Academic literature on the topic 'Un/self-supervised learning'

Lately, deep learning technology has been extensively investigated for accelerating dynamic magnetic resonance (MR) imaging, with encouraging progresses achieved. However, without fully sampled reference data for training, the current approaches may have limited abilities in recovering fine details or structures. To address this challenge, this paper proposes a self-supervised collaborative learning framework (SelfCoLearn) for accurate dynamic MR image reconstruction from undersampled k-space data directly. The proposed SelfCoLearn is equipped with three important components, namely, dual-network collaborative learning, reunderampling data augmentation and a special-designed co-training loss. The framework is flexible and can be integrated into various model-based iterative un-rolled networks. The proposed method has been evaluated on an in vivo dataset and was compared to four state-of-the-art methods. The results show that the proposed method possesses strong capabilities in capturing essential and inherent representations for direct reconstructions from the undersampled k-space data and thus enables high-quality and fast dynamic MR imaging.

We introduce Perceiving Stroke-Semantic Context (PerSec), a new approach to self-supervised representation learning tailored for Scene Text Recognition (STR) task. Considering scene text images carry both visual and semantic properties, we equip our PerSec with dual context perceivers which can contrast and learn latent representations from low-level stroke and high-level semantic contextual spaces simultaneously via hierarchical contrastive learning on unlabeled text image data. Experiments in un- and semi-supervised learning settings on STR benchmarks demonstrate our proposed framework can yield a more robust representation for both CTC-based and attention-based decoders than other contrastive learning methods. To fully investigate the potential of our method, we also collect a dataset of 100 million unlabeled text images, named UTI-100M, covering 5 scenes and 4 languages. By leveraging hundred-million-level unlabeled data, our PerSec shows significant performance improvement when fine-tuning the learned representation on the labeled data. Furthermore, we observe that the representation learned by PerSec presents great generalization, especially under few labeled data scenes.

Machine learning approaches, such as soft independent modeling of class analogy (SIMCA) and pathway analysis, were introduced in depression research in the 1990s (Maes et al.) to construct neuroimmune endophenotype classes. The goal of this paper is to examine the promise of precision psychiatry to use information about a depressed person’s own pan-omics, environmental, and lifestyle data, or to tailor preventative measures and medical treatments to endophenotype subgroups of depressed patients in order to achieve the best clinical outcome for each individual. Three steps are emerging in precision medicine: (1) the optimization and refining of classical models and constructing digital twins; (2) the use of precision medicine to construct endophenotype classes and pathway phenotypes, and (3) constructing a digital self of each patient. The root cause of why precision psychiatry cannot develop into true sciences is that there is no correct (cross-validated and reliable) model of clinical depression as a serious medical disorder discriminating it from a normal emotional distress response including sadness, grief and demoralization. Here, we explain how we used (un)supervised machine learning such as partial least squares path analysis, SIMCA and factor analysis to construct (a) a new precision depression model; (b) a new endophenotype class, namely major dysmood disorder (MDMD), which is a nosological class defined by severe symptoms and neuro-oxidative toxicity; and a new pathway phenotype, namely the reoccurrence of illness (ROI) index, which is a latent vector extracted from staging characteristics (number of depression and manic episodes and suicide attempts), and (c) an ideocratic profile with personalized scores based on all MDMD features.

Cellulose is the main raw material for the production of paper. Companies that produce it present in their production line the cutting of the cellulose sheet. This failure is sporadic and has a high economic impact since it paralyzes the production line for several hours, incurring unproductive hours and a large deployment of human and financial resources. In this research, the use of Data Mining is proposed to define a machine learning algorithm that allows predicting the cutting of the cellulose sheet in a production line of a cellulose plant in Chile. The results show that by applying this technique it is possible to predict the cutting of the cellulose sheet well in advance to take corrective actions to avoid cutting and thus minimize the economic impact associated with the failure. Keywords: Data Mining, machine learning, cellulose, productivity. References [1]B. Ranaganth y G. Viswanath, «Application of artificial neural network for optimizing cutting variables in laser cutting of 304 grade stainless steel,» International Journal of Applied Engineering and Technology, vol. 1, nº 1, pp. 106-112, 2011. [2]M. Durica, J. Frnda y L. Svabova, «Decision tree based model of business failure prediction for Polish companies,» Oeconomia Copernicana, vol. 10, nº 3, pp. 453-469, 2019. [3]G. Köksal, İ. Batmaz y M. C. Testik, «A review of data mining applications for quality improvement in manufacturing industry,» Expert systems with Applications, vol. 38, nº 10, pp. 13448-13467, 2011. [4]H. Poblete y R. Vargas, «Relacion entre densidad y propiedades de tableros HDF producidos por un proceso seco,» Maderas. Ciencia y tecnología, vol. 8, nº 3, pp. 169-182, 2006. [5]B. Kovalerchuk y E. Vityaev, «Data mining for financial applications,» Data Mining and Knowledge Discovery Handbook, pp. 1203-1224, 2005. [6]U. Fayyad, G. Piatetsky-Shapiro, P. Smyth y R. Uthurusamy, «Advances in knowledge discovery and data mining,» American Association for Artificial Intelligence, 1996. [7]A. K. Pandey y A. K. Dubey, «Neuro fuzzy modeling of laser beam cutting process,» Applied Mechanics and Materials, vol. 110, pp. 4109-4117, 2012. [8]M. Németh y G. Michaľčonok, «Preparation and cluster analysis of data from the industrial production process for failure prediction,» Research Papers Faculty of Materials Science and Technology Slovak University of Technology, vol. 24, nº 39, pp. 111-116, 2016. [9]S. Ballı, «A data mining approach to the diagnosis of failure modes for two serial fastened sandwich composite plates,» Journal of Composite Materials, vol. 51, nº 20, pp. 2853-2862, 2017. [10]S. Dindarloo y E. Siami-Irdemoosa, «Data mining in mining engineering: results of classification and clustering of shovels failures data,» International Journal of Mining, Reclamation and Environment, vol. 31, nº 2, pp. 105-118, 2017. [11]E. e Oliveira, V. Miguéis, L. Guimarães y J. L. Borges, «Power Transformer Failure Prediction: Classification in Imbalanced Time Series,» U. Porto Journal of Engineering, vol. 3, nº 2, pp. 34-48, 2017. [12]A. Taghizadeh y N. Demirel, «Application of Machine Learning for Dragline Failure Prediction,» E3S Web of Conferences, vol. 15, p. 03002, 2017. [13]W. Chang, Z. Xu, M. You, S. Zhou, Y. Xiao y Y. Cheng, «A Bayesian Failure Prediction Network Based on Text Sequence Mining and Clustering,» Entropy, vol. 20, nº 12, p. 923, 2018. [14]K. Halteh, K. Kumar y A. Gepp, «Financial distress prediction of Islamic banks using tree-based stochastic techniques,» Managerial Finance, vol. 44, nº 6, pp. 759-773, 2018. [15]C.-H. Liu, C.-J. Lin, Y.-H. Hu y Z.-H. You, «Predicting the failure of dental implants using supervised learning techniques,» Applied Sciences, vol. 8, nº 5, p. 698, 2018. [16]B. Mohammed, I. Awan, H. Ugail y M. Younas, «Failure prediction using machine learning in a virtualised HPC system and application,» Cluster Computing, vol. 22, nº 2, pp. 471-485, 2019. [17]O. Sukhbaatar, T. Usagawa y L. Choimaa, «An artificial neural network based early prediction of failure-prone students in blended learning course,» International Journal of Emerging Technologies in Learning (iJET)}, vol. 14, nº 19, pp. 77-92, 2019. [18]Z. Wang, W. Zhao y X. Hu, «Analysis of prediction model of failure depth of mine floor based on fuzzy neural network,» Geotechnical and Geological Engineering, vol. 37, nº 1, pp. 71-76, 2019. [19]V. S. Gujre y R. Anand, «Machine learning algorithms for failure prediction and yield improvement during electric resistance welded tube manufacturing,» Journal of Experimental \& Theoretical Artificial Intelligence, vol. 32, nº 4, pp. 601-622, 2020. [20]P. du Jardin, «Forecasting corporate failure using ensemble of self-organizing neural networks,» European Journal of Operational Research, vol. 288, nº 3, pp. 869-885, 2021. [21]R. Brachman y T. Anand, «The process of knowledge discovery in databases,» Advances in knowledge discovery and data mining, pp. 37-57, 1996. [22]W. Frawley, G. Piatetsky-Shapiro y C. Matheus, «Knowledge discovery in databases: An overview,» AI magazine, vol. 13, nº 3, p. 57, 1992. [23]F. H. Troncoso Espinosa y J. V. Ruiz Tapia, «Predicción de fuga de clientes en una empresa de distribución de gas natural mediante el uso de minería de datos,» Universidad Ciencia y Tecnología, vol. 24, nº 106, pp. 79-87, 2020. [24]F. H. Troncoso, «Prediction of Recidivism in Thefts and Burglaries Using Machine Learning,» Indian Journal of Science and Technology, vol. 13, nº 6, pp. 696-711, March 2020. [25]M. Kantardzic, Data mining: concepts, models, methods, and algorithms, John Wiley & Sons, 2011. [26]F. H. Troncoso Espinosa, P. G. Fuentes Figueroa y I. R. Belmar Arriagada, «Predicción de fraudes en el consumo de agua potable mediante el uso de Minería de Datos,» Universidad Ciencia y Tecnología, vol. 24, nº 104, pp. 58-66, 2020. [27]C. Romero y S. Ventura, «Data mining in education,» Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 3, nº 1, pp. 12-27, 2013. [28]D. Larose y C. Larose, Discovering knowledge in data: an introduction to data mining, John Wiley & Sons, 2014.

Abstract Computer vision algorithms have been utilized for 3-D road imaging and pothole detection for over two decades. Nonetheless, there is a lack of systematic survey articles on state-of-the-art (SoTA) computer vision techniques, especially deep learning models, developed to tackle these problems. This article first introduces the sensing systems employed for 2-D and 3-D road data acquisition, including camera(s), laser scanners and Microsoft Kinect. It then comprehensively reviews the SoTA computer vision algorithms, including (1) classical 2-D image processing, (2) 3-D point cloud modelling and segmentation and (3) machine/deep learning, developed for road pothole detection. The article also discusses the existing challenges and future development trends of computer vision-based road pothole detection approaches: classical 2-D image processing-based and 3-D point cloud modelling and segmentation-based approaches have already become history; and convolutional neural networks (CNNs) have demonstrated compelling road pothole detection results and are promising to break the bottleneck with future advances in self/un-supervised learning for multi-modal semantic segmentation. We believe that this survey can serve as practical guidance for developing the next-generation road condition assessment systems.

Objetivo: conhecer os significados e as percepções sobre o processo de avaliação dos estágios supervisionados sob a ótica de enfermeiros docentes. Método: trata-se de estudo qualitativo, descritivo, exploratório e transversal, com 15 enfermeiros na função de docente e supervisores de estágios do curso de graduação em Enfermagem. Utilizaram-se dois instrumentos para a coleta de dados. Empregou-se, para a análise, o Discurso do Sujeito Coletivo. Resultados: emergiram-se, do primeiro tema, as ideias centrais: “meio de formação profissional”; “processo contínuo para melhorar a prática”; “preparo para a vida profissional”; “relação teoria e a prática”; “diversos significados”; “reta final”; “feedback”; “amadurecimento do conhecimento”; “valor dado ao desempenho do aluno”. Identificaram-se, do segundo tema, as ideias centrais: “diversas percepções”; “instrumentos e processo pontual”; “presença de critérios”; “dinâmico, gradativo e contínuo”; “diversos aspectos”; “associação teoria e a prática”; “maturidade”; “coerência e autoavaliação”. Conclusão: demonstrou-se a preocupação em formar pessoas críticas e comprometidas em desenvolver a futura profissão, portanto, tendo a avaliação por competência identificando erros e acertos no ensino e aprendizagem. Descritores: Avaliação; Docente; Estágio; Supervisão; Ensino; Aprendizagem.AbstractObjective: to know the meanings and perceptions about the evaluation process of supervised internships from the perspective of teaching nurses. Method: this is a qualitative, descriptive, exploratory and cross-sectional study, with 15 nurses working as teachers and supervisors of undergraduate nursing courses. Two instruments were used for data collection. For the analysis, the Collective Subject Discourse was used. Results: emerged from the first theme, the central ideas: “means of vocational training”; “Continuous process to improve practice”; “Preparation for professional life”; “Relationship theory and practice”; “Various meanings”; “Final straight”; Feedback; “Maturation of knowledge”; “Value given to student performance”. From the second theme, the main ideas were identified: “diverse perceptions”; “Instruments and timely process”; “Presence of criteria”; “Dynamic, gradual and continuous”; “Various aspects”; “Association theory and practice”; "maturity"; “Coherence and self-assessment”. Conclusion: The concern with forming critical people committed to developing the future profession was demonstrated, therefore, having the competence assessment identifying errors and successes in teaching and learning. Descriptors: Evaluation; Teacher; Internship; Supervision; Teaching; Learning. ResumenObjetivo: conocer los significados y percepciones sobre el proceso de evaluación de pasantías supervisadas desde la perspectiva de los profesores enfermeros. Método: este es un estudio cualitativo, descriptivo, exploratorio y transversal, con 15 enfermeros que trabajan como profesores y supervisores de cursos de pregrado en Enfermería. Se utilizaron dos instrumentos para la recopilación de datos. Para el análisis, se utilizó el Discurso del Sujeto Colectivo. Resultados: surgieron del primer tema, las ideas centrales: "medios de formación profesional"; "proceso continuo para mejorar la práctica"; "preparación para la vida profesional"; " relación teoría y práctica"; "varios significados"; "recta final"; “feedback”; "maduración del conocimiento"; "valor dado al rendimiento del alumno". A partir del segundo tema, se identificaron las ideas centrales: "percepciones diversas"; "instrumentos y proceso puntual"; "presencia de criterios"; "dinámico, gradual y continuo"; "varios aspectos"; "asociación teoría y práctica"; "madurez"; "coherencia y autoevaluación". Conclusión: se demostró la preocupación por formar personas críticas comprometidas con el desarrollo de la profesión futura, por lo tanto, tener la evaluación de competencia identificando errores y éxitos en la enseñanza y el aprendizaje. Descriptores: Evaluación; Docente; Prácticas Profesionales; Supervisión; Enseñanza; Aprendiendo.

Background: Canadian residency programs have adopted competency-based medical education, where time-based learning systems are replaced with core competency “milestones” that must be achieved before a student progresses. Assessment tools must be developed to predict performance prior to high-stakes milestones, so interventions can be targeted to improve performance.Objectives: 1. To characterize how well each of three practice written exams predicts passing the Canadian Internal Medicine Royal College (RC) exam. 2. To determine if writing practice exams is perceived to improve performance on the RC exam.Methods: Three 105-question multiple choice question exams were created from a range of internal medicine topics, and offered one month apart to 35 residents. Percentile ranks on each practice exam were compared to the result (pass/fail) on the RC exam. Surveys were completed within 1 month after the RC exam.Results: There were 35 residents invited to participate. Practice exams (PE) 1, 2, and 3 were taken by 33, 26, and 22 residents, for an exam participation rate of 94.3, 74.3, and 62.9%, respectively. Failure on the RC exam could be predicted by percentile ranking <15% on PE1 (OR 19.5, p=0.017) or PE2 (OR 63.0, p=0.006), and by percentile ranking <30% on PE1 (OR 28.8, p=0.003), PE2 (OR 24.0, p=0.010) or PE3 (OR 15.0, p=0.046). The survey was sent out to the 33 participants. Of those, the total number of respondents was 25, with a response rate of 75.5%. Survey takers agreed that practice written exams improved performance on the RC exam (18/25, 88%).Conclusions: Performance in the Canadian Internal Medicine RC Exam can be predicted by performance on any of three practice written exams. This tool can therefore identify trainees for whom additional resources should be invested, to prevent failure of a high-stakes milestone within the competency based medical education framework.RÉSUMÉContexte : Les programmes canadiens de résidence ont choisi de diffuser un enseignement médical axé sur les compétences dans lequel les systèmes d’apprentissage structurés en fonction du temps sont remplacés par des « jalons » liés aux compétences fondamentales que l’étudiant doit atteindre pour aller de l’avant. Il faut élaborer des outils d’évaluation pour prédire la probabilité de résultats escomptés par un étudiant avant que celui-ci ne se présente à certains événements dont les enjeux sont élevés. Ainsi, il devient possible d’intervenir de manière à améliorer les résultats escomptés.Objectifs : 1. Déterminer dans quelle mesure chacun des trois examens de pratique écrits prédit la réussite à l’examen du Collège royal des médecins et chirurgiens du Canada (CRMCC) en médecine interne; 2. Évaluer si le fait de se soumettre à des examens de pratique écrits est perçu comme un élément qui améliore les résultats à l’examen du CRMCC.Méthodologie : Trois examens écrits comportant chacun 105 questions à choix de réponses portant sur un éventail de sujets relatifs à la médecine interne ont été préparés et proposés à 35 résidents à intervalle d’un mois. Les rangs-centiles de chaque examen de pratique ont été comparés avec le résultat obtenu à l’examen du CRMCC (succès/échec). Les sondages ont été effectués dans le mois suivant l’examen du CRMCC.Résultats : Trente-cinq résidents ont été invités aux examens de pratique écrits (EP) 1, 2 et 3. La participation a été respectivement de 33, 26 et 22 résidents, soit de 94,3 %, 74,3 % et 62,9 %. L’échec à l’examen du CRMCC pouvait être prédit par un rang-centile < 15 % à l’EP1 (OR 19,5 et p = 0,017) ou à l’EP2 (OR 63,0 et p = 0,006) et un rang-centile < 30 % à l’EP1 (OR 28,8 et p = 0,003), à l’EP2 (OR 24,0 et p = 0,010) ou à l’EP3 (OR 15,0, et p = 0,046). Le sondage a été envoyé aux 33 participants. Le nombre total de répondants a été de 25, pour un taux de réponse de 75,5 %. La majorité des répondants (18/25, 88 %) sont d’avis que les examens de pratique écrits leur ont permis d’obtenir de meilleurs résultats à l’examen du CRMCC.Conclusions : Les résultats à l’examen du Collège royal des médecins et chirurgiens du Canada (CRMCC) en médecine interne peuvent être prédits par les résultats obtenus à l’un des examens de pratique écrits. Par conséquent, cet outil peut être utilisé dans le cadre de l’enseignement de la médecine axé sur les compétences pour identifier sur qui l’on devrait investir des ressources additionnelles en vue d’éviter un échec à cet événement aux enjeux élevés.Competency-based medical education (CBME) has generated increased attention over the last decade,1–3 and become entrenched within several national medical education frameworks including Canada.4 Proponents of CBME suggest that older medical education models focus on medical knowledge rather than skills, or higher order aspects of practice. 5 Focus on time spent in training can take away from the abilities acquired during that time frame.6 Furthermore, flexible time periods may be more efficient and focused, compared to time-based curriculum.3,6,7 In light of these advantages, the Royal College of Physicians and Surgeons of Canada (RCPSC) has committed to transform medical education to a CBME model for all residency programs by 2017.4While residency programs reorganize toward the CBME model, residents will still be required to perform oral and written exams. It is thus essential that CBME-based programs incorporate assessment tools to predict performance on high-stakes milestones, such as RC exams.We created three written PEs, and evaluated how well each predicted performance on a high-stakes milestone, the RCPSC Internal Medicine exam (RC exam). We also evaluated how well PE were perceived to improve performance on the same high-stakes milestone RC exam.METHODSSetting and ParticipantsThe RC exam contains both written and oral components. All residents sitting both components of the RC exam, within 12 months, who were post-graduate medical residents at Western University (London, Ontario, Canada), were invited to participate. The study was conducted in 2013-2014.Western University Health Sciences Research Ethics Board provided an ethics waiver for this study, since the study was performed as part of the standard operations of an educational program.Intervention: ExamsTwo authors (HS, BT) separately created multiple choice questions (MCQ) reflecting all areas of internal medicine, based on the Objectives of Training of the RC Internal Medicine exam. RC exam questions are not available for purchase, and examinees are forbidden to share RC exam questions. Therefore, PE content and question style was informed by questions purchased for American Board of Internal Medicine (ABIM) course reviews.8,9 MCQ creators had each completed the RC exam within 3 years, and were thus familiar with MCQ and exam format.All authors independently reviewed each PE question to assure quality of content, grammar, spelling, and syntax. Each PE covered all subspecialty areas within internal medicine, including allergy and immunology ( n=4), cardiology (n=13), dermatology (n=2), endocrinology (n=8), gastroenterology (n=10), hematology (n=10), infectious diseases (n=15), nephrology ( n=9), neurology (n=6), oncology (n=4), respiratory and critical care medicine (n=7), rheumatology ( n=14) and statistics (n=3). This topic allocation included 7 questions for JAMA Rational Clinical Exam, and 5 for interpretation of medical images (e. g., chest X-ray, electrocardiogram). PE size (105 questions) and length (3 hours) were chosen to reflect the RC exam.Each PE was offered at two separate times, to assure flexibility within ongoing clinical responsibilities. PE1, PE2, and PE3 were offered approximately 7, 6, and 5 months prior to RC exam, respectively. This timing was chosen so that trainees had sufficient time to improve their performance before the RC exam if a poor PE result was found.Examinees were provided a personalized report for each exam, within 7 days of completing the PE. The personalized report included the examinee’s overall mark, average within each subspecialty, and percentile rank within the entire cohort of examinees. Two separate 1-hour periods were available to review each PE results, with the questions and key, supervised by BT.Intervention: SurveyAll study participants were invited to participate in a survey. The survey assessed how well PE simulated the RC exam, whether the PE were recommended to the next year’s cohort of examinees, and whether the PE improved performance on the RC exam.Outcomes: ExamsEach study participant agreed to provide the RC exam result (“pass” or “fail”) once he or she had received it. Each candidates verbally communicated RC exam result was confirmed online 3 months after the RC exam results were reported (cpso.on.ca).Odds ratios were calculated. The adverse outcome was failure on the RC exam. Exposures evaluated included percentile rank < 15% and <30%. Odds ratios of infinity were prevented by adding 1 adverse outcome to any exposure group without any adverse outcomes; this was performed for 3 exposure groups, but did not impact whether statistical significance was attained. Results are detailed in Table 1.Outcome: SurveySurvey results were on a Likert Scale. The proportion of those respondents who agreed or disagreed were calculated.All data was analyzed using Statistical Package for the Social Sciences (SPSS) version 21.0.RESULTSSetting and ParticipantsThere were 35 residents invited to participate, the total number of possible participants. PE1, PE2, and PE3 were taken by 33, 26, and 22 residents, for an exam participation rate of 94.3, 74.3, and 62.9%, respectively. The majority of invitees took 3 (n=17) or 2 ( n=14), while a minority took 1 (n=2) or 0 (n=2) practice exams.ExamsOf all examinees of the RC exam (n=35), 7 failed. RC exam pass rates were lower when PE1 percentile rank was lower than 15% (40.0 vs. 92.9%, p<0.001) or 30% (44.4 vs. 100%, p<0.004), when PE2 percentile rank was lower than 15% (0.0 vs. 100.0%, p <0.001) or 30% (42.9 vs. 100.0%, p=0.038), and when PE3 percentile rank was lower than 30% (50 vs. 93.75%, p=0.046) (Figure 1). Figure 1. License exam practice pass rate versus percentile rate (PR) on practice exams. Examinees were more likely to fail the RC exam if percentile rank was less than 15% (OR 19.5, p=0.017) or 30% (OR 28.8, p=0.003) in PE1, less than 15% (OR 63.0, p=0.006) or 30% (OR 24.0, p =0.010) in PE2, or less than 30% (OR 15.0, p=0.046) in PE3.SurveyOnly residents who had taken at least 1 practice exam were invited to participate. The survey was sent out to the 33 participants, the total number of possible participants. Of those, the total number of respondents was 25, with a response rate of 75.5%. Most survey respondents agreed that the PEs were an accurate simulation of the written component of the RC examination (20/25, 80%) (Figure 2A). Most survey respondents agreed that the PEs improved performance on the RC written examinations (18/25, 72%) (Figure 2B). Most survey respondents recommended future residents to take the PEs (22/25, 88%) (Figure 2C).DISCUSSIONWe describe the creation of a tool to assess performance on a high-stakes milestone examination, the RC exam. This tool is easy to create, affordable, and is administered on a voluntary basis with high uptake amongst candidates writing the RC exam. The assessment tool has been shown to predict performance well so that resources can be invested in those at risk for failing.There is a possibility that mere participation in the assessment tool itself improves performance on this high-stakes exam. There were insufficient numbers of study participants to determine a correlation between number of exams taken and pass rates. Even still, unwillingness to participate in the study may reflect a general unwillingness to prepare, which means the results would be confounded and correlative rather than causative. One way to look into this is to perform a randomized trial in which half of residents take the assessment tool and the other half doesn’t. Unfortunately, almost all invited residents were anxious to participate, rendering such a possible study impossible. On the other hand, exam takers were able to communicate usefulness of the exam and to provide feedback on how it might be improved for future years.As CBME develops and becomes entrenched, there will continue to be a need to prepare for knowledge based written exams. This exam will continue to be considered a core competency between the stages to transition to practice. Thus, tools are needed to assess exam competence. This study confirms that such tools can and should be developed to assure that trainees are prepared.Ideally, residents with low performance would be identified early enough to intervene to change the outcome. It is uncertain what the ideal time frame is or what the intervention should be. It is reasonable to assume that taking the examination earlier in their training may allow candidates to become aware of their performance and implications thereof and implement earlier changes in learning strategies. For example, in past years, candidates contacted their program directors to ease the clinical workload to allow more study time. Others sought counselling and mentorship from staff, while others were self-directed in their learning and became more motivated to study. On the other hand, poor performance on this formative examination could potentially discourage some residents from studying if they felt their studying was futile. Future research efforts should focus on identifying which intervention is optimal to modify exam performance.The failure rate of 20% on the RC exam the year the study was conducted was unusually high for the program; however, this allowed for a correlation to be established between the PEs and the RC exam. The PEs were able to identify all those who failed the RC exam. However, there were those who scored below the 30th percentile on the PEs and still passed the RC exam. It must be kept in mind that the purpose of these formative exams is to identify those at high risk of failure so they can receive remedial support and improve their chances of passing. Thus, it is possible that through increased remedial support, those candidates who did poorly on the in-training exam managed to pass the RC exam.Only 2 trainees chose to write none of the PEs. While both of these trainees ultimately failed the RC exam, statistical significance could not be established due to the small sample size. It thus remains uncertain whether the act of writing PEs predicts passing the RC exam. However, the study objective was to identify candidates at high risk of failing the RC exam; the next step will be to determine which interventions can improve RC exam result. However, it must be acknowledged that PEs could not only identify candidates at risk of failing RC exam, but also improve their performance. This requires future study before any firm conclusions can be found.This study confirms that formative exams’ results can predict failure on the RC exam. The questions were written by authors who had recently written the RC exam, familiar with its format, and knowledgeable of the current Canadian guidelines, which are a focus of the actual RC exam. Because of confidentiality agreements with the RCPSC, actual RC exam questions can’t be shared, and thus can’t be used as part of the practice exams. However, we attempted to overcome this limitation by having all PE questions reviewed by at least 3 physicians who’d recently successfully completed the RC exam, to assure syntax and format was as similar as possible between PE and RC exam. Furthermore, this limitation does not impact the PEs statistically significant prediction of candidates at risk of failing the RC exam. The study objective was to identify candidates at risk of failing the RC exam, and the PEs are indeed a valid predictor of RC exam performance. There are limitations to this study. Firstly, this was a single centre study. However, Western University has a wide range of subspecialty programs available, and the trainees’ demographics resemble that at other Canadian centers. Secondly, new questions need to be created annually to reflect updated literature and guidelines; this requires ongoing commitment and dedication from staff. These “updated” exams could become more difficult to validate if candidates no longer fail the RC exam. However, if the act of taking the PE predicts passing RC, future research could focus on comparing RC pass rates at programs with and without the PEs. Thirdly, it's entirely possible that the use of questions from old RC exam would be more predictive, but these questions cannot be shared or used for PE due to the confidentiality agreement with the RCPSC. Therefore, creation of independent questions is still required. This is the first study of an assessment tool to predict performance on the Canadian internal medicine examination within the CBME framework. This strategy can easily be replicated and feedback is rapidly provided in a time sensitive manner. This could help trainees direct their preparation and identify knowledge gaps more easily.CONCLUSIONWe report an assessment tool to predict performance on the RC exam that can be a valid and useful form of feedback. This strategy can easily be replicated for other subspecialties or internal medicine programs. Future efforts need to focus on how the results can determine which interventions or learning strategies improve the results of candidates identified to be at risk for failing.DisclaimersThe authors declare they have no competing interest.The authors report no external funding source for this study.The authors declare no previous abstract or poster or research presentation or any online presentation of this study.REFERENCES 1. Caccia N, Nakajima A, Kent N. Competency-based medical education: the wave of the future. J Obstet Gynaecol Can 2015;37:349–53. 2. Carraccio C, Englander R, Gilhooly J, et al. Building a framework of entrustable professional activities, supported by competencies and milestones, to bridge the educational continuum. Acad Med 2016 ;92(3):324–30. doi: 10.1097/ACM.0000000000001141. 3. Carraccio C, Wolfsthal SD, Englander R, Ferentz K, Martin C. Shifting paradigms: from Flexner to competencies. Acad Med 2002;77:361–67. 4. Johnston C. Residents prepare for switch to competency-based medical education. CMAJ2013;185:1029. 5. Talbot M. Monkey see, monkey do: a critique of the competency model in graduate medical education. Med Educ 2004;38:587–92. 6. Long DM. Competency-based residency training: the next advance in graduate medical education. Acad Med 2000;75:1178–83. 7. Bell HS, Kozakowski SM, Winter RO. Competency-based education in family practice. Fam Med 1997;29:701–704.8. Mittman B. Frontrunners 2016: Internal Medicine Q&A Review: Syllabus Companion for Board Review/Practice Questions & Answers for the ABIM Exam. Aliso Viejo, CA: Frontrunners Publishing; 2016.9. Fischer C. Internal Medicine Question Book: Second Edition: Complete Preparation for the American Board of Internal Medicine Exam. New York, NY: Kaplan Publishing; 2009.

This thesis focuses on two fundamental problems in robotic vision, scene geometry understanding and camera tracking. While both tasks have been the subject of research in robotic vision, numerous geometric solutions have been proposed in the past decades. In this thesis, we cast the geometric problems as machine learning problems, specifically, deep learning problems. Differ from conventional supervised learning methods that using expensive annotations as the supervisory signal, we advocate for the use of geometry as a supervisory signal to improve the perceptual capabilities in robots, namely Geometry Self-supervision. With the geometry self-supervision, we allow robots to learn and infer the 3D structure of the scene and ego-motion by watching videos, instead of expensive ground-truth annotation in traditional supervised learning problems. Followed by showing the use of geometry for deep learning, we show the possibilities of integrating self-supervised models with traditional geometry-based methods as a hybrid solution for solving the mapping and tracking problem. We focus on an end-to-end mapping problem from stereo data in the first part of this thesis, namely Deep Stereo Matching. Stereo matching is one of the oldest problems in computer vision. Classical approaches to stereo matching typically rely on handcrafted features and a multiple-step solution. Recent deep learning methods utilize deep neural networks to achieve end-to-end trained approaches while significantly outperforming classic methods. We propose a novel data acquisition pipeline using an untethered device (Microsoft HoloLens) with a Time-of-Flight (ToF) depth camera and stereo cameras to collect real-world data. A novel semi-supervised method is proposed to train networks with ground-truth supervision and self-supervision. The large scale real-world stereo dataset with semi-dense annotation and dense self-supervision allow our deep stereo matching network to generalize better when compared to prior arts. Mapping and tracking using a single camera (Monocular) is a harder problem when compared to that using a stereo camera due to varies well-known challenges. In the second part of this thesis, We decouple the problem into single view depth estimation (mapping) and two view visual odometry (tracking) and propose a self-supervised framework, namely SelfTAM, which jointly learns the depth estimator and the odometry estimator. The self-supervised problem is usually formulated as an energy minimization problem consist of an energy of data consistency in multi-view (e.g. photometric) and an energy of prior regularization (e.g. depth smoothness prior). We strengthen the supervision signal with a deep feature consistency energy term and a surface normal regularization term. Though our method trains models with stereo sequence such that a real-world scaling factor is naturally incorporated, only monocular data is required in the inference stage. In the last part of this thesis, we revisit the basics of visual odometry and explore the best practice to integrate deep learning models with geometry-based visual odometry methods. A robust visual odometry system, DF-VO, is proposed. We use deep networks to establish 2D-2D/3D-2D correspondences and pick the best correspondences from the dense predictions. Feeding the high-quality correspondences into traditional VO methods, e.g. Epipolar Geometry and Prospective-n-Points, we can solve visual odometry problem within a more robust framework. With the proposed self-supervised training, we can even allow the models to perform online adaptation in the run-time and take a step toward a lifelong learning visual odometry system.
Thesis (Ph.D.) -- University of Adelaide, School of Computer Science, 2020