Introduction

Deep learning (DL) is a subset of machine learning (ML) that is created using complex algorithms that are inspired by the organization of the human brain with many discrete nodes or neurons and can identify important patterns or features in a dataset.1 DL and ML refer to two different technologies, and DL is considered an advanced structure of ML. Convolutional neural networks, long and short-term memory networks, recurrent neural networks, transformer models, and attention mechanisms are all common u DL technologies.2 ML techniques are a collection of mathematical and statistical concepts such as support vector machine, random forest, and K-nearest neighbors.3 Whereas DL algorithms are specialized techniques that are a subset of ML.1 The most important difference between the two approaches is that ML requires a feature engineering process that eliminates unnecessary variables and pre-selects only those that will be used for learning.4 This process is disadvantaged by the requirement that experienced professionals pre-select critical variables. Conversely, DL algorithms overcomes this shortfall by a process in which have built-in mechanisms for assessing and addressing the root of any inaccuracies and do not require guidance.4 To interpret an image, for example, to deconstruct the image into specific features, such as brightness, curvature, sharpness, etc., the extraction process of the support vector machine, a ML technique, requires digital input into the computer algorithms to extract these image features.5 Whereas this feature extraction process is completely different in DL algorithms. By varying the weights of the given features, DL applies a series of convolutional filters to the image, and the DL algorithm can then be trained to recognize a specific type of image and ultimately achieve the extraction of features from the image.6

Patients in the critical care medicine (CCM) field usually have aggressive and complex conditions, complicated medical record data, and clear trends in personalized treatment, resulting in a huge need for automated and reliable health information processing and analysis.7 DL algorithms and models enable machines to mimic human activities such as seeing, hearing, and thinking, helping to solve many complex pattern recognition challenges, and such features seem to help bring breakthroughs to the CCM. In addition to its well-known use in image processing and analysis, DL is also widely used in the medical field for health-record analysis, clinical diagnosis, health monitoring, personalized medicine, and drug development.7 Managing most diseases in the field of CCM, such as sepsis, acute respiratory distress syndrome, and severe stroke, largely requires the application of DL technologies.810 Currently, various algorithms and models based on DL have been widely used in the management of common diseases in CCM (such as sepsis and acute respiratory distress syndrome), including early detection of diseases, severity score estimation, facilitating ICU liberation through early successful extubation, early mobility, and survival prediction.1115 Accumulating evidence suggests that the application of DL technology promotes intelligence in CCM, not only effectively improving the quality of medical care, but also helping to increase the efficiency of clinicians.16

Publications on the application of DL in CCM have continued to grow in recent years. The continuous increase in publications is positive for the updating of knowledge, but it also poses a challenge for researchers, as the process of acquiring knowledge makes it difficult to avoid the heavy work of combing publications.17 As a quantitative research method used to analyze the scholarly characteristics of the literature in certain scientific fields, bibliometrics helps researchers to grasp research hotspots and trends in their fields of interest and to predict their prospects.18 Therefore, the present bibliometric and visualized study was performed to provide a comprehensive overview of current hotspots, future trends in the use of DL in CCM, to showcase the contributions of leading countries, authorities, and prominent scholars, and to provide clues to potential future collaborations and research directions.

Data were obtained from the Web of Science on 15 March, 2023 using the followed strategy: TS=(critical care OR critically ill OR intensive care OR ICU OR high dependency) AND TS=(deep learning OR convolutional neural network). It should be noted that TS=Topic. The inclusion criteria were as follows: (a) literature published between 2012 and 2022; (b) articles as the type of literature; and (c) literature published in English. Duplicate publications were excluded. A manual check of the included literature was independently performed by two authors (clinicians).

Full records and cited references of the obtained publications were downloaded in BibTex or txt formats for further analysis. Information on the publications, including title, abstract, key words, country, author, institution, source, count of citations, cited references, and the 2021 IF of the top 10 core journals as well as the H-index of the top 10 most productive authors were recorded. Data extraction were conducted by two independent authors.

Bibliometric analysis was performed by Bibliometrix package in R software (4.2.2), Microsoft Excel 2019, VOSviewer (1.6.18), and CiteSpace (5.8.R3). In the present study, publication trends in the literature were analyzed using Microsoft Excel 2019. A polynomial-fitting curve in Microsoft Excel 2019 was applied to predict the number of future publications. The National collaborative networks, author collaborative networks, institutional collaborative networks, and journal publication trends were constructed using Bibliometrix package. Furthermore, co-occurrence of keywords and co-citation relationship of references were analyzed using VOSviewer. Finally, using CiteSpace to analyze keyword bursts.

A total of 1708 articles on DL in CCM were published in the past 11 years. Overall, there was an overall upward trend in the number of publications (Figure 1A), with 3 in 2012 to 651 in 2022. Notably, research activity peaked in 2017, where 95.67% (1634/1708) articles were published during the past six years.

Figure 1 (A) The number of publications and annual citations over time. (B) Curve fitting of the e total annual growth trend of publications (R2 = 0.9773).

Furthermore, the polynomial-fitting curve suggested that research in this area will continue to grow, with an R2 value of 0.9773 (Figure 1B).

Publications on this topic were contributed by 62 different countries/regions. Table 1 shows the top ten most productive countries/regions. China ranked first with 804 publications, followed by the USA with 420 publications. USA had the highest number of total citations and average citations of all publications. Furthermore, China maintained close ties with USA, Korea, and France, whereas USA had strong cooperative bonds with China, England, and Australia (Figure 2).

Table 1 The Top 10 Publishing Countries/Regions

Figure 2 The international collaboration between countries/regions.

Notes: The different colors of arcs represent different countries/regions, and the larger the arc area, the wider the international cooperation of the country/region. Line thickness between countries/regions reflects the intensity of the closeness.

For the analysis of institutions, 2379 institutions made contributions to this field. The top ten most productive institutions were Chinese Academy of Sciences, Harvard University, University of Chinese, Academy of Sciences, University of California System, Wuhan University, Harvard Medical School, Tsinghua University, Centre National De La Recherche Scientifique CNRS, Shanghai Jiao Tong University, and Harbin Institute of Technology (Table 2). Notably, the top 10 institutions were from the China (n = 6), USA (n = 3), and France (n = 1). Furthermore, University of Chinese Academy of Sciences, Harvard University, and Tsinghua University have more connections to other affiliations (Figure 3).

Table 2 The Top 10 Publishing Affiliations

Figure 3 Collaboration between affiliations.

Notes: Each circle represents an affiliation, and the larger the circle, the wider the cooperative relationship. Affiliations with frequent cooperative relationships are clustered into plates of the same color. Line thickness between affiliations reflects the intensity of the closeness.

Over the last ten years, a total of 6211 authors have made significant contributions to the field. Based on publication counts, Wang Y was the most productive author (n = 37), followed by Liu Y, Li Y, Wang J, Wang L, Zhang J, Zhang Y, Li L, Liu J, and Yang Y (Table 3). Furthermore, Zhang Y with the highest total citations and average citations, whereas Li Y has the highest H-index. Interestingly, the top ten most productive authors are all from China. These findings agree with the total productions for the nations mentioned above, showing that China is leading the way in this area. In addition, Wang Y, Wang X, Wang J and Zhang Y have more connections to other authors (Figure 4).

Table 3 The Top 10 Publishing Authors

Figure 4 Collaboration between authors.

Notes: Each circle represents an author, and the larger the circle, the wider the cooperative relationship. Authors with frequent cooperative relationships are clustered into plates of the same colour. Line thickness between authors reflects the intensity of the closeness.

A total of 260 journals have made contributions to this field. As shown in Table 4, IEEE Access, Scientific Reports, Remote Sensing were the top three. When it came to journal impact, IEEE Transactions on Geoscience and Remote Sensing ranked first, with an IF of 8.125, followed by Journal of Biomedical Informatics (IF = 8.000), and Computers in Biology and Medicine (IF = 6.698). These journals were therefore valuable resources for research in this field. Additionally, over the last five years, the top five most active journals have displayed a sharp rise in the number of annual publications (Figure 5).

Table 4 The Top 10 Most Active Journals

Figure 5 Publications of the top 5 most active journals over time.

A total of 59,659 co-cited references were identified. After setting the minimum number of citations to 30, 52 of them were selected to form the cited reference network, which contained four clusters (Figure 6). Cluster 1 (in red) primarily centered on the model development and validation, cluster 2 (in green) primarily centered on the application of the models, cluster 3 (in blue) and cluster 4 (in yellow) primarily centered on the application of models in the medical field. These results suggest that DL-based model development and applications are the foundation of current research in this area. All literature included in Figure 6 is provided in Supplementary Material 1.

Figure 6 Network visualization map of co-citation references.

Notes: Cluster 1 (in red), cluster 2 (in green), cluster 3 (in blue), cluster 4 (in yellow). The lines between the circles represent the co-citation relationship. The thickness and number of connections between the nodes indicate the strength of links between references.

A total of 4864 keywords were identified. Deep learning, machine learning, and feature extraction were the keywords with the highest frequency (Figure 7A). After setting the minimum number of occurrences to ten, 84 of them were selected to form the keyword network, which contained six clusters (Figure 7A). Cluster 1 (in red) primarily centered on the model development and validation, cluster 2 (in green) and cluster 4 (in yellow) primarily centered on the extraction of clinical characteristics of critically ill patients, cluster 3 (in blue) primarily centered on the prognosis of critically ill patients, cluster 5 (in purple) primarily centered on monitoring changes in the condition of critically ill patients; and cluster 6 (in light blue) primarily centered on big data analysis in CCM. It was worth noting that adaptation models, computed tomography, and electronic medical records were recent emerging hot topics (Figure 7B). These topics offer potential research directions on DL in CCM for the future.

Figure 7 (A) Network map of keywords on DL in CCM. (B) Visualization map of top 15 keywords with the strongest citation bursts.

Notes: Cluster 1 (in red), cluster 2 (in green), cluster 3 (in blue), cluster 4, cluster 5 (in purple), cluster 6 (in light blue). The node size reflects the co-occurrence frequencies and the link indicates the co-occurrence relationship. The thickness of the link is proportional to the number of times two keywords co-occur. The blue bars indicate that the keywords have been published and the red bars indicate citation burstness.

The present study used a bibliometric approach to analyze publications of DL in CCM by exploring the expansion of research interest, publication output, top nations, international cooperation, top institutions, authoritative scholars, preferred journals, keywords, and citation analysis.

Publications on DL in CCM have grown steadily in recent years since the concept of DL was introduced in 2006.19 In 2018, there was a significant increase in interest in DL in CCM, which marked the turning point. Interest in DL in general medicine has been gradually increasing since 2012, but in the field of CCM, it was significantly delayed by six years.20 Indeed, the DL models are not commonly used in CCM daily practice. The reason is that few models have external validation, clinical interpretability and high predictability.21,22 Furthermore, most models are developed in a single institution and are do not perform well when applied to other institutions.23 There are also limited venues to incorporate the models. Ideally, they would be embedded in electronic health record systems, but this is challenging to implement due to the limitations of these systems and the corporate disincentives to do so.24,25 In addition, privacy issues are also one of the challenges faced by the adoption of artificial intelligence in the medical field.26 Based on the above evidence, we therefore believe that the safety and accountability of DL models applied to critically ill patients has not yet fully gained gain enough trust from the people. Since 2016, with the continuous development of DL technologies and the development of DL models that begin to pay attention to multi-center data sources and external validation, the accuracy and clinical adaptability of the models have been strengthened, which may help to establish patient confidence in the DL model.27,28 The emergence of new DL technologies, such as convolutional neural networks, long and short-term memory networks, recurrent neural networks, transformer models, and attention mechanisms, offers previously unheard-of possibilities for disease management, diagnosis, and prediction. In addition, MIMIC and eICU, two large public intensive care databases launched in 2016 and 2018, respectively, became available to researchers.29,30 Especially, the release of the MIMIC III database was a large contributing factor to the development of DL models in CCM. Types of DL models developed based on MIMIC III typically include diagnostic models, disease severity score models, real-time monitoring models, hospital length of stay prediction models, readmission prediction models, survival prediction models, and automated adverse drug reaction reporting models.12,15,3135 Common conditions in the field of CCM that these DL models are applied to manage include sepsis, acute respiratory distress syndrome, acute kidney injury, and cardiovascular disease.3135 The common variables used in these models can be divided into 4 categories: history information, admission information, vital signs, laboratory results, and arterial blood gas analysis.15 However, MIMIC-III is only an extensive single-center database spanning from 2001 to 2012 of electronic medical records of patients admitted to the ICU at Beth Israel Deaconess Medical Center, an academic teaching hospital of Harvard Medical School in Boston, USA.36 Therefore, the establishment of a continuously updated CCM database of multicenter admitted ICU patient data, or even electronic medical records of ICU patients admitted globally, would be more conducive to the promotion of the field of CCM as well as the application of DL models.

It was discovered that high-income nations predominate in DL research on CCM after analyzing the distribution of publications across nations. Notably, the top 30 countries in the world in terms of GDP include these ten most productive countries, suggesting that the number of publications is closely linked to the economic power of each country. This finding is consistent with the bibliometric findings of many other medical disciplines.3739 Furthermore, over 70% of the publications came from the USA and China, indicating that these two nations are the main contributors to the DL in CCM research. Additionally, the highest citation rates are also found in these two nations, though China has a lower average citation rate per article than the USA. Analysis of collaborative networks showed that the USA and China are the countries with the most collaborative network relationships. A stable and adaptable policy is a prerequisite to ensure that international collaborations are successfully achieved. Adequate research funding, a wide range of research collaborators, as well as a significant proportion of visiting scholars all contribute to improved international partnerships. Furthermore, the top 100 universities in the world include seven of the top ten most productive institutions, indicating that the use of DL in CCM has gained the attention of leading universities. Researchers may be encouraged to consider conducting conjoint research or applying for educational programs or visiting scholars with these top institutions in the USA or China.

Notably, current articles on this topic are not in the top-tier clinical journals in this space, such as Critical Care Medicine, JAMA, or the New England Journal. It is likely due to the fact that there is still a gap between the DL models and clinical applicability. It needs to be acknowledged that DL has the advantage of responding to the challenges faced by CCM. DL algorithms and models enable machines to mimic human activities such as vision, hearing, and thinking and can automate and reliably process and analyze health information.7 Therefore, various algorithms and models based on DL have been widely used in the management of common diseases in CCM, including early detection of diseases, severity score estimation, facilitating ICU liberation through early successful extubation, early mobility, and survival prediction.1115 Accumulating evidence suggests that the application of DL technology promotes intelligence in CCM, not only effectively improving the quality of medical care, but also helping to increase the efficiency of clinicians.16 DL implementation will support clinicians in the decision-making processes. Benefits comprehend earlier diagnoses, detection of subclinical deteriorations and generation of new medical knowledge.7 To improve the clinical utility of DL models in CCM, the following challenges need to be addressed. First, DL models that lack external validation automatically move away from clinical applicability. It should be advocated that researchers should develop external validation from a multidimensional perspective to continuously improve the scientific validity of the model and enhance the predictive performance, which in turn will promote the clinical interpretability and applicability of DL models in the CMM field.21 Models developed in a single institution are not always applicable to other institutions. Therefore, constructing models based on multicenter shared data should be advocated to increase the breadth of their applicability.23 If achievable, the creation of a global shared database of electronic medical records in the field of CCM is expected to bring a major breakthrough in this area.36 Furthermore, privacy protection through policies remains the cornerstone of health data, with the addition of special safeguards for personal health data addressed by the new innovative principles of the General Data Protection Regulation.40

Based on the authors keywords in the identified categories, CCM-related DL research mainly focused on the model development and validation, the extraction of clinical characteristics of critically ill patients, the prognosis of critically ill patients, monitoring changes in the condition of critically ill patients, and big data analysis in CCM. Furthermore, the primary disease domains addressed in CCM-related DL research were COVID-19, ARDS, sepsis, cardiac arrest, and acute kidney injury. The common targets of DL algorithms are these common diseases in CCM. It was worth noting that adaptation models, computed tomography, and electronic medical records were recent emerging hot topics. These topics offer potential research directions on DL in CCM for the future. DL has demonstrated potential applications in various areas of CCM. However, the development and implementation of DL in CCM remains challenging. Firstly, the absence of external validation and prospective assessment to confirm the repeatability of DL protocols both limit the utility of DL in clinical practice.41 Secondly, implementing artificial intelligence models in clinical practice may entail high initial costs, which is a significant barrier to implementing artificial intelligence (AI) in low- and middle-income countries.42 Furthermore, the World Health Organization released guidelines on the ethics and management of AI for health in 2021, emphasizing the key role of privacy, transparency, informed consent, and regulation of data protection frameworks.43 Thus, the legal protection of patient privacy also limits the current widespread use of AI in medicine.42

The application of AI is beneficial for the advancement of medicine.44,45 This study provides a systematic review of hot spots and trends in CCM-related DL research, highlights leading countries and institutions, reveals potential partnership networks, and provides insights into the direction of future research. However, limitations should be acknowledged. The COVID-19 pandemic has impacted various industries around the world, including the DL field and the CCM field. From the outset of the COVID-19 pandemic, it was clear that the greatest challenge was the unavailability of fully equipped and staffed ICU beds.46 With the expansion of the Internet, the amount of content on COVID-19 has exploded in the last three years. In addition to fact-based content, a large amount of COVID-19 content is being manipulated, and it leads to people spending more time online and getting more invested in this false content.47 Potentially preventing its spread by using DL to identify uninformative information early has also raised concerns.47 Therefore, the COVID-19 pandemic may introduce a publication bias to the publication trend of CCM-related DL research, which may lead to unstable research hotspots and trends identified in this study. Furthermore, given that our search strategy was constructed based on Topic, the search strategy used in the present study may have resulted in missing some relevant literature.

Hot spots in research on the application of DL in CCM have focused on classifying disease phenotypes, predicting early signs of clinical deterioration, and forecasting disease progression, prognosis, and death. Extensive collaborative research to improve the maturity and robustness of the model remains necessary to make DL-based model applications sufficiently compelling for conventional CCM practice.

This work was supported by Science and Technology Development Fund of Hospital of Chengdu University of Traditional Chinese Medicine (No.21ZL08).

The authors report no conflicts of interest in this work.

1. Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. Commun ACM. 2017;60:8490. doi:10.1145/3065386

2. Girshick R, Donahue J, Darrell T, Malik J. Rich feature hierarchies for accurate object detection and semantic segmentation. 27th IEEE Conference on Computer Vision and Pattern Recognition (CVPR); Columbus, OH; 2014.

3. Goecks J, Jalili V, Heiser LM, Gray JW. How machine learning will transform biomedicine. Cell. 2020;181(1):92101. doi:10.1016/j.cell.2020.03.022

4. Shinde P, Shah S. A review of machine learning and deep learning applications; 2019.

5. Irisson JO, Ayata SD, Lindsay DJ, Karp-Boss L, Stemmann L. Machine learning for the study of plankton and marine snow from images. Ann Rev Mar Sci. 2022;14:277301. doi:10.1146/annurev-marine-041921-013023

6. Wang S, Yang DM, Rong R, Zhan X, Xiao G. Pathology image analysis using segmentation deep learning algorithms. Am J Pathol. 2019;189(9):16861698. doi:10.1016/j.ajpath.2019.05.007

7. Egger J, Gsaxner C, Pepe A, et al. Medical deep learning-A systematic meta-review. Comput Methods Programs Biomed. 2022;221:106874. doi:10.1016/j.cmpb.2022.106874

8. Zhang Z, Pan Q, Ge H, Xing L, Hong Y, Chen P. Deep learning-based clustering robustly identified two classes of sepsis with both prognostic and predictive values. EBioMedicine. 2020;62:103081. doi:10.1016/j.ebiom.2020.103081

9. Reamaroon N, Sjoding MW, Gryak J, Athey BD, Najarian K, Derksen H. Automated detection of acute respiratory distress syndrome from chest X-rays using directionality measure and deep learning features. Comput Biol Med. 2021;134:104463. doi:10.1016/j.compbiomed.2021.104463

10. Sharma N, Simmons LH, Jones PS, et al. Motor imagery after subcortical stroke: a functional magnetic resonance imaging study. Stroke. 2009;40(4):13151324. doi:10.1161/STROKEAHA.108.525766

11. Lauritsen SM, Kalr ME, Kongsgaard EL, et al. Early detection of sepsis utilizing deep learning on electronic health record event sequences. Artif Intell Med. 2020;104:101820. doi:10.1016/j.artmed.2020.101820

12. Aurolu T, Oul H. A deep learning approach for sepsis monitoring via severity score estimation. Comput Methods Programs Biomed. 2021;198:105816. doi:10.1016/j.cmpb.2020.105816

13. Jia Y, Kaul C, Lawton T, Murray-Smith R, Habli I. Prediction of weaning from mechanical ventilation using convolutional neural networks. Artif Intell Med. 2021;117:102087. doi:10.1016/j.artmed.2021.102087

14. Yeung S, Rinaldo F, Jopling J, et al. A computer vision system for deep learning-based detection of patient mobilization activities in the ICU. NPJ Digit Med. 2019;2:11. doi:10.1038/s41746-019-0087-z

15. Tang H, Jin Z, Deng J, et al. Development and validation of a deep learning model to predict the survival of patients in ICU. J Am Med Inform Assoc. 2022;29(9):15671576. doi:10.1093/jamia/ocac098

16. Datta R, Singh S. Artificial intelligence in critical care: its about time! Med J Armed Forces India. 2021;77(3):266275. doi:10.1016/j.mjafi.2020.10.005

17. Dan F, Kudu E. The evolution of cardiopulmonary resuscitation: global productivity and publication trends. Am J Emerg Med. 2022;54:151164. doi:10.1016/j.ajem.2022.01.071

18. Niu B, Hong S, Yuan J, Peng S, Wang Z, Zhang X. Global trends in sediment-related research in earth science during 19922011: a bibliometric analysis. Scientometrics. 2013;98(1):511529. doi:10.1007/s11192-013-1065-x

19. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436444. doi:10.1038/nature14539

20. Guo Y, Hao Z, Zhao S, Gong J, Yang F. Artificial intelligence in health care: bibliometric analysis. J Med Internet Res. 2020;22(7):e18228. doi:10.2196/18228

21. Ozrazgat-Baslanti T, Loftus TJ, Ren Y, Ruppert MM, Bihorac A. Advances in artificial intelligence and deep learning systems in ICU-related acute kidney injury. Curr Opin Crit Care. 2021;27(6):560572. doi:10.1097/MCC.0000000000000887

22. Ching T, Himmelstein DS, Beaulieu-Jones BK, et al. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface. 2018;15(141):20170387.

23. Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform. 2018;19(6):12361246. doi:10.1093/bib/bbx044

24. Wang L, Tong L, Davis D, Arnold T, Esposito T. The application of unsupervised deep learning in predictive models using electronic health records. BMC Med Res Methodol. 2020;20(1):37. doi:10.1186/s12874-020-00923-1

25. Golas SB, Shibahara T, Agboola S, et al. A machine learning model to predict the risk of 30-day readmissions in patients with heart failure: a retrospective analysis of electronic medical records data. BMC Med Inform Decis Mak. 2018;18(1):44. doi:10.1186/s12911-018-0620-z

26. Harvey HB, Gowda V. Regulatory issues and challenges to artificial intelligence adoption. Radiol Clin North Am. 2021;59(6):10751083. doi:10.1016/j.rcl.2021.07.007

27. Zhong L, Dong D, Fang X, et al. A deep learning-based radiomic nomogram for prognosis and treatment decision in advanced nasopharyngeal carcinoma: a multicentre study. EBioMedicine. 2021;70:103522. doi:10.1016/j.ebiom.2021.103522

28. Hiremath A, Shiradkar R, Fu P, et al. An integrated nomogram combining deep learning, Prostate Imaging-Reporting and Data System (PI-RADS) scoring, and clinical variables for identification of clinically significant prostate cancer on biparametric MRI: a retrospective multicentre study. Lancet Digit Health. 2021;3(7):e445e454. doi:10.1016/S2589-7500(21)00082-0

29. Pollard TJ, Johnson AE, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU collaborative research database, a freely available multi-center database for critical care research. Sci Data. 2018;5:180178. doi:10.1038/sdata.2018.178

30. Johnson AE, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3:160035. doi:10.1038/sdata.2016.35

31. Guo F, Zhu X, Wu Z, Zhu L, Wu J, Zhang F. Clinical applications of machine learning in the survival prediction and classification of sepsis: coagulation and heparin usage matter. J Transl Med. 2022;20(1):265. doi:10.1186/s12967-022-03469-6

32. Alfieri F, Ancona A, Tripepi G, et al. External validation of a deep-learning model to predict severe acute kidney injury based on urine output changes in critically ill patients. J Nephrol. 2022;35(8):20472056. doi:10.1007/s40620-022-01335-8

33. Wu J, Lin Y, Li P, Hu Y, Zhang L, Kong G. Predicting Prolonged Length of ICU stay through machine learning. Diagnostics. 2021;11(12):2242. doi:10.3390/diagnostics11122242

34. Pishgar M, Theis J, Del Rios M, Ardati A, Anahideh H, Darabi H. Prediction of unplanned 30-day readmission for ICU patients with heart failure. BMC Med Inform Decis Mak. 2022;22(1):117. doi:10.1186/s12911-022-01857-y

35. McMaster C, Chan J, Liew DFL, et al. Developing a deep learning natural language processing algorithm for automated reporting of adverse drug reactions. J Biomed Inform. 2023;137:104265. doi:10.1016/j.jbi.2022.104265

36. Rsli E, Bozkurt S, Hernandez-Boussard T. Peeking into a black box, the fairness and generalizability of a MIMIC-III benchmarking model. Sci Data. 2022;9(1):24. doi:10.1038/s41597-021-01110-7

37. Qiang W, Xiao C, Li Z, et al. Impactful publications of critical care medicine research in China: a bibliometric analysis. Front Med. 2022;9:974025. doi:10.3389/fmed.2022.974025

38. Liu YX, Zhu C, Wu ZX, Lu LJ, Yu YT. A bibliometric analysis of the application of artificial intelligence to advance individualized diagnosis and treatment of critical illness. Ann Transl Med. 2022;10(16):854. doi:10.21037/atm-22-913

39. Cui X, Chang Y, Yang C, Cong Z, Wang B, Leng Y. Development and trends in artificial intelligence in critical care medicine: a bibliometric analysis of related research over the period of 20102021. J Pers Med. 2022;13(1):50. doi:10.3390/jpm13010050

40. Castelluccia Claude, le Metayer Daniel, European Parliament. European Parliamentary Research Service. Scientific foresight unit. Understanding algorithmic decision-making: opportunities and challenges; 2019. Available from: https://www.europarl.europa.eu/RegData/etudes/STUD/2019/624261/EPRS_STU. Accessed January 27, 2023.

41. Yoon JH, Pinsky MR, Clermont G. Artificial intelligence in critical care medicine. Crit Care. 2022;26(1):75. doi:10.1186/s13054-022-03915-3

42. Caruso PF, Greco M, Ebm C, Angelotti G, Cecconi M. Implementing artificial intelligence: assessing the cost and benefits of algorithmic decision-making in critical care. Crit Care Clin. 2023;6072(1):1253.

43. Ethics and governance of artificial intelligence for health ethics and governance of artificial intelligence for health 2; 2021. Available from: http://apps.who.int/bookorders. Accessed January 27, 2023.

44. Wu Q, Liu S, Zhang R, et al. ACU&MOX-DATA: a platform for fusion analysis and visual display acupuncture multi-omics heterogeneous data. Acupunct Herbal Med. 2023;3(1):5962. doi:10.1097/HM9.0000000000000051

45. Jiang C, Qu H. In-line spectroscopy combined with multivariate analysis methods for endpoint determination in column chromatographic adsorption processes for herbal medicine. Acupunct Herbal Med. 2022;2(4):253260. doi:10.1097/HM9.0000000000000035

46. Arabi YM, Myatra SN, Lobo SM. Surging ICU during COVID-19 pandemic: an overview. Curr Opin Crit Care. 2022;28(6):638644. doi:10.1097/MCC.0000000000001001

47. Machov K, Mach M, Porezan M. Deep learning in the detection of disinformation about COVID-19 in online space. Sensors. 2022;22(23):9319. doi:10.3390/s22239319

Read the rest here:
Research hotspots of deep learning in critical care medicine | JMDH - Dove Medical Press

Westend61/Getty Images

When you're buying a bulky item like a sofa on the internet, the last thing you need is for the product to arrive and for it to look nothing like you expected -- that kind of error is costly and frustrating not only for the person buying the sofa, but also for the company selling the item.

So, what if there was a way to use technology to make the buying experience more enjoyable and to reduce the risk of an unpleasant surprise on delivery day?

Also: This retailer is using RFID tags to make in-person clothes shopping less frustrating

That's exactly what Wayfair is doing with the help of machine learning. The e-commerce company, which sells furniture and home goods online, is using a specially designed platform that's been built alongside technology specialist Snorkel AI.

Tulia Plumettaz, director of machine learning at Wayfair, says the platform is helping her company boost the quality of the online search experience it provides to consumers and to help ensure the sofa you receive looks like the sofa you ordered.

"We have these bulky items that are hard to transport," she says. "We want you to get inspired and feel confident that what you're going to get is what you're buying. And we want that to happen without you even touching the product."

Also: Is Temu legit? What to know about this shopping app before placing your first order

Delivering this kind of high-quality online experience is far from straightforward. Wayfair's site includes thousands of products with a huge number of potential variables, including size, color, and texture.

An added complication comes from the fact that the e-commerce company provides a platform for its suppliers to sell goods to customers. Plumettaz says Wayfair sometimes receives a limited amount of information on products from its suppliers, so providing detailed descriptions to customers can be tough.

It's at that point that Snorkel's platform plays a key role in providing enriched product information.

"We want suppliers to find that it's easy to work with us using our advanced technology. We want them to say, 'I gave Wayfair a picture, some information, and -- with not a lot of effort -- my item just started selling,'" Plumettaz says.

Plumettaz also says machine learning supports "fast-labeling operations" through a bespoke solution that's been developed through a design partnership.

Also: The best online thrifting apps

Snorkel already had its key product called Snorkel Flow, which is a data-centric AI platform for automated data labeling, integrated model training, and analysis.

But while Snorkel Flow is focused on text, Wayfair needed a solution that would support the programmatic labeling of images.

Plumettaz says the solution, which was developed over a twelve-month period by the two companies in combination, provides benefits for both companies: Wayfair gets to shape the technology it's using, and Snorkel gets a route into a new and fast-emerging market.

"We engaged together, and the result is a novel development that brings programmatic labeling into computer vision," says Plumettaz.

Also: Here's why everything on Temu is so cheap

Now, with the bespoke technology in place, Wayfair's team can label and re-label products quickly and effectively.

Rather than having to rely on humans to label up to 40 million products manually, automation deals with a lot of the heavy lifting before specialists within the business -- such as category managers -- ensure the right images are served to online shoppers, says Plumettaz. "With programmatic image labeling, we can match products in the catalogue to the items that customers are looking for as new trends emerge."

Machine learning is also a productivity enhancement -- with less time being spent on labeling images, employees can now focus on higher-value activities. "It's making what we do a lot more interesting," she says. "At Wayfair, our employees don't lack activities to do -- think about maintaining such a rich catalogue. So, now we can be more productive. It's helped make our lives easier and our work a lot more cost-effective."

While Wayfair has chosen to work with Snorkel, Plumettaz recognizes there are other technology players who continue to develop their own machine-learning solutions.

Also: I bought four brand-name tech gadgets on Temu for work. Here's how it went

She says each company has its own stack and, in such a fast-developing market, it's tough to know where machine learning goes next. Plumettaz advises other professionals who are looking at emerging technology to make early inroads and build strong partnerships.

"The field is moving so fast," she says. "Five years ago, it was a lot harder to integrate with a vendor in machine learning. Now, the hurdles to get a vendor approved are disappearing fast."

While machine learning can provide a big boon to customer and employee experiences, Plumettaz says professionals shouldn't let emerging technology work in isolation.

Left to its own devices, an automated system might start labeling products wrongly, leading to unhappy customers and what she refers to as "tremendous consequences".

"You can have an amazing model, but the noise that can come your way through a 1% error rate -- such as when a bulky item gets delivered to your home and it's wrong -- is huge."

The lesson for all business leaders is to ensure the human stays in the loop in what remains a nascent area of development.

Also: I bought some off-brand geeky stuff from Temu (and wasn't mad about it)

"It's a journey with a lot of these applications," she says. "Let's automate, but let's still have a layer that is checking that the automation is working."

Plumettaz provides more details about how that process works at Wayfair. "When we're not confident, we put the outputs in front of a human and get some feedback," she says. "I call it the path to automation. It's like a toddler; it's not yet an adult who can run. And that's the framework that we've been using for those kinds of applications."

Another lesson for professionals who are thinking about dabbling in machine learning is to focus on cross-organization integration and processes, especially in terms of how the technology is implemented, used, and exploited.

Plumettaz says the takeaway will be a familiar one for professionals who introduce new systems or systems: Don't just implement technology for the sake of it. "Partnering really closely with business owners and product owners is key," she says. "I think the blocker is less around the technology and more around thinking about machine learning as a business-value driver from the get-go."

See the article here:
Machine learning helps this company deliver a better online shopping experience - ZDNet

July 26, 2023 -A new study described that despite the continued threat of infectious diseases on public health, the capabilities of artificial intelligence (AI) and machine learning (ML) can help handle this issue and provide a framework for future pandemics.

Regardless of research and biological advancements, infectious diseases remain an issue. To keep up with the conflict, common methods that are applied include therapies and diagnostics. Often, synthetic biology approaches provide a platform for innovation. Research indicated that synthetic biology is often divided into two development categories: quantitative biological hypotheses and data from experimentation, and the comprehension of the factors such as nucleic acids and peptides, which allow for the control of biology.

According to research, advancements in AI have considered these factors. Given the complexities of biology and infectious disease, there is a high level of potential. Thus, researchers reviewed how the relationship between AI and synthetic biology can battle infectious diseases.

The review described three uses of AI in infectious diseases: anti-infective drug discovery, infection biology, and diagnostics.

Despite the pre-existence of various anti-infective drugs, drug resistance often outmatches their effectiveness. AI and ML can play a large role in developing new drugs by searching small-molecule databases while using training models to define new drugs or apply existing drugs.

The complications of infection biology are extensive, largely due to the activity of bacterial, eukaryotic, and viral pathogens. These factors can affect host responses, and, therefore, the course of infection.

ML models, however, can analyze nucleic acid, protein, and other variables to determine the aspects of hostpathogen interactions and immune responses. Research also indicates they can define genes and interactions between proteins that link to host cell changes, immunogenicity prediction, and other activities.

Also, gene expression optimization and antigen prediction has assisted the development of vaccines and drugs through supervised models.

AI and ML have applications in diagnostics. As prior instances have shown, the speed of infectious disease detection plays a large role in how spreading takes place. However, through AI and ML, researchers can identify infections and foresee drug resistance. This is primarily because of its ability to program elements well and highlight essential information from biomolecular networks.

Regardless of the opportunities and challenges that these methods may pose, they are essential to the future of infectious disease treatment. As the development of AI continues, it is critical to consider a wide range of datasets to avoid bias.

Various research efforts have also showcased the capabilities of AI and how it may advance healthcare.

Research from April 2022, for example, involved the creation of an AI model that uses non-contrast abdominal CT images to analyze factors related to pancreatic health, determining type 2 diabetes risk.

Using hundreds of images and various measurements, researchers defined the factors that correlated with diabetes. Consistent and accurate results allowed researchers to determine this analysis was an effective approach to detecting diabetes.

This study is a step towards the wider use of automated methods to address clinical challenges, said study authors Ronald M. Summers, MD, PhD, and Hima Tallam, an MD and PhD student, in apress release.It may also inform future work investigating the reason for pancreatic changes that occur in patients with diabetes.

Research efforts such as these are integral examples of how AI continues to play a role in healthcare.

See the rest here:
AI and Machine Learning Hold Potential in Fighting Infectious Disease - HealthITAnalytics.com

When it comes to tech skills, machine learning is only getting hotter. Companies of all sizes want tech professionals who can build and manage self-learning A.I. models, and then integrate those models into all kinds of next-generation apps and services.

According tolevels.fyi, which crowdsources compensation information for various tech roles,compensation for those specializing in machine learning and A.I. has increased 2.7 percent over the past six months, from an average of $246,000 per year to $252,535. Those with lots of experience and skill in machine learning can command exponentially higher salaries, of course, especially at big companies known for extremely high pay.

But what does it take to launch yourself onto a machine learning engineer career path, and once youre there, what sort of options are available to you? Lets dive in.

Before embarking on this career path, its important to have a solid foundation in computer science and math, including an understanding of how computers and algorithms work.

Programming is an essential skill, and multiple coding languages may be required, depending on the role and company. Python and JavaScript are often the most popular programming languages that aspiring machine learning engineers focus on first, followed by supporting frameworks like TensorFlow and PyTorch.

Once you've built that foundation and developed your core skill sets, the next step is to start applying what youve learned, explains Neil Ouellette, senior machine learning engineer at Skillsoft. Its important to gain hands-on practice and experience by experimenting with different algorithms and creating small projects on Github.

This is a good way to not only sharpen your skills, but to build a portfolio of work that you can eventually share with prospective employers, he explains.

Do you need formal education to become a machine learning engineer? Thats a great question. Given the demand for ML and AI engineers, many companies are willing to hire tech professionals who dont have a formal two- or four-year degree, provided they can prove during the interview process that they have the skills necessary to succeed in the role. Before you begin applying for jobs, make sure you have a solid grasp on the following, which pop up frequently as requirements for machine learning engineer roles:

In order to carry out these tasks, youll need to have mastered the following:

Mehreen Tahir, software engineer at New Relic, says that entry-level machine learning engineers are often responsible for preprocessing and cleaning data, implementing and testing different machine learning models, and possibly deploying these models.

This involves a lot of data wrangling and debugging, but it's an essential part of the learning process, he says. I always recommend beginners to start working on their own projects or participate in online competitions like those on Kaggle.

These experiences can give you invaluable insights into the practical challenges of machine learning; theyll also help bulk out your resume and application materials when you begin applying for roles in earnest.

An entry-level machine learning engineer (often titled as a junior machine learning engineer or machine learning intern) typically fits into the data science or engineering department of an organization. Some of the typical tasks might include assisting in the development of machine learning models, with lots of collaboration with data analysts and data scientists. As with most tech jobs, a solid grasp of soft skills such as communication and empathy is essential for anyone who wants to make a career out of machine learning.

You might help in building and testing models under the supervision of senior team members, gathering and cleaning data, and learning to interpret and present results, Ouellette says. Many organizations expect their machine learning teams to stay current with the latest techniques and methodologies, and you may be asked to help with this.

There are many pathways for career advancement as a machine learning engineer, whether one is interested in being a manager or individual contributor.

"After gaining some years of experience and expertise, you can advance to a senior role," Ouellette says. "These engineers usually oversee project management, design systems on a larger scale, and may mentor junior engineers."

These potential roles could include a senior machine learning engineer, lead machine learning engineer or team lead, data scientist, AI specialist, machine learning architect, or research scientist. Given the popularity of machine learning, mastering its fundamentals can open an incredible number of career tracks that increasingly rely on the technology.

In the role of team lead, you would oversee and lead a team of machine learning engineers, Ouellette explains. This includes making key decisions on behalf of the team and owning the whole machine learning development process.

In companies that heavily rely on data or A.I., advancing to the executive roles of chief data officer or chief A.I. officer means one is responsible for establishing A.I. and data-related strategies at the highest level.If you find you have a knack for handling clients and translating business problems into data problems, a move into a data science role could be a good fit, Tahir notes.

Data scientists often do a bit of everything, from understanding the business context to data analysis to communicating results in a way that non-technical folks can understand. Soft skills matter more than ever if youre interested in management and want to eventually run your own team.

In this role, you'd be less hands-on with the code and more involved in strategic decisions, team management, and liaising between your team and the rest of the organization, Tahir says. If you're deeply interested in the theoretical side of machine learning and want to push the boundaries of what's possible, you might consider going back to school to get a PhD and become a researcher.

Its important to remember these pathways aren't strictly linear, and the beauty of this field is that there's a lot of flexibility to shape one's own career based on a personal interests and skills. What do you want machine learning to do for you?

To ensure continuous progression in a career as a machine learning engineer, it's crucial to stay updated with the latest advancements in the field, especially as it evolves at a rapid pace. Tools, languages, and frameworks enjoy frequent iterations and updates; if you ignore them for too long, youll fall behind.

Maintaining your baseline knowledge involves taking online courses, attending workshops, webinars, or conferences, and regularly reading relevant research papers. Another key is to constantly work on challenging projects, either at work or in your spare time, that push the boundaries of your current skill set, Tahir says. This hands-on experience is invaluable and can often expose you to new tools and techniques.

Networking is also essential: joining professional groups, online communities, and attending industry events can help you machine learning pros connected, learn from peers, and open new opportunities.

From Tahir's perspective, it's also important to develop soft skills, including communication, teamwork, and problem-solving skills: These are vital, particularly as you move into more senior or managerial roles Demonstrating your ability to effectively communicate complex ideas to non-technical team members or stakeholders can significantly boost your career progression.

Ouellette agrees it's critical to know how to communicate with non-technical audiences. Although machine learning is inherently complex, youll often need to explain how algorithms and statistical models work with stakeholders or clients who may not have a technical background, he says. Strong communication skills are a must.

Go here to read the rest:
Machine Learning Engineer Career Path: What You Need to Know - Dice Insights

USC researchers will present nine papers at the 40th International Conference on Machine Learning (ICML 2023).

This year, USC researchers will showcase nine papers at the 40th International Conference on Machine Learning (ICML 2023), one of the most prestigious machine learning conferences, taking place July 23 29 in Honolulu, Hawaii. ICML brings together the artificial intelligence (AI) community to share new ideas, tools, and datasets, and make connections to advance the field.

Accepted papers with USC affiliation:

Moccasin: Efficient Tensor Rematerialization for Neural Networks Tu, Jul 25, 17:00Poster Session 2

Burak Bartan,Haoming Li,Harris Teague,Christopher Lott,Bistra Dilkina

Abstract: The deployment and training of neural networks on edge computing devices pose many challenges. The low memory nature of edge devices is often one of the biggest limiting factors encountered in the deployment of large neural network models. Tensor rematerialization or recompute is a way to address high memory requirements for neural network training and inference. In this paper we consider the problem of execution time minimization of compute graphs subject to a memory budget. In particular, we develop a new constraint programming formulation called textsc{Moccasin} with onlyO(n)integer variables, wherenis the number of nodes in the compute graph. This is a significant improvement over the works in the recent literature that propose formulations withO(n2)Boolean variables. We present numerical studies that show that our approach is up to an order of magnitude faster than recent work especially for large-scale graphs.

Refined Regret for Adversarial MDPs with Linear Function Approximation Tu, Jul 25, 17:00 Poster Session 2

Yan Dai,Haipeng Luo,Chen-Yu Wei,Julian Zimmert

Abstract: We consider learning in an adversarial Markov Decision Process (MDP) where the loss functions can change arbitrarily overKepisodes and the state space can be arbitrarily large. We assume that the Q-function of any policy is linear in some known features, that is, a linear function approximation exists. The best existing regret upper bound for this setting (Luo et al., 2021) is of order(K2/3)(omitting all other dependencies), given access to a simulator. This paper provides two algorithms that improve the regret to(K)in the same setting. Our first algorithm makes use of a refined analysis of the Follow-the-Regularized-Leader (FTRL) algorithm with the log-barrier regularizer. This analysis allows the loss estimators to be arbitrarily negative and might be of independent interest. Our second algorithm develops a magnitude-reduced loss estimator, further removing the polynomial dependency on the number of actions in the first algorithm and leading to the optimal regret bound (up to logarithmic terms and dependency on the horizon). Moreover, we also extend the first algorithm to simulator-free linear MDPs, which achieves(K8/9)regret and greatly improves over the best existing bound(K14/15). This algorithm relies on a better alternative to the Matrix Geometric Resampling procedure by Neu & Olkhovskaya (2020), which could again be of independent interest.

Searching Large Neighborhoods for Integer Linear Programs with Contrastive Learning We, Jul 26, 14:00 Poster Session 3

Taoan Huang,Aaron Ferber,Yuandong Tian,Bistra Dilkina,Benoit Steiner

Abstract: Integer Linear Programs (ILPs) are powerful tools for modeling and solving a large number of combinatorial optimization problems. Recently, it has been shown that Large Neighborhood Search (LNS), as a heuristic algorithm, can find high-quality solutions to ILPs faster than Branch and Bound. However, how to find the right heuristics to maximize the performance of LNS remains an open problem. In this paper, we propose a novel approach, CL-LNS, that delivers state-of-the-art anytime performance on several ILP benchmarks measured by metrics including the primal gap, the primal integral, survival rates and the best-performing rate. Specifically, CL-LNS collects positive and negative solution samples from an expert heuristic that is slow to compute and learns a new one with a contrastive loss. We use graph attention networks and a richer set of features to further improve its performance.

Fairness in Matching under UncertaintyWe, Jul 26, 17:00Poster Session 4

Siddartha Devic,David Kempe,Vatsal Sharan,Aleksandra Korolova

Abstract: The prevalence and importance of algorithmic two-sided marketplaces has drawn attention to the issue of fairness in such settings. Algorithmic decisions are used in assigning students to schools, users to advertisers, and applicants to job interviews. These decisions should heed the preferences of individuals, and simultaneously be fair with respect to their merits (synonymous with fit, future performance, or need). Merits conditioned on observable features are always emph{uncertain}, a fact that is exacerbated by the widespread use of machine learning algorithms to infer merit from the observables. As our key contribution, we carefully axiomatize a notion of individual fairness in the two-sided marketplace setting which respects the uncertainty in the merits; indeed, it simultaneously recognizes uncertainty as the primary potential cause of unfairness and an approach to address it. We design a linear programming framework to find fair utility-maximizing distributions over allocations, and we show that the linear program is robust to perturbations in the estimated parameters of the uncertain merit distributions, a key property in combining the approach with machine learning techniques.

On Distribution Dependent Sub-Logarithmic Query Time of Learned Indexing Th, Jul 27, 13:30Poster Session 5

Sepanta Zeighami,Cyrus Shahabi

Abstract: A fundamental problem in data management is to find the elements in an array that match a query. Recently, learned indexes are being extensively used to solve this problem, where they learn a model to predict the location of the items in the array. They are empirically shown to outperform non-learned methods (e.g., B-trees or binary search that answer queries inO(logn)time) by orders of magnitude. However, success of learned indexes has not been theoretically justified. Only existing attempt shows the same query time ofO(logn), but with a constant factor improvement in space complexity over non-learned methods, under some assumptions on data distribution. In this paper, we significantly strengthen this result, showing that under mild assumptions on data distribution, and the same space complexity as non-learned methods, learned indexes can answer queries inO(loglogn)expected query time. We also show that allowing for slightly larger but still near-linear space overhead, a learned index can achieveO(1)expected query time. Our results theoretically prove learned indexes are orders of magnitude faster than non-learned methods, theoretically grounding their empirical success.

SurCo: Learning Linear SURrogates for COmbinatorial Nonlinear Optimization ProblemsTh, Jul 27, 16:30Poster Session 6

Aaron Ferber,Taoan Huang,Daochen Zha,Martin Schubert,Benoit Steiner,Bistra Dilkina,Yuandong Tian

Abstract: Optimization problems with nonlinear cost functions and combinatorial constraints appear in many real-world applications but remain challenging to solve efficiently compared to their linear counterparts. To bridge this gap, we proposeSurCothat learns linearSurrogate costs which can be used in existingCombinatorial solvers to output good solutions to the original nonlinear combinatorial optimization problem. The surrogate costs are learned end-to-end with nonlinear loss by differentiating through the linear surrogate solver, combining the flexibility of gradient-based methods with the structure of linear combinatorial optimization. We propose threevariants:for individual nonlinear problems,for problem distributions, andto combine both distribution and problem-specific information. We give theoretical intuition motivating, and evaluate it empirically. Experiments show thatfinds better solutions faster than state-of-the-art and domain expert approaches in real-world optimization problems such as embedding table sharding, inverse photonic design, and nonlinear route planning.

Emergent Asymmetry of Precision and Recall for Measuring Fidelity and Diversity of Generative Models in High Dimensions Th, Jul 27, 13:30Poster Session 5

Mahyar Khayatkhoei,Wael AbdAlmageed

Abstract: Precision and Recall are two prominent metrics of generative performance, which were proposed to separately measure the fidelity and diversity of generative models. Given their central role in comparing and improving generative models, understanding their limitations are crucially important. To that end, in this work, we identify a critical flaw in the common approximation of these metrics using k-nearest-neighbors, namely, that the very interpretations of fidelity and diversity that are assigned to Precision and Recall can fail in high dimensions, resulting in very misleading conclusions. Specifically, we empirically and theoretically show that as the number of dimensions grows, two model distributions with supports at equal point-wise distance from the support of the real distribution, can have vastly different Precision and Recall regardless of their respective distributions, hence an emergent asymmetry in high dimensions. Based on our theoretical insights, we then provide simple yet effective modifications to these metrics to construct symmetric metrics regardless of the number of dimensions. Finally, we provide experiments on real-world datasets to illustrate that the identified flaw is not merely a pathological case, and that our proposed metrics are effective in alleviating its impact.

Conformal Inference is (almost) Free for Neural Networks Trained with Early Stopping Th, Jul 27, 13:30Poster Session 5

Ziyi Liang,Yanfei Zhou,Matteo Sesia

Abstract: Early stopping based on hold-out data is a popular regularization technique designed to mitigate overfitting and increase the predictive accuracy of neural networks. Models trained with early stopping often provide relatively accurate predictions, but they generally still lack precise statistical guarantees unless they are further calibrated using independent hold-out data. This paper addresses the above limitation with conformalized early stopping: a novel method that combines early stopping with conformal calibration while efficiently recycling the same hold-out data. This leads to models that are both accurate and able to provide exact predictive inferences without multiple data splits nor overly conservative adjustments. Practical implementations are developed for different learning tasks outlier detection, multi-class classification, regression and their competitive performance is demonstrated on real data.

A Critical View of Vision-Based Long-Term Dynamics Prediction Under Environment MisalignmentTu, Jul 25, 17:00Poster Session 2

Hanchen Xie,Jiageng Zhu,Mahyar Khayatkhoei,Jiazhi Li,Mohamed Hussein,Wael AbdAlmageed

Abstract: Dynamics prediction, which is the problem of predicting future states of scene objects based on current and prior states, is drawing increasing attention as an instance of learning physics. To solve this problem, Region Proposal Convolutional Interaction Network (RPCIN), a vision-based model, was proposed and achieved state-of-the-art performance in long-term prediction. RPCIN only takes raw images and simple object descriptions, such as the bounding box and segmentation mask of each object, as input. However, despite its success, the models capability can be compromised under conditions of environment misalignment. In this paper, we investigate two challenging conditions for environment misalignment: Cross-Domain and Cross-Context by proposing four datasets that are designed for these challenges: SimB-Border, SimB-Split, BlenB-Border, and BlenB-Split. The datasets cover two domains and two contexts. Using RPCIN as a probe, experiments conducted on the combinations of the proposed datasets reveal potential weaknesses of the vision-based long-term dynamics prediction model. Furthermore, we propose a promising direction to mitigate the Cross-Domain challenge and provide concrete evidence supporting such a direction, which provides dramatic alleviation of the challenge on the proposed datasets.

Published on July 25th, 2023

Last updated on July 26th, 2023

Go here to see the original:
USC at the International Conference on Machine Learning (ICML ... - USC Viterbi School of Engineering