No abstract

The El Niño-Southern Oscillation (ENSO) is a quasi-periodic climate type that occurs near the equatorial Pacific Ocean. Extreme periods of this climate type can cause terrible weather and climate anomalies on a global scale. Therefore, it is critical to accurately, quickly, and effectively predict the occurrence of ENSO events. Most existing research methods rely on the powerful data-fitting capability of deep learning which does not fully consider the spatio-temporal evolution of ENSO and its quasi-periodic character, resulting in neural networks with complex structures but a poor prediction. Moreover, due to the large magnitude of ocean climate variability over long intervals, they also ignored nearby prediction results when predicting the Niño 3.4 index for the next month, which led to large errors. To solve these problem, we propose a spatio-temporal transformer network to model the inherent characteristics of the sea surface temperature anomaly map and heat content anomaly map along with the changes in space and time by designing an effective attention mechanism, and innovatively incorporate temporal index into the feature learning procedure to model the influence of seasonal variation on the prediction of the ENSO phenomenon. More importantly, to better conduct long-term prediction, we propose an effective recurrent prediction strategy using previous prediction as prior knowledge to enhance the reliability of long-term prediction. Extensive experimental results show that our model can provide an 18-month valid ENSO prediction, which validates the effectiveness of our method.

IntroductionExisting behavioral and neuroimaging studies revealed inter-individual variability in the selection of the two phonological routes in word reading. However, it is not clear how individuals’ preferred reading pathways/strategies modulate the involvement of a certain brain region for phonological learning in a new language, and consequently affect their behavioral performance on phonological access.MethodsTo address this question, the present study recruited a group of native Chinese speakers to learn two sets of artificial language characters, respectively, in addressed-phonology training (i.e., whole-word mapping) and assembled-phonology training conditions (i.e., grapheme-to-phoneme mapping).ResultsBehavioral results showed that the more lexical pathways participants preferred, the better they performed on newly-acquired addressed characters relative to assembled characters. More importantly, neuroimaging results showed that participants who preferred lexical pathway in phonological access show less involvement of brain regions for addressed phonology (e.g., the bilateral orbitofrontal cortex and right pars triangularis) in the processing of newly-acquired addressed characters.ConclusionThese results indicated that phonological access via the preferred pathway required less neural resources to achieve better behavioral performance. These above results provide direct neuroimaging evidence for the influence of reading pathway preference on phonological learning.