Multi-Task Learning for Dense Prediction Tasks: A Survey Vandenhende, Simon; Georgoulis, Stamatios; Van Gansbeke, Wouter ...
IEEE transactions on pattern analysis and machine intelligence,
07/2022, Letnik:
44, Številka:
7
Journal Article
Recenzirano
Odprti dostop
With the advent of deep learning, many dense prediction tasks, i.e., tasks that produce pixel-level predictions, have seen significant performance improvements. The typical approach is to learn these ...tasks in isolation, that is, a separate neural network is trained for each individual task. Yet, recent multi-task learning (MTL) techniques have shown promising results w.r.t. performance, computations and/or memory footprint, by jointly tackling multiple tasks through a learned shared representation. In this survey, we provide a well-rounded view on state-of-the-art deep learning approaches for MTL in computer vision, explicitly emphasizing on dense prediction tasks. Our contributions concern the following. First, we consider MTL from a network architecture point-of-view. We include an extensive overview and discuss the advantages/disadvantages of recent popular MTL models. Second, we examine various optimization methods to tackle the joint learning of multiple tasks. We summarize the qualitative elements of these works and explore their commonalities and differences. Finally, we provide an extensive experimental evaluation across a variety of dense prediction benchmarks to examine the pros and cons of the different methods, including both architectural and optimization based strategies.
Hypothetical Purchase Tasks (HPTs) simulate demand for a substance as a function of escalating price. HPTs are increasingly used to examine relationships between substance-related correlates and ...outcomes and demand typically characterized using a common battery of indices (Intensity, Omax, Pmax, Breakpoint, Elasticity). This review examines the relative sensitivity of the HPT indices. Reports were identified using the search term “purchase task” in PubMed and Web of Science. For inclusion, reports had to be original studies in English, examine relationships between HPT indices and substance-related correlates or outcomes, and appear in a peer-reviewed journal through December 2017. Indices were compared using effect sizes (Cohen's d) and the proportion of studies in which statistically significant relationships were observed. The search identified 1274 reports with 114 (9%) receiving full-text review and 82 (6%) meeting inclusion criteria. 41 reports examined alcohol, 34 examined cigarettes/nicotine products, and 10 examined other substances. Overall, statistically significant relationships between HPT indices and substance-related correlates and outcomes were most often reported for Intensity (88.61%, 70/79), followed by Omax (81.16%, 56/69), Elasticity (72.15%, 57/59), Breakpoint (62.12%, 41/66), and Pmax (48.08%; 25/52). The largest effect sizes were observed for Intensity (0.75 ± 0.04, CI 0.67–0.84) and Omax (0.64 ± 0.04, CI 0.56–0.71), followed by Elasticity (0.44 ± 0.04, CI 0.37–0.51), Breakpoint (0.30 ± 0.03, CI 0.25–0.36), and Pmax (0.25 ± 0.04, CI 0.18–0.33). Patterns were largely consistent across substances. In conclusion, HPTs can be highly effective in revealing relationships between demand and substance-related correlates and outcomes, with Intensity and Omax exhibiting the greatest sensitivity.
•A systematic literature review of studies using hypothetical purchase tasks (HPTs)•The review exclusively focused on studies using HPTs to assess substance use.•Overarching aim was to assess the relative sensitivity of the five HPT indices.•Intensity and Omax were more sensitive than Elasticity, Breakpoint, or Pmax.•That pattern had generality across multiple substances and study types.
With the increasing number of services and industries including nuclear, chemical, aerospace, and automotive sectors in cyber‐physical systems (CPSs), systems are being severely overloaded. CPSs ...comprises mixed‐critical tasks which are of either safety‐critical (high) or non‐safety critical (low). In traditional task scheduling, most of the existing scheduling algorithms provide poor performance for high‐criticality tasks when the system experiences overload and do not show explicit separation among different criticality tasks to take advantage of using cloud resources. Here, we propose a framework to schedule the mixed‐criticality tasks by analyzing their deadlines and execution times which leverage the performance of parallel processing through OpenMP. The proposed framework introduces a machine learning‐based prediction for a task offloading in the cloud. Moreover, it illustrates to execute a selected number of low‐criticality tasks in the cloud while the high‐criticality tasks are run on the local processors during the system overload. As a result, the high‐criticality tasks meet all their deadlines and the system achieves a significant improvement in the overall execution time and better throughput. In addition, the experimental results employing OpenMP show the effectiveness of using the partitioned scheduling over the global scheduling method upon multiprocessor systems to achieve the tasks isolation.
Crowdsourcing information systems are socio-technical systems that provide informational products or services by harnessing the diverse potential of large groups of people via the Web. Interested ...individuals can contribute to such systems by selecting among a wide range of open tasks. Arguing that current approaches are suboptimal in terms of matching tasks and contributors' individual interests and capabilities, this article advocates the introduction of personalized task recommendation mechanisms. We contribute to a conceptual foundation for the design of such mechanisms by conducting a systematic review of the corresponding academic literature. Based on the insights derived from this analysis, we identify a number of issues for future research. In particular, our findings highlight the need for more significant empirical results through large-scale online experiments, an improved dialog with mainstream recommender systems research, and the integration of various sources of knowledge that exceed the boundaries of individual systems.
In software crowdsourcing, task prize is a primary incentive for engaging crowd developers. One of the main challenges in crowdsourcing task pricing is to determine appropriate prizes in order to ...attract qualified workers. Few studies proposed methods to address this challenge. However, they are either too theoretical or too restricted to be applied for early crowdsourcing planning. In this study, we propose a novel approach, i.e., PTMA, to support early task pricing in software crowdsourcing from textual task requirements. PTMA consists of three phases, namely data pre-processing, topic extraction, and topic-based task pricing analysis, integrating 6 machine learning algorithms and 3 analogy-based models for topic-based pricing analysis. PTMA is evaluated using data from 2016 software crowdsourcing tasks extracted from TopCoder, the largest software crowdsourcing platform. The results show that: 1) textual requirement information can aid early task pricing in software crowdsourcing; 2) the best predictor in PTMA, based on logistic regression, achieves an accuracy of 88.3% in Pred (30); and 3) PTMA outperforms the existing baseline models by 9% in Pred (30). PTMA greatly simplifies the pricing process by only leveraging textual task description as inputs, and can achieve better prediction accuracy in making task pricing decisions.
Objective
The aim of this pilot study was to investigate differences on dual‐ and triple‐task performance in institutionalized prefrail and frail older adults. Performance on these tasks is relevant ...since many activities of daily living involve simultaneous motor and cognitive tasks.
Methods
We used a phenotypic description of frailty based on the presence or absence of five criteria related to physical fitness and metabolism (unintentional weight loss, self‐reported exhaustion, muscle weakness, low gait speed, and low physical activity). Thirty‐three institutionalized older adults (≥ 65 years, 78.8% females) were divided according to their frailty status. Participants completed cognitive tasks (a phonemic verbal fluency task and a visuospatial tracking task) while cycling on a stationary cycle (upper‐ and lower‐extremity function was assessed). Cycling (number of arm and foot cycles) and cognitive (number of correct answers) performances were measured during single‐, dual‐, and triple‐task conditions. Performances and costs of dual ‐and triple‐ tasking on cycling and cognitive performances were compared between prefrail and frail groups.
Results
Prefrail and frail older adults did not differ in their performance in dual‐tasks; however, frail older adults showed a poorer performance in the triple‐task.
Conclusions
Although future studies need to confirm our observations in larger samples, this pilot study suggests that developing new tools based on triple tasking could be useful for the comprehensive assessment of frailty.
Engaging the retrieval state (Tulving, 1983) impacts processing and behavior (Long and Kuhl, 2019, 2021; Smith et al., 2022), but the extent to which top-down factors—explicit instructions and ...goals—versus bottom-up factors—stimulus properties such as repetition and similarity—jointly or independently induce the retrieval state is unclear. Identifying the impact of bottom-up and top-down factors on retrieval state engagement is critical for understanding how control of task-relevant versus task-irrelevant brain states influence cognition. We conducted between-subjects recognition memory tasks on male and female human participants in which we varied test phase goals. We recorded scalp electroencephalography and used an independently validated mnemonic state classifier (Long, 2023) to measure retrieval state engagement as a function of top-down task goals (recognize old vs detect new items) and bottom-up stimulus repetition (hits vs correct rejections (CRs)). We find that whereas the retrieval state is engaged for hits regardless of top-down goals, the retrieval state is only engaged during CRs when the top-down goal is to recognize old items. Furthermore, retrieval state engagement is greater for low compared to high confidence hits when the task goal is to recognize old items. Together, these results suggest that top-down demands to recognize old items induce the retrieval state independent from bottom-up factors, potentially reflecting the recruitment of internal attention to enable access of a stored representation.
OpenAI has released the Chat Generative Pre-trained Transformer (ChatGPT) and revolutionized the approach in artificial intelligence to human-model interaction. The first contact with the chatbot ...reveals its ability to provide detailed and precise answers in various areas. Several publications on ChatGPT evaluation test its effectiveness on well-known natural language processing (NLP) tasks. However, the existing studies are mostly non-automated and tested on a very limited scale. In this work, we examined ChatGPT’s capabilities on 25 diverse analytical NLP tasks, most of them subjective even to humans, such as sentiment analysis, emotion recognition, offensiveness, and stance detection. In contrast, the other tasks require more objective reasoning like word sense disambiguation, linguistic acceptability, and question answering. We also evaluated GPT-4 model on five selected subsets of NLP tasks. We automated ChatGPT and GPT-4 prompting process and analyzed more than 49k responses. Our comparison of its results with available State-of-the-Art (SOTA) solutions showed that the average loss in quality of the ChatGPT model was about 25% for zero-shot and few-shot evaluation. For GPT-4 model, a loss for semantic tasks is significantly lower than for ChatGPT. We showed that the more difficult the task (lower SOTA performance), the higher the ChatGPT loss. It especially refers to pragmatic NLP problems like emotion recognition. We also tested the ability to personalize ChatGPT responses for selected subjective tasks via Random Contextual Few-Shot Personalization, and we obtained significantly better user-based predictions. Additional qualitative analysis revealed a ChatGPT bias, most likely due to the rules imposed on human trainers by OpenAI. Our results provide the basis for a fundamental discussion of whether the high quality of recent predictive NLP models can indicate a tool’s usefulness to society and how the learning and validation procedures for such systems should be established.
•The results of ChatGPT and GPT-4 evaluation on 25 tasks using 48k+ prompts.•Context-awareness and personalization are valuable capabilities of ChatGPT.•ChatGPT and GPT-4 are always worse compared to SOTA methods from 4% to over 70%.•ChatGPT loss tends to be higher for more difficult reasoning problems.•ChatGPT can boost AI development and change our daily lives.
In the past few years, with the rapid development of heterogeneous computing systems (HCS), the issue of energy consumption has attracted a great deal of attention. How to reduce energy consumption ...is currently a critical issue in designing HCS. In response to this challenge, many energy-aware scheduling algorithms have been developed primarily using the dynamic voltage-frequency scaling (DVFS) capability which has been incorporated into recent commodity processors. However, these techniques are unsatisfactory in minimizing both schedule length and energy consumption. Furthermore, most algorithms schedule tasks according to their average-case execution times and do not consider task execution times with probability distributions in the real-world. In realizing this, we study the problem of scheduling a bag-of-tasks (BoT) application, made of a collection of independent stochastic tasks with normal distributions of task execution times, on a heterogeneous platform with deadline and energy consumption budget constraints. We build execution time and energy consumption models for stochastic tasks on a single processor. We derive the expected value and variance of schedule length on HCS by Clark's equations. We formulate our stochastic task scheduling problem as a linear programming problem, in which we maximize the weighted probability of combined schedule length and energy consumption metric under deadline and energy consumption budget constraints. We propose a heuristic energy-aware stochastic task scheduling algorithm called ESTS to solve this problem. Our algorithm can achieve high scheduling performance for BoT applications with low time complexity O(n(M + logn)), where n is the number of tasks and M is the total number of processor frequencies. Our extensive simulations for performance evaluation based on randomly generated stochastic applications and real-world applications clearly demonstrate that our proposed heuristic algorithm can improve the weighted probability that both the deadline and the energy consumption budget constraints can be met, and has the capability of balancing between schedule length and energy consumption.
Traffic flow prediction is a fundamental problem in transportation modeling and management. Many existing approaches fail to provide favorable results due to being: 1) shallow in architecture; 2) ...hand engineered in features; and 3) separate in learning. In this paper we propose a deep architecture that consists of two parts, i.e., a deep belief network (DBN) at the bottom and a multitask regression layer at the top. A DBN is employed here for unsupervised feature learning. It can learn effective features for traffic flow prediction in an unsupervised fashion, which has been examined and found to be effective for many areas such as image and audio classification. To the best of our knowledge, this is the first paper that applies the deep learning approach to transportation research. To incorporate multitask learning (MTL) in our deep architecture, a multitask regression layer is used above the DBN for supervised prediction. We further investigate homogeneous MTL and heterogeneous MTL for traffic flow prediction. To take full advantage of weight sharing in our deep architecture, we propose a grouping method based on the weights in the top layer to make MTL more effective. Experiments on transportation data sets show good performance of our deep architecture. Abundant experiments show that our approach achieved close to 5% improvements over the state of the art. It is also presented that MTL can improve the generalization performance of shared tasks. These positive results demonstrate that deep learning and MTL are promising in transportation research.