), regression algorithms, Artificial Neural Network (ANN), and MLP to determine enhanced model accuracy of our model. The experiments demonstrate that our ST-NN regression model significantly improves model evaluation indicators by producing a 63.4-82.37% decrease in Mean Absolute Error. Furthermore, the accuracy outperforms the best benchmark models regarding train delay prediction.
Sepsis, leading to an estimated 11 million deaths per year, is often left undiagnosed due to its heterogeneity and lack of a single diagnostic test [3]. Every hour of delay in sepsis treatment increases the mortality rate by 4-8%, making early diagnosis and medical intervention critical to saving lives [1].Although several machine learning models have been developed using clinical data, their performance has been unsatisfactory, with low sensitivity scores leading to high mortality. To overcome this, a unique segmentation method is applied to a large time series clinical dataset of 40,336 patients, including 2,932 sepsis and 37,404 nonsepsis cases, comprising 41 variables of laboratory values, vital signs, and demographic data. Multiple experiments are conducted using different machine learning algorithms such as K-Nearest Neighbors, Random Forest, Multi-Layer Perceptron, and Gradient Boosting. The findings reveal that the XGB algorithm with a six-hour early prediction outperforms other models with a recall value of 0.98 and AUROC of 0.98 in predicting sepsis onset. Additionally, the use of data from 12 hours before onset results in a performance recall of 0.86 and AUROC of 0.95. These results demonstrate the potential of utilizing machine learning algorithms for early sepsis detection and highlight the importance of time series data segmentation and feature engineering for improved model performance.
