Training Pipeline Research Articles

Determining the Required Training Capacity Within a Military Establishment

We address the problem of deciding how many positions to set aside for military recruits undergoing training. Within a cap on the total number of military members, we vary the ratio between positions allocated to the training pipeline versus those required in the trained effective establishment. This is done with the goal of determining the extent to which given ratios are sustainable. We use a Markovian model of the training pipeline, with parameters derived from historical personnel data. Through Monte Carlo simulation, we predict how often a given ratio allows the required trained force to be fully generated, as well as the surplus of trained personnel, it is expected to generate. We extend our previous work in this area by considering an alternative Human Resources policy that uncaps the training pipeline. Our modelling results have informed ongoing initiatives to optimize the force mix and structure of the Canadian Armed Forces.

Digital Forensics Investigation for Attacks on Artificial Intelligence

The new research approaches are needed to be adopted to deal with security threats in Artificial Intelligence (AI)-based systems. This research is aimed at investigating the AI attacks that are “malicious by design.” It also deals with conceptualization of the problem and strategies for attacks on AI using digital forensic tools. A specific class of problems in adversarial attacks are tampering of images for computational processing in applications of digital photography, computer vision, pattern recognition (facial capping algorithms). State-of-the-art developments in forensics, such as 1. Application of end-to-end Neural Network training pipeline for image rendering and provenance analysis. 2. Deep fake image analysis using frequency methods, wavelet analysis, and tools like Amped Authenticate. 3. Capsule networks for detecting forged images. 4. Information transformation for feature extraction via image forensic tools, such as EXIF-SC, Splice Radar, and Noiseprint. 5. Application of generative adversarial networks (GAN) based models as anti-image forensics [8], will be studied in great detail and a new research approach will be designed incorporating these advancements for utility of digital forensics.

Using modeling and scenario analysis to support evidence-based health workforce strategic planning in Malawi

BackgroundA well-trained and equitably distributed workforce is critical to a functioning health system. As workforce interventions are costly and time-intensive, investing appropriately in strengthening the health workforce requires an evidence-based approach to target efforts to increase the number of health workers, deploy health workers where they are most needed, and optimize the use of existing health workers. This paper describes the Malawi Ministry of Health (MoH) and collaborators’ data-driven approach to designing strategies in the Human Resources for Health Strategic Plan (HRH SP) 2018–2022.MethodsThree modelling exercises were completed using available data in Malawi. Staff data from districts, central hospitals, and headquarters, and enrollment data from all health training institutions were collected between October 2017 and February 2018. A vacancy analysis was conducted to compare current staffing levels against established posts (the targeted number of positions to be filled, by cadre and work location). A training pipeline model was developed to project the future available workforce, and a demand-based Workforce Optimization Model was used to estimate optimal staffing to meet current levels of service utilization.ResultsAs of 2017, 55% of established posts were filled, with an average of 1.49 health professional staff per 1000 population, and with substantial variation in the number of staff per population by district. With current levels of health worker training, Malawi is projected to meet its establishment targets in 2030 but will not meet the WHO standard of 4.45 health workers per 1000 population by 2040. A combined intervention reducing attrition, increasing absorption, and doubling training enrollments would allow the establishment to be met by 2023 and the WHO target to be met by 2036. The Workforce Optimization Model shows a gap of 7374 health workers to optimally deliver services at current utilization rates, with the largest gaps among nursing and midwifery officers and pharmacists.ConclusionsGiven the time and significant financial investment required to train and deploy health workers, evidence needs to be carefully considered in designing a national HRH SP. The results of these analyses directly informed Malawi’s HRH SP 2018–2022 and have subsequently been used in numerous planning processes and investment cases in Malawi. This paper provides a practical methodology for evidence-based HRH strategic planning and highlights the importance of strengthening HRH data systems for improved workforce decision-making.

Responding to COVID-19: What we learned in 2020

Introduction: This article summarises a report published in July 2021 by Medical Deans Australia and New Zealand (Medical Deans), the peak body representing the 23 medical programs in Australia and New Zealand. It explores how medical schools responded to the early impacts of COVID-19 in 2020, and how they might build on some of the changes to achieve fundamental improvements in medical education in Australia and New Zealand. Methods: The Medical Deans report was based on: Responses to a survey midway through 2020 of its member schools about their experiences during the initial lockdowns; the contributions of presenters at the Medical Deans 2020 Annual Conference; subsequent discussions among communities of practice. Results: Innovations introduced in response to the pandemic, included greater equity of learning opportunities across geographical locations; health services taking more responsibility for clinical placements; greater emphasis on competencies and less on clinical rotations in specific disciplines; strong collaboration between medical schools, and with providers in the medical training and research pipeline. Challenges include balancing the benefits of online learning with the need for human connection and a chronic lack of clinical training opportunities in community-based care. Conclusion: While the impact of the pandemic on medical education and training was costly for all involved – both financially and personally – the scale of disruption provided a unique opportunity for step change. To fully realise this potential moving forward, medical schools will need to work in partnership with all those involved: students, health services, prevocational training, specialist colleges, regulators and governments.

Sorting biotic and abiotic stresses on wild rocket by leaf-image hyperspectral data mining with an artificial intelligence model

BackgroundWild rocket (Diplotaxis tenuifolia) is prone to soil-borne stresses under intensive cultivation systems devoted to ready-to-eat salad chain, increasing needs for external inputs. Early detection of the abiotic and biotic stresses by using digital reflectance-based probes may allow optimization and enhance performances of the mitigation strategies.MethodsHyperspectral image analysis was applied to D. tenuifolia potted plants subjected, in a greenhouse experiment, to five treatments for one week: a control treatment watered to 100% water holding capacity, two biotic stresses: Fusarium wilting and Rhizoctonia rotting, and two abiotic stresses: water deficit and salinity. Leaf hyperspectral fingerprints were submitted to an artificial intelligence pipeline for training and validating image-based classification models able to work in the stress range. Spectral investigation was corroborated by pertaining physiological parameters.ResultsWater status was mainly affected by water deficit treatment, followed by fungal diseases, while salinity did not change water relations of wild rocket plants compared to control treatment. Biotic stresses triggered discoloration in plants just in a week after application of the treatments, as evidenced by the colour space coordinates and pigment contents values. Some vegetation indices, calculated on the bases of the reflectance data, targeted on plant vitality and chlorophyll content, healthiness, and carotenoid content, agreed with the patterns of variations observed for the physiological parameters. Artificial neural network helped selection of VIS (492–504, 540–568 and 712–720 nm) and NIR (855, 900–908 and 970 nm) bands, whose read reflectance contributed to discriminate stresses by imaging.ConclusionsThis study provided significative spectral information linked to the assessed stresses, allowing the identification of narrowed spectral regions and single wavelengths due to changes in photosynthetically active pigments and in water status revealing the etiological cause.

Multi-Task Semi-Supervised Adversarial Autoencoding for Speech Emotion Recognition

Inspite the emerging importance of Speech Emotion Recognition (SER), the state-of-the-art accuracy is quite low and needs improvement to make commercial applications of SER viable. A key underlying reason for the low accuracy is the scarcity of emotion datasets, which is a challenge for developing any robust machine learning model in general. In this article, we propose a solution to this problem: a multi-task learning framework that uses auxiliary tasks for which data is abundantly available. We show that utilisation of this additional data can improve the primary task of SER for which only limited labelled data is available. In particular, we use gender identifications and speaker recognition as auxiliary tasks, which allow the use of very large datasets, e. g., speaker classification datasets. To maximise the benefit of multi-task learning, we further use an adversarial autoencoder (AAE) within our framework, which has a strong capability to learn powerful and discriminative features. Furthermore, the unsupervised AAE in combination with the supervised classification networks enables semi-supervised learning which incorporates a discriminative component in the AAE unsupervised training pipeline. This semi-supervised learning essentially helps to improve generalisation of our framework and thus leads to improvements in SER performance. The proposed model is rigorously evaluated for categorical and dimensional emotion, and cross-corpus scenarios. Experimental results demonstrate that the proposed model achieves state-of-the-art performance on two publicly available datasets.

Keeping an Eye on Things: Deep Learned Features for Long-Term Visual Localization

In this letter, we learn visual features that we use to first build a map and then localize a robot driving autonomously across a full day of lighting change, including in the dark. We train a neural network to predict sparse keypoints with associated descriptors and scores that can be used together with a classical pose estimator for localization. Our training pipeline includes a differentiable pose estimator such that training can be supervised with ground truth poses from data collected earlier, in our case from 2016 and 2017 gathered with multi-experience Visual Teach and Repeat (VT&R). We insert the learned features into the existing VT&R pipeline to perform closed-loop path following in unstructured outdoor environments. We show successful path following across all lighting conditions despite the robot’s map being constructed using daylight conditions. Moreover, we explore generalizability of the features by driving the robot across all lighting conditions in new areas not present in the feature training dataset. In all, we validated our approach with 35.5 km of autonomous path following experiments in challenging conditions.

Improving Classification Performance of Fully Connected Layers by Fuzzy Clustering in Transformed Feature Space

Fully connected (FC) layers are used in almost all neural network architectures ranging from multilayer perceptrons to deep neural networks. FC layers allow any kind of symmetric/asymmetric interaction between features without making any assumption about the structure of the data. However, success of convolutional and recursive layers and findings of many studies have proven that the intrinsic structure of a dataset holds a great potential to improve the success of a classification problem. Leveraging clustering to explore and exploit this intrinsic structure in classification problems has been the subject of various studies. In this paper, we propose a new training pipeline for fully connected layers which enables them to make more accurate classification predictions. The proposed method aims to reflect the clustering patterns in the original feature space of the training dataset to the transformed feature space created by the FC layer. In this way, we intend to enhance the representation ability of the extracted features and accordingly increase the classification accuracy. The Fuzzy C-Means algorithm is employed in this study as the clustering tool. To evaluate the performance of the proposed method, 11 experiments were conducted on 9 benchmark UCI datasets. Empirical results show that the proposed method works well in practice and gives higher classification accuracies compared to a regular FC layer in most datasets.

Aligning Air Force Ground Surgical Team (GST) Training With the Deployment Experience.

The Air Force Ground Surgical Team (GST) Phase 1 course is a two-week pre-deployment training for the Air Force's conventional austere surgical care platform. Since the creation of the GST platform and associated training pipeline in 2017, course adjustments and improvements have relied on expert opinion and anecdote. To gain a more robust understanding of GST deployment clinical and operational activities, we conducted a survey of all surgeons who completed GST Phase 1 training from its inception in April 2017 to September 2020. 87 surgeons took the course from April 2017 to September 2020. 60 of those surgeons were still on active duty and were candidates to complete the survey sent from the Air Force Survey Office. 31 individuals responded and their identification was kept blinded. An IRB exemption was issued before study initiation. Of the 31 respondents, 9 took the GST Phase 1 course but never deployed, and were excluded. The remaining 22 surgeons deployed at some point from 2017 to 2020. Four surgeons reported providing no surgical care during their deployment. 68% of the surgeons deployed to an actual standalone GST platform, while the rest were retained at larger military treatment facilities (MTFs) or Role III facilities. The median number of surgeries performed was 2 for surgeons at standalone GSTs and 7 for those at larger MTFs/Role IIIs. A holding time of greater than 12 hours was reported for 15% of operative patients and 58% of nonoperative patients at standalone GSTs. 28% (n = 5) of surgeons reported taking care of patients in the Golden Hour of surgery, and 23% of teams reported a patient death. Two surgeons cared for a military working dog, and four surgeons cared for pediatric patients. 50% of surgeons had more than one patient present simultaneously for care. 50% of surgeons' resupply were in the greater than 30 days or never received category. The GST Phase 1 course has a unique role in preparing students to provide austere surgical care. This includes both preparing to function in the operational military environment as well as applying sound in-garrison trauma surgical care techniques to the austere, resource variable environment. The results of this survey suggest that a broadening of content-specific deployment-related topics, the formalization of documentation education, incorporating formal evidence-based nontechnical skills training, and identifying optimal GST context-specific behaviors will strengthen the effectiveness of the course in preparing students for deployment.

Geographic practice preferences of graduating medical students pursuing careers in dermatology

Geographic practice preferences of graduating medical students pursuing careers in dermatology

DilUnet: A U-net based architecture for blood vessels segmentation

Background and objectiveRetinal image segmentation can help clinicians detect pathological disorders by studying changes in retinal blood vessels. This early detection can help prevent blindness and many other vision impairments. So far, several supervised and unsupervised methods have been proposed for the task of automatic blood vessel segmentation. However, the sensitivity and the robustness of these methods can be improved by correctly classifying more vessel pixels. MethodWe proposed an automatic, retinal blood vessel segmentation method based on the U-net architecture. This end-to-end framework utilizes preprocessing and a data augmentation pipeline for training. The architecture utilizes multiscale input and multioutput modules with improved skip connections and the correct use of dilated convolutions for effective feature extraction. In multiscale input, the input image is scaled down and concatenated with the output of convolutional blocks at different points in the encoder path to ensure the feature transfer of the original image. The multioutput module obtains upsampled outputs from each decoder block that are combined to obtain the final output. Skip paths connect each encoder block with the corresponding decoder block, and the whole architecture utilizes different dilation rates to improve the overall feature extraction. ResultsThe proposed method achieved an accuracy: of 0.9680, 0.9694, and 0.9701; sensitivity of 0.8837, 0.8263, and 0.8713; and Intersection Over Union (IOU) of 0.8698, 0.7951, and 0.8184 with three publicly available datasets: DRIVE, STARE, and CHASE, respectively. An ablation study is performed to show the contribution of each proposed module and technique. ConclusionThe evaluation metrics revealed that the performance of the proposed method is higher than that of the original U-net and other U-net-based architectures, as well as many other state-of-the-art segmentation techniques, and that the proposed method is robust to noise.

VPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training

The increasing computational complexity of DNNs achieved unprecedented successes in various areas such as machine vision and natural language processing (NLP), e.g., the recent advanced Transformer has billions of parameters. However, as large-scale DNNs significantly exceed GPU's physical memory limit, they cannot be trained by conventional methods such as data parallelism. Pipeline parallelism that partitions a large DNN into small subnets and trains them on different GPUs is a plausible solution. Unfortunately, the layer partitioning and memory management in existing pipeline parallel systems are fixed during training, making them easily impeded by out-of-memory errors and the GPU under-utilization. These drawbacks amplify when performing neural architecture search (NAS) such as the evolved Transformer, where different network architectures of Transformer needed to be trained repeatedly. vPipe is the first system that transparently provides dynamic layer partitioning and memory management for pipeline parallelism. vPipe has two unique contributions, including (1) an online algorithm for searching a near-optimal layer partitioning and memory management plan, and (2) a live layer migration protocol for re-balancing the layer distribution across a training pipeline. vPipe improved the training throughput of two notable baselines (Pipedream and GPipe) by 61.4-463.4 percent and 24.8-291.3 percent on various large DNNs and training settings.

Fostering a diverse regional community of physician-scientist trainees

Among the many academic challenges faced by dual-degree MD-PhD students is access to professional support networks designed to overcome the unique academic and personal barriers to physician-scientist training. In the current study, we hypothesized that regional access to a student MD-PhD conference, termed the Southeastern Medical Scientist Symposium (SEMSS), would enhance medical and/or graduate training by fostering such relationships between physician-scientist trainees, doing so by discussing both the challenges of physician-scientist training and effective strategies to overcome them. In the current study, we used a mixed-methods approach to evaluate the overall usefulness of SEMSS over a ten-year period (2010-2020) to identify key areas of particular benefit to trainees. The authors used conference registration data to compile self-reported demographic and regional attendance, followed by a post-conference survey to gauge attendee satisfaction. Over the reporting period, SEMSS was attended by equivalent proportions of MD-PhD and undergraduate students, among which were a high-percentage of students from underrepresented minority (URM) groups relative to the national MD-PhD applicant pool; nearly one-third of URM students attendees later matriculated into MD-PhD programs, far exceeding the national MD-PhD matriculation rate. Among the benefits reported by students were “opportunities to network with peers” and opportunities to learn about the physician-scientist career track. Therefore, we therefore propose regional MD-PhD conferences as an effective model to promote diversity within the physician-scientist training pipeline.

PrimeNet: Adaptive Multi-Layer Deep Neural Structure for Enhanced Feature Selection in Early Convolution Stage

The colossal depths of the deep neural network sometimes suffer from ineffective backpropagation of the gradients through all its depths, whereas the strong performance of shallower multilayer neural structures proves their ability to increase the gradient signals in the early stages of training, which easily gets backpropagated for global loss corrections. Shallow neural structures are always a good starting point for encouraging the sturdy feature characteristics of the input. In this research, a shallow, deep neural structure called PrimeNet is proposed. PrimeNet is aimed to dynamically identify and encourage the quality visual indicators from the input to be used by the subsequent deep network layers and increase the gradient signals in the lower stages of the training pipeline. In addition to this, the layer-wise training is performed with the help of locally generated errors, which means the gradient is not backpropagated to previous layers, and the hidden layer weights are updated during the forward pass, making this structure a backpropagation free variant. PrimeNet has obtained state-of-the-art results on various image datasets, attaining the dual objective of: (1) a compact dynamic deep neural structure, which (2) eliminates the problem of backward-locking. The PrimeNet unit is proposed as an alternative to traditional convolution and dense blocks for faster and memory-efficient training, outperforming previously reported results aimed at adaptive methods for parallel and multilayer deep neural systems.

Examination of Intersectionality and the Pipeline for Black Academic Surgeons

The lack of underrepresented in medicine physicians within US academic surgery continues, with Black surgeons representing a disproportionately low number. To evaluate the trend of general surgery residency application, matriculation, and graduation rates for Black trainees compared with their racial and ethnic counterparts over time. In this nationwide multicenter study, data from the Electronic Residency Application Service (ERAS) for the general surgery residency match and Graduate Medical Education (GME) surveys of graduating general surgery residents were retrospectively reviewed and stratified by race, ethnicity, and sex. Analyses consisted of descriptive statistics, time series plots, and simple linear regression for the rate of change over time. Medical students and general surgery residency trainees of Asian, Black, Hispanic or Latino of Spanish origin, White, and other races were included. Data for non-US citizens or nonpermanent residents were excluded. Data were collected from 2005 to 2018, and data were analyzed in March 2021. Primary outcomes included the rates of application, matriculation, and graduation from general surgery residency programs. Over the study period, there were 71 687 applicants, 26 237 first-year matriculants, and 24 893 graduates. Of 71 687 applicants, 24 618 (34.3%) were women, 16 602 (23.2%) were Asian, 5968 (8.3%) were Black, 2455 (3.4%) were Latino, and 31 197 (43.5%) were White. Women applicants and graduates increased from 29.4% (1178 of 4003) to 37.1% (2293 of 6181) and 23.5% (463 of 1967) to 33.5% (719 of 2147), respectively. When stratified by race and ethnicity, applications from Black women increased from 2.2% (87 of 4003) to 3.5% (215 of 6181) (P < .001) while applications from Black men remained unchanged (3.7% [150 of 4003] to 4.6% [284 of 6181]). While the matriculation rate for Black women remained unchanged (2.4% [46 of 1919] to 2.3% [52 of 2264]), the matriculation rate for Black men significantly decreased (3.0% [57 of 1919] to 2.4% [54 of 2264]; P = .04). Among Black graduates, there was a significant decline in graduation for men (4.3% [85 of 1967] to 2.7% [57 of 2147]; P = .03) with the rate among women remaining unchanged (1.7% [33 of 1967] to 2.2% [47 of 2147]). Findings of this study show that the underrepresentation of Black physicians at every stage in surgical training pipeline persists. Black men are especially affected. Identifying factors that address intersectionality and contribute to the successful recruitment and retention of Black trainees in general surgery residency is critical for achieving racial and ethnic as well as gender equity.

Gender, psychiatric disability, and dropout from peer support specialist training.

Although research indicates that the prevalence of psychiatric disability differs depending on gender, a paucity of information exists as to whether men and women with psychiatric disability also differ regarding service program outcomes. For a United States Southwest peer support specialist training program, this study examines whether gender moderates the association between psychiatric disability and a key outcome-training dropout. Data were collected for 78 men and 157 women with psychiatric disability and 137 men and 203 women with mental illness only. Logistic regression was used to examine the association between psychiatric disability and dropout, with gender as a moderator variable, and age, education, race/ethnicity, and substance use disorder as control variables. Of trainees with psychiatric disability, dropout was greater among men than women (34.6% and 20.4%, respectively; p < .05). Dropout was also greater among men with psychiatric disability than among men with mental illness only (34.6% and 15.3%; p < .01). In contrast, dropout was similar for women with psychiatric disability and mental illness only (20.4% and 18.7%; p > .05), and dropout was comparable among men and women with mental illness only (15.3% and 18.7%; p > .05). In summary, risk of dropout was substantially higher among men with psychiatric disability than women with psychiatric disability. Gender tailoring of the program's services should be considered to better support training completion. This study's findings also raise questions as to possible underrepresentation of men with psychiatric disability in the peer support workforce training pipeline. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Fast Panoptic Segmentation with Soft Attention Embeddings.

Panoptic segmentation provides a rich 2D environment representation by unifying semantic and instance segmentation. Most current state-of-the-art panoptic segmentation methods are built upon two-stage detectors and are not suitable for real-time applications, such as automated driving, due to their high computational complexity. In this work, we introduce a novel, fast and accurate single-stage panoptic segmentation network that employs a shared feature extraction backbone and three network heads for object detection, semantic segmentation, instance-level attention masks. Guided by object detections, our new panoptic segmentation head learns instance specific soft attention masks based on spatial embeddings. The semantic masks for stuff classes and soft instance masks for things classes are pixel-wise coherent and can be easily integrated in a panoptic output. The training and inference pipelines are simplified and no post-processing of the panoptic output is necessary. Benefiting from fast inference speed, the network can be deployed in automated vehicles or robotic applications. We perform extensive experiments on COCO and Cityscapes datasets and obtain competitive results in both accuracy and time. On the Cityscapes dataset we achieve 59.7 panoptic quality with an inference speed of more than 10 FPS on high resolution 1024 × 2048 images.

Emerging Technology: Preparing Tomorrow's MCH Workforce to Innovate for Equity

PurposeThis commentary proposes a new direction to train the MCH workforce by leveraging today’s rapidly changing innovation and technology to address persistent health inequities.DescriptionWe outline the creation of an MCH technology and innovation training pipeline developed by harnessing creative funding opportunities, diversifying training modalities, and expanding partnerships beyond traditional academic-practice partners, that be replicated and adapted by other academic programs.AssessmentTechnology and innovation will continue to be a growing intersection between health and equity, and we must create a robust pipeline of MCH leaders prepared to collaborate with entrepreneurial and innovation leaders.ConclusionTechnology offers an important opportunity to improve MCH outcomes and reduce disparities, but only if we train the MCH workforce to seize these opportunities.

Taris: An online speech recognition framework with sequence to sequence neural networks for both audio-only and audio-visual speech

It is widely accepted that the visual modality of speech provides complementary information to the speech recognition task, and many models have been introduced in order to make good use of the visual channel. This article develops Taris, a fully differentiable neural network model capable of decoding both audio-only and audio-visual speech in real time. We achieve this by connecting our previously proposed models AV Align and Taris, which are both end-to-end differentiable approaches to audio-visual speech integration and online speech recognition respectively. We evaluate AV Taris under the same conditions as AV Align and Taris on one of the largest publicly available audio-visual speech datasets, LRS2. Our results show that AV Taris is superior to the audio-only variant of Taris, demonstrating the utility of the visual modality to speech recognition within the real time decoding framework defined by Taris. Compared to an equivalent Transformer-based AV Align model that takes advantage of full sentences without meeting the real-time requirement, we report an absolute degradation of approximately 3% with AV Taris. As opposed to the more popular alternative for online speech recognition, namely the RNN Transducer, Taris offers a greatly simplified fully differentiable training pipeline. We speculate that AV Taris has the potential to popularise the adoption of Audio-Visual Speech Recognition (AVSR) technology and overcome the inherent limitations of the audio modality in less optimal listening conditions.11Our code is publicly available at https://github.com/georgesterpu/Taris.

Neural Acoustic-Phonetic Approach for Speaker Verification With Phonetic Attention Mask

Traditional acoustic-phonetic approach makes use of both spectral and phonetic information when comparing the voice of speakers. While phonetic units are not equally informative, the phonetic context of speech plays an important role in speaker verification (SV). In this paper, we propose a neural acoustic-phonetic approach that learns to dynamically assign differentiated weights to spectral features for SV. Such differentiated weights form a phonetic attention mask (PAM). The neural acoustic-phonetic framework consists of two training pipelines, one for SV and another for speech recognition. Through the PAM, we leverage the phonetic information for SV. We evaluate the proposed neural acoustic-phonetic framework on the RSR2015 database Part III corpus, that consists of random digit strings. We show that the proposed framework with PAM consistently outperforms baseline with an equal error rate reduction of 13.45% and 10.20% for female and male data, respectively.