Yusuke Iida

Tomoki Kubo, Ryuken Uda, Yusuke Iida

Deep double descent is one of the key phenomena underlying the generalization capability of deep learning models. In this study, epoch-wise double descent, which is delayed generalization following overfitting, was empirically investigated by focusing on the evolution of internal structures. Fully connected neural networks of three different sizes were trained on the CIFAR-10 dataset with 30% label noise. By decomposing the loss curves into signal contributions from clean and noisy training data, the epoch-wise evolutions of internal signals were analyzed separately. Three main findings were obtained from this analysis. First, the model achieved strong re-generalization on test data even after perfectly fitting noisy training data during the double descent phase, corresponding to a "benign overfitting" state. Second, noisy data were learned after clean data, and as learning progressed, their corresponding internal activations became increasingly separated in outer layers; this enabled the model to overfit only noisy data. Third, a single, very large activation emerged in the shallow layer across all models; this phenomenon is referred as "outliers," "massive activa-tions," and "super activations" in recent large language models and evolves with re-generalization. The magnitude of large activation correlated with input patterns but not with output patterns. These empirical findings directly link the recent key phenomena of "deep double descent," "benign overfitting," and "large activation", and support the proposal of a novel scenario for understanding deep double descent.

Jargalmaa Batmunkh, Yusuke Iida, Takayoshi Oba et al.

The complex structure and extensive details of solar spectral data, combined with a recent surge in volume, present significant processing challenges. To address this, we propose a deep learning-based compression technique using deep autoencoder (DAE) and 1D-convolutional autoencoder (CAE) models developed with Hinode SOT/SP data. We focused on compressing Stokes I and V polarization spectra from the quiet Sun, as well as from active regions, providing a novel insight into comprehensive spectral analysis by incorporating spectra from extreme magnetic fields. The results indicate that the CAE model outperforms the DAE model in reconstructing Stokes profiles, demonstrating greater robustness and achieving reconstruction errors around the observational noise level. The proposed method has proven effective in compressing Stokes I and V spectra from both the quiet Sun and active regions, highlighting its potential for impactful applications in solar spectral analysis, such as detection of unusual spectral signals.

Jargalmaa Batmunkh, Yusuke Iida, Takayoshi Oba

Detecting unusual signals in observational solar spectra is crucial for understanding the features associated with impactful solar events, such as solar flares. However, existing spectral analysis techniques face challenges, particularly when relying on pre-defined, physics-based calculations to process large volumes of noisy and complex observational data. To address these limitations, we applied deep learning to detect anomalies in the Stokes V spectra from the Hinode/SP instrument. Specifically, we developed an autoencoder model for spectral compression, which serves as an anomaly detection method. Our model effectively identifies anomalous spectra within spectro-polarimetric maps captured prior to the onset of the X1.3 flare on May 5, 2024, in NOAA AR 13663. These atypical spectral points exhibit highly complex profiles and spatially align with polarity inversion lines in magnetogram images, indicating their potential as sites of magnetic energy storage and possible triggers for flares. Notably, the detected anomalies are highly localized, making them particularly challenging to identify in magnetogram images using current manual methods.