Short-term Traffic Prediction with Deep Neural Networks and Adaptive Transfer Learning

Forecasting reservoir water levels is essential in water supply management, impacting both operations and intervention strategies. This paper examines the short-term and long-term forecasting performance of several statistical and machine learning-based methods for predicting the water levels of the Angat Dam in the Philippines. A total of six forecasting methods are compared: naïve/persistence; seasonal mean; autoregressive integrated moving average (ARIMA); gradient boosting machines (GBM); and two deep neural networks (DNN) using a long short-term memory-based (LSTM) encoder-decoder architecture: a univariate model (DNN-U) and a multivariate model (DNN-M). Daily historical water levels from 2001 to 2021 are used in predicting future water levels. In addition, we include meteorological data (rainfall and the Oceanic Niño Index) and irrigation data as exogenous variables. To evaluate the forecast accuracy of our methods, we use a time series cross-validation approach to establish a more robust estimate of the error statistics. Our results show that our DNN-U model has the best accuracy in the 1-day-ahead scenario with a mean absolute error (MAE) and root mean square error (RMSE) of 0.2 m. In the 30-day-, 90-day-, and 180-day-ahead scenarios, the DNN-M shows the best performance with MAE (RMSE) scores of 2.9 (3.3), 5.1 (6.0), and 6.7 (8.1) meters, respectively. Additionally, we demonstrate that further improvements in performance are possible by scanning over all possible combinations of the exogenous variables and only using a subset of them as features. In summary, we provide a comprehensive framework for evaluating water level forecasting by defining a baseline accuracy, analyzing performance across multiple prediction horizons, using time series cross-validation to assess accuracy and uncertainty, and examining the effects of exogenous variables on forecasting performance. In the process, our work addresses several notable gaps in the methodologies of previous works.

Download Full-text

Natural Images Allow Universal Adversarial Attacks on Medical Image Classification Using Deep Neural Networks with Transfer Learning

10.21203/rs.3.rs-757225/v1 ◽

2021 ◽

Author(s):

Akinori Minagi ◽

Hokuto Hirano ◽

Kazuhiro Takemoto

Keyword(s):

Neural Networks ◽

Image Classification ◽

Transfer Learning ◽

Medical Image ◽

Deep Neural Networks ◽

Disease Diagnosis ◽

Natural Images ◽

Fine Tuning ◽

Security And Privacy ◽

Medical Image Classification

Abstract Transfer learning from natural images is well used in deep neural networks (DNNs) for medical image classification to achieve computer-aided clinical diagnosis. Although the adversarial vulnerability of DNNs hinders practical applications owing to the high stakes of diagnosis, adversarial attacks are expected to be limited because training data — which are often required for adversarial attacks — are generally unavailable in terms of security and privacy preservation. Nevertheless, we hypothesized that adversarial attacks are also possible using natural images because pre-trained models do not change significantly after fine-tuning. We focused on three representative DNN-based medical image classification tasks (i.e., skin cancer, referable diabetic retinopathy, and pneumonia classifications) and investigated whether medical DNN models with transfer learning are vulnerable to universal adversarial perturbations (UAPs), generated using natural images. UAPs from natural images are useful for both non-targeted and targeted attacks. The performance of UAPs from natural images was significantly higher than that of random controls, although slightly lower than that of UAPs from training images. Vulnerability to UAPs from natural images was observed between different natural image datasets and between different model architectures. The use of transfer learning causes a security hole, which decreases the reliability and safety of computer-based disease diagnosis. Model training from random initialization (without transfer learning) reduced the performance of UAPs from natural images; however, it did not completely avoid vulnerability to UAPs. The vulnerability of UAPs from natural images will become a remarkable security threat.

Download Full-text

Optimization of Dynamic Neural Network Performance for Short-Term Traffic Prediction

Transportation Research Record Journal of the Transportation Research Board ◽

10.3141/1836-07 ◽

2003 ◽

Vol 1836 (1) ◽

pp. 45-56 ◽

Cited By ~ 66

Author(s):

Sherif Ishak ◽

Prashanth Kotha ◽

Ciprian Alecsandru

Keyword(s):

Neural Networks ◽

Network Performance ◽

Information Dissemination ◽

Target Location ◽

Prediction Performance ◽

Time Factor ◽

Traffic Prediction ◽

Short Term ◽

Traffic Condition ◽

Dynamic Neural Networks

An approach is presented for optimizing short-term traffic-prediction performance by using multiple topologies of dynamic neural networks and various network-related and traffic-related settings. The conducted study emphasized the potential benefit of optimizing the prediction performance by deploying multimodel approaches under parameters and traffic-condition settings. Emphasis was placed on the application of temporal-processing topologies in short-term speed predictions in the range of 5-min to 20-min horizons. Three network topologies were used: Jordan–Elman networks, partially recurrent networks, and time-lagged feedforward networks. The input patterns were constructed from data collected at the target location and at upstream and downstream locations. However, various combinations were also considered. To encourage the networks to associate with historical information on recurrent conditions, a time factor was attached to the input patterns to introduce time-recognition capabilities, in addition to information encoded in the recent past data. The optimal prediction settings (type of topology and input settings) were determined so that performance was maximized under different traffic conditions at the target and adjacent locations. The optimized performance of the dynamic neural networks was compared to that of a statistical nonlinear time series approach, which was outperformed in most cases. The study showed that no single topology consistently outperformed the others for all prediction horizons considered. However, the results showed that the significance of introducing the time factor was more pronounced under longer prediction horizons. A comparative evaluation of performance of optimal and nonoptimal settings showed substantial improvement in most cases. The applied procedure can also be used to identify the prediction reliability of information-dissemination systems.

Download Full-text