Jiha Kim 김지하

3D-CNN action recognizers learn appearance and motion in an entangled way, so they lean on static visual cues instead of true motion — which hurts generalization and interpretability on resource-constrained AIoT edge devices. ASTRA is a dual-pathway architecture that separates semantic appearance from pure motion through three components: an STFormer module that models space and time adaptively through cross-attention and estimates frame-wise importance; Importance-Guided Temporal Perturbation, which isolates appearance-free motion by comparing original and perturbed inputs; and a Motion Gated Unit that refines the motion features through dual-gate modulation. Under a leak-free, video-level protocol ASTRA reaches 97.88% on UCF-101 and 81.05% on HMDB-51 with an MViT-v2-S backbone. Against its own R(2+1)D-18 backbone, the motion pathway adds a significant +2.81%p on HMDB-51 (McNemar p < 0.001) while staying equivalent on appearance-saturated UCF-101, at 32.9M parameters, 41.9 GFLOPs and 8.2 ms GPU latency.

Massive machine-type communications congest the random-access channel and burn power on repeated preamble transmissions. DTDO-RA adjusts the backoff indicator (BI) according to the number of transmissions, and uses reinforcement learning — Q-learning and DDPG — to tune the BI together with the maximum number of preamble transmissions (Max Tx) as the network state changes. Widening the backoff range under heavy traffic keeps the procedure succeeding: the success rate, which falls at 10,000 UEs under the standard procedure, holds until 15,000 UEs. At 15,000 UEs a larger BI cuts transmit power by about 98% (3.71 to 0.072 mJ) at the cost of longer access delay, while letting Max Tx widen the backoff range lowers the delay by up to 4.3× relative to BI adjustment. Simulations run to 50,000 UEs and also show that shrinking Max Tx too far raises power rather than saving it.

GANs are widely used to augment scarce data, but a discriminator that leans too hard on the training set overfits, and the generator ends up reproducing images that look like the training images — augmentation that adds nothing. LDGAN opens the discriminator with LIME to show which image regions drive its real/fake decision, and finds that an unconstrained discriminator relies on the whole image, background included. It then limits the discriminator's training with a threshold, set either on the discriminator's mean loss (about 0.5) or on the generator's loss, so that the decision rests on the object rather than the background and the generated images stay diverse. Of 500 images generated per model, 254 from DCGAN show usable eye and nose regions, against 324 and 360 under the two LDGAN limits.