Keep Skills in Mind: Understanding and Implementing Skills in Commonsense Question Answering

The cerebellum plays a vital role in motor learning and control with supervised learning capability, while neuromorphic engineering devises diverse approaches to high-performance computation inspired by biological neural systems. This article presents a large-scale cerebellar network model for supervised learning, as well as a cerebellum-inspired neuromorphic architecture to map the cerebellar anatomical structure into the large-scale model. Our multinucleus model and its underpinning architecture contain approximately 3.5 million neurons, upscaling state-of-the-art neuromorphic designs by over 34 times. Besides, the proposed model and architecture incorporate 3411k granule cells, introducing a 284 times increase compared to a previous study including only 12k cells. This large scaling induces more biologically plausible cerebellar divergence/convergence ratios, which results in better mimicking biology. In order to verify the functionality of our proposed model and demonstrate its strong biomimicry, a reconfigurable neuromorphic system is used, on which our developed architecture is realized to replicate cerebellar dynamics during the optokinetic response. In addition, our neuromorphic architecture is used to analyze the dynamical synchronization within the Purkinje cells, revealing the effects of firing rates of mossy fibers on the resonance dynamics of Purkinje cells. Our experiments show that real-time operation can be realized, with a system throughput of up to 4.70 times larger than previous works with high synaptic event rate. These results suggest that the proposed work provides both a theoretical basis and a neuromorphic engineering perspective for brain-inspired computing and the further exploration of cerebellar learning.

Advantages and disadvantages of the use of Digital Technologies (DT) and especially of Computer Algebra Systems (CAS) in mathematics lessons are worldwide discussed controversially. Many empirical studies show the benefit of the use of DT in classrooms and there are also many useful examples concerning their use. However, despite these inspiring results and the countless ideas, classroom suggestions, lesson plans and research reports, the use of DT—and especially CAS—has not succeeded, as many had expected during the last decades see Hoyles & Lagrange, (2010). The thesis of this article is that we have not been able to convince teachers, lecturers at university and parents of the benefit of CAS in the classrooms in a sufficient way. What are the arguments that justify the use of CAS in the classroom? The article gives examples of a fruitful use of CAS with regard to the generally accepted goals or standards of mathematics education—like fostering students’ abilities in problem solving, modelling, proving or communicating—and to the subjects taught in high school. The basis of the argumentation is a competence model which classifies the relation between contents or topics: sequences and limits, functions and equations; representations of DT or CAS: static isolated, static multiple, dynamic isolated and dynamic multiple representations; and classroom activities: calculate, consult, control, communicate and discover.

Topological optimization of the DenseNet with pretrained-weights inheritance and genetic channel selection

Objective. Computed tomography (CT) and magnetic resonance imaging (MRI) are widely used in medical imaging modalities, and provide valuable information for clinical diagnosis and treatment. However, due to hardware limitations and radiation safety concerns, the acquired images are often limited in resolution. Super-resolution reconstruction (SR) techniques have been developed to enhance the resolution of CT and MRI slices, which can potentially improve diagnostic accuracy. To capture more useful feature information and reconstruct higher quality super-resolution images, we proposed a novel hybrid framework SR model based on generative adversarial networks. Approach. The proposed SR model combines frequency domain and perceptual loss functions, which can work in both frequency domain and image domain (spatial domain). The proposed SR model consists of 4 parts: (i) the discrete Fourier transform (DFT) operation transforms the image from the image domain to frequency domain; (ii) a complex residual U-net performs SR in the frequency domain; (iii) the inverse discrete Fourier transform (iDFT) operation based on data fusion transforms the image from the frequency domain to image domain; (iv) an enhanced residual U-net network is used for SR of image domain. Main results. Experimental results on bladder MRI slices, abdomen CT slices, and brain MRI slices show that the proposed SR model outperforms state-of-the-art SR methods in terms of visual quality and objective quality metric such as the structural similarity (SSIM) and the peak signal-to-noise ratio (PSNR), which proves that the proposed model has better generalization and robustness. (Bladder dataset: upscaling factor of 2: SSIM = 0.913, PSNR = 31.203; upscaling factor of 4: SSIM = 0.821, PSNR = 28.604. Abdomen dataset: upscaling factor of 2: SSIM = 0.929, PSNR = 32.594; upscaling factor of 4: SSIM = 0.834, PSNR = 27.050. Brain dataset: SSIM = 0.861, PSNR = 26.945). Significance. Our proposed SR model is capable of SR for CT and MRI slices. The SR results provide a reliable and effective foundation for clinical diagnosis and treatment.

The main purpose of this article is to review the course titled as ”The History of Medical Education in Taiwan”. There were two teachers for collaborative teaching in 2007 academic year. The action plan of the course was to promote students’ information literacy by using the biographies of medical figures as the basic materials for topic inquiry. The Big6 approach was incorporated into the course, directing students in solving their information problems. On the whole, there are some learning outcomes from the course implementation. Students acquired profound understanding of the selected figures they investigated and were markedly inspired by the deeds of these medical figures. Their attitude of active and spontaneous learning was improved. Their problem-solving abilities were enhanced. However, there are still rooms of improvement: the restructure of the action programs, the appointment of teaching assistances selected among distinguished students, the stipulations of autonomous learning agreement, and etc.

This study investigated the residual axial strength of reinforced concrete (RC) columns after exposure to standard and non-standard fires. Experiments and theoretical analyses were performed on 30 RC columns, which were divided into five groups: one group (the control group) was not exposed to fire; two groups were exposed to 30-min and 45-min ISO 834 standard fires; and the other two groups were exposed to 60-min and 75-min non-standard fires. These postfire specimens were axially loaded until they failed. The experimental results showed that 30-min and 45-min ISO 834 fires decreased the residual strength by 11.6% and 17.4%, respectively. 60-min and 75-min non-ISO 834 fires caused similar reductions in the residual axial strength of postfire RC columns, although the durations of these non-standard fires were considerably longer. This experimental result confirmed that the maximum temperature, rather than the fire durations, of non-standard fires governed the reduction in the residual axial strength. In theoretical analysis, the axial strength of postfire RC columns can be calculated using the proposed model, which combines the confinement model, the 500°C isotherm method, and the distribution model of temperature. The proposed model can appropriately predict the residual axial strength of postfire columns when the fire closely follows the ISO 834 fire. The inaccuracy of the model significantly increases when the fires are non-ISO 834 fires. Two concepts, namely equivalent area and maximum temperature, were introduced to address the issue. The concept of maximum temperature successfully addressed the above-mentioned issue and exhibited superiority over the concept of equivalent area. This result reaffirmed the superior governing characteristic of the maximum temperature compared with the duration of non-standard fires, as evidenced in the experimental results. The proposed models would be useful for engineers in practice when evaluating the residual axial strength of both standard and non-standard postfire RC columns.

The traditional laboratory models for the hydroelasticity and seakeeping performance of ships are tested in calm water and in uni-directional, artificially generated waves. A new alternative to the tank model measurement methodology is to conduct experiments using large-scale models in actual sea conditions. To implement the tests, a large-scale segmented self-propelling model and testing system were designed and assembled. A buoy wave meter was adopted to record the coastal waves that the model encountered during the tests. The analysis of the results of waves in sheltered waters by the spectral method shows good agreement with ISSC spectra. To investigate the difference between this new methodology and the traditional towing tank tests, a small-scale model, whose type and configuration are the same as those of the large-scale model ship, was used and tests were conducted in a towing tank. Comparison of the two experimental results shows that there is a remarkable difference in the response characteristics between the large-scale model at sea and the small-scale model in the tank. Numerical simulations of the responses of the ship under equivalent sea states were also carried out. The influence of directional spreading functions on the results was analyzed by a numerical approach. The classical model tests under long-crested waves in the towing tank over-estimate the motion and wave load responses; however, large-scale model tests carried out at sea are more reasonable for ship design and scientific research.

Speaker identification is the process of automatically determining who is speaking from the known speakers by the model. It is crucial in voice-based authentication, forensic investigations, security and surveillance. In recent studies, the combination of convolutional neural network (CNN) and recurrent neural network (RNN) variants performed better than separate models of both. However, only limited studies are conducted in speaker identification using a combination of CNN and RNN variants. In this study, we proposed speaker identification using hybrid two-dimensional CNN (2DCNN) and bidirectional gated recurrent unit (BiGRU) to improve performance. The proposed model integrates the advantage of 2DCNN and BiGRU layers to improve the performance of the model. 2DCNN layers have the advantage of extracting short-term spatial features from input data and it has a limited number of parameters for computation. BiGRU layers have an advantage in extracting long-term temporal dependency between the features in both directions (i.e. backward and forward) and it is efficient in achieving convergence during training. Spectrograms of the speech were used as input in our proposed model because of the rich acoustic features of the speaker. To compare the performance of the proposed model, additional experiments are conducted using the models 2DCNN, CNN-LSTM, CNN-BiLSTM and CNN-GRU. The experiments were conducted on the VoxCeleb1 audio dataset, which consists of 153,516 utterances collected from the 1251 speakers. The accuracy, precision, recall and f1 score of the proposed model are 98.28%, 99.08%, 98.92% and 98.97% respectively. The proposed model was compared with the existing works to show the effectiveness of the proposed model. The experiment results and the comparison with the existing works show that the proposed model has higher performance than both existing works and other models experimented in this study.

One of the goals of machine learning is learning a sequence of tasks more naturally. Continual learning imitates the real learning mode of human beings and continually learns new knowledge without forgetting old knowledge at the same time. In the past decades, considerable attention has been paid to this learning method. However, avoiding forgetting old knowledge in the process of learning new knowledge remains an ongoing challenge due to catastrophic forgetting. Therefore, it is desirable to exploit a new method to improve the stability in a continual learning scenario. In this paper, we specifically focus on the shared feature extraction between two consecutive tasks. To this end, we explore a continual learning paradigm by using the task-wise shared hidden representation alignment module, which contrasts shared representations from the current task and shared representations from reconstruction pseudo samples of previous tasks. Our proposed TSHRA model grasps similarity features provided by the alignment module to learn shared representations more consummately when the model is learning the current task. To verify our proposed model, we conduct experiments on Split-MNIST and Fashion-MNIST. The experimental results show that our proposed TSHRA’s performance is outstanding which justify that the alignment module has a positive effect on learning shared representations among different tasks for the continual learning scenario.KeywordsContinual learningThe task-wise shared hidden representationCatastrophic forgettingAlignment

During the process of aerodynamic shape design of supersonic and hypersonic space planes, laminar flow design and boundary layer transition prediction play important roles in aero-thermal numerical simulations and aero-thermal protection design. Therefore, in this study, a computational fluid dynamics compatible transition closure model for high speed laminar-to-turbulent transitional flows is formulated with consideration of the analysis results from stability theory. The proposed model contains two transport equations to describe the transition mechanism using local variables. Specifically, the eddy viscosity of laminar fluctuations and intermittency factor are chosen to be the characteristic parameters and modeled by transport equations. Accounting for the dominant instability modes at supersonic/hypersonic conditions, the first- and second- modes are modeled using local variables through the analysis of laminar self-similar boundary layers. Then, the present transition model is applied with compressibility corrected k $k$ - ω $\omega$ shear stress transport turbulence model. Thus, as the main significance of the current work, the present model is enabled to capture the overshoot phenomena as well as predict the transition onset position. Finally, comparisons between the predictions using the present model and the wind tunnel experimental results of several well-documented flow cases are provided to validate the proposed transition turbulence model.

Abstract. Understanding and predicting the ocean environment and marine ecosystem status depends on accurate representations of regional ocean dynamics. Recently, the Modular Ocean Model version 6 (MOM6) has been configured to span the Northeast Pacific Ocean from Baja California to the Chukchi Sea (MOM6-NEP). In this study we present a physical hindcast (1993–2018) simulation of MOM6-NEP where it is coupled to a thermodynamic-dynamic sea-ice module and includes tides. We evaluate the performance of this model in the Bering Sea. Various model metrics are benchmarked against in-situ mooring data and satellite observations. The simulation captures the general characteristics of Bering Sea dynamics, particularly with respect to seasonal and interannual variability of the middle shelf water mass properties. Modeling of shear induced mixing was found to be critical to the model's ability to reproduce the observed sharp summer thermocline and its depth. The hindcast simulation reproduces the long-term mean timing of sea-ice arrival and retreat in both the northern and southern Bering Sea, with the remaining mild biases primarily occurring in May over the northern shelf – the model captures the mean timing of sea-ice retreat, though it tends to retreat earlier in colder years and later in warmer years compared to observations. This pattern in biases suggests that the melting rate in the model likely underestimates the well-known melt-rate dependency on ice property whereby thicker (thinner) ice melts more slowly (quickly). As a result of high skills in reproducing sea-ice areal coverage, the interannual variability of the cold pool (the cold-water mass present on the bottom of the Bering Sea shelf in summer) extent is accurately reproduced by the model. Skillful representation of sea ice and the cold pool is essential for understanding ecosystem dynamics and successful fisheries management in the Bering Sea. The findings of this study contribute to the development of reliable oceanographic modeling and forecasting of marine ecosystem conditions to support fisheries management decision making.

In this paper, combining multiple classifiers based on the Mixture of Experts investigated, as Multi Layer Perceptrons, MLPs, is used as gating and expert networks. We call this method Mixture of Multilayer Perceptron Experts, MOME. Large experiments established upon combining classifiers on Persian handwritten words are reported and discussed. Here, to evaluate our proposed model we use a real-world database: Iranshahr. The experimental results using our database support claim that implementing a mixture of some simple MLPs improves the performance, in which the best result of our proposed model is 90.50%, demonstrating 60% decline of error rate with regards to single MLPs. Furthermore, comparison test with other combination methods indicates that the proposed model yields excellent recognition rate in handwritten word recognition.

Self-adaptive robust nonlinear regression for unknown noise via mixture of Gaussians

There1 is often a difference between the training library and the test library in the inclusion of noise, which affects the practical application of speech endpoint detection. We propose to use a generalized regularized online sequential extreme learning machine with forgetting factor (GR-OSELM-FF) for voice activation detection to adapt to the difference between test and training samples. Practical usability is our driving motivation; the proposed model should be easily adapted to new conditions. When a new voice stream arrives in the test or the actual application phase, the proposed model can directly adjust the output weight. To overcome the weakness of ELM's vulnerability to random hidden layer parameters, we use an extreme learning machine-based autoencoder (ELM-AE) to initialize the model parameters instead of using random initialization. The experimental results show that the pretrained models achieve better performance with ELM-AE, which can obtain the potential information from the data. The experimental results also show that the proposed algorithm maintains good accuracy and omission rates in different SNR noise environments and real-world voice samples.

Hybrid Attention for Chinese Character-Level Neural Machine Translation