Gradient descent algorithms play a crucial role in machine learning applications and beyond, but they might exhibit highly unstable trajectories and strong dependency on their hyperparameters, especially when working with rugged objective functions. In light of these challenges, this work considers an optimization paradigm based on Jackson q -calculus and explores if q-deformed gradients could offer improved properties for optimization problems. Building on previous q -gradient methods with iteration-dependent deformation parameters, we develop an exponentially annealed Variable Jackson q -Gradient Descent (VJqGD) framework. The proposed method employs the deterministic schedule q_k = 1-q_0exp( -k/τ) , which provides explicit control over the initial deformation magnitude and the transition rate toward the classical gradient regime. Our first contribution is to derive theoretical properties for the underlying Jackson q -gradient method. Specifically, under standard smoothness assumptions, we show convergence to a stationary point as well as an 𝒪(1/T) convergence-rate bound for the minimum squared q-gradient norm. Under more stringent conditions on the q-deformation, we achieve additional convergence rates along with logarithmic iteration complexities. To evaluate the practical performance of VJqGD, we conduct an experimental analysis on four benchmark supervised classification datasets from the UCI repository, namely Arrhythmia, Fertility, Pima Indians Diabetes, and SPECT Heart. VJqGD is compared with SGD, Adam, RMSprop, AdamW, AdaBelief, Lion, and Sophia under an identical experimental protocol using the same single-layer model architecture, learning-rate tuning procedure, and fixed training budget of 50 epochs, with no optimizer-specific stopping criterion. Evaluated using accuracy, Macro-F1, balanced accuracy, and AUC-ROC, VJqGD achieves the highest or tied-highest accuracy on all four benchmark datasets in the noise-free setting, reaching its highest accuracy of 0.8700 on Fertility. It also attains the best Macro-F1 score on Arrhythmia ( 0.3981 ), the best balanced accuracy on Fertility ( 0.4947 ), and an AUC-ROC of 0.8345 on the Pima Indians Diabetes dataset, while maintaining competitive performance under 10% label noise.
Image fusion integrates multiple images into a single, information-rich image, utilizing methods such as exposure fusion and visible/near-infrared (NIR) image fusion. In low-light environments, visible images often suffer from significant information loss, whereas near-infrared (NIR) images preserve structural details. To address this challenge, we propose a novel visible–NIR image fusion framework that integrates a modified bright prior (MBP) enhancement model with edge-preserving filtering techniques. The key contribution lies in the combination of MBP-based illumination enhancement and a contrast-aware fusion strategy that efficaciously preserves structural details, textures, and edge information while improving overall image visibility. The framework employs weighted guided image filtering (WGIF) and fast bilateral filtering (FBF) to decompose images into base and detail parts, facilitating adaptive fusion of complementary information from visible and NIR modalities. In addition, a weight-guided detail fusion mechanism is introduced to retain fine textures and salient features. Compared with existing fusion frameworks, including recent deep learning-based methods, the proposed method produces images with enhanced contrast, improved detail preservation, and superior visual quality, particularly under challenging low-light conditions. The performance of the proposed method was assessed using both visual and quantitative evaluations against existing image fusion techniques. Six no-reference image quality metrics were utilized for objective comparison. The obtained results demonstrate that the proposed approach achieves superior fusion quality, providing enhanced visual appearance and improved metric scores compared to state-of-the-art methods. Overall, both qualitative and quantitative analyses validate the effectiveness of the proposed framework.
Skeleton-based hand gesture recognition with graph convolutional networks (GCNs) has achieved strong benchmark performance, yet a critical preprocessing inconsistency undermines the fidelity of temporal modeling. Existing methods normalize variable-length gesture sequences to a fixed temporal extent via uniform linear interpolation before network ingestion, erasing absolute velocity information and rendering per-frame topology semantically inconsistent across instances. We expose this contradiction through a duration-stratified accuracy analysis, demonstrating that state-of-the-art models suffer an accuracy gap between fast and slow-executed instances of the same gesture class, a monotonically speed-dependent degradation consistent with the velocity-compression hypothesis. To address this, we propose Speed-Invariant Phase Normalization (SIPN), a two-stage, architecture-agnostic preprocessing pipeline that replaces only the data loader’s resampling step. The first stage aligns each input sequence to a class-specific DTW Barycenter prototype, placing all instances on a shared gestural phase axis so that temporal position carries consistent semantic meaning across samples. The second stage, Motion-Gated Temporal Sampling (MGTS), resamples the phase-aligned sequence to the network’s required fixed length via stratified sampling biased toward high-motion frames, governed by a single learnable scalar per modality stream, with the sampling distribution scaled by temperature. SIPN raises DSTSA-GCN from 97.74 https://github.com/pytorchteam/SIPN .
This paper outlines an efficient hybrid method to encrypt audio streams by integrating ChaCha20 as a stream-cipher into Sine/Cosine Non-Linear Transformations. The proposed method works with large audio files in WAV format (.wav) having high bit depth (16 bits, 24 bits, 32 bits), applies multi-stage diffusion/confusion in each sample of the signal, and does this at very low computation costs. The scheme is evaluated over a corpus of 1,500 uncompressed PCM WAV files (500 per bit-depth category) assembled from MUSAN and the Free Music Archive by stratified sampling across speech, music and ambient content, using Peak Signal-to-Noise Ratio (PSNR), Signal-to-Noise Ratio (SNR), Mean Squared Error (MSE) and correlation coefficient. Two distinct comparisons are reported: (i) obfuscation quality (original plaintext audio vs. ciphertext), which is expected to be low, confirming strong obfuscation; and (ii) reconstruction fidelity (original plaintext audio vs. decrypted audio), which is expected to be near-perfect. Reconstruction is near-lossless at every bit depth: correlation between original and decrypted audio is 1.000000 with zero variance across all 1,500 files, the maximum absolute sample error anywhere in the corpus is one least-significant bit, 337 of 1,500 files (22.5
Text-to-image (T2I) generative AI models enable scalable visual content creation but often embed or even amplify societal biases, particularly around gender and object associations. For example, T2I models may reinforce restrictive stereotypes by depicting objects like flowers as feminine or laptops as masculine. Current bias assessment methods often involve multi-step evaluation procedures and typically emphasize aggregate demographic distributions, prompt adherence, or profession-level associations rather than object-level gender–object association analysis. This paper proposes a bias assessment pipeline that combines object detection, perceived-gender identification, object-association scoring, and heatmap visualization to analyse gender–object association bias in T2I-generated images. To support evaluation, we introduce the Object Bias Image Dataset (OBID), designed with a 10,000-image generation target, and evaluate the proposed pipeline on 3,247 retained images that satisfy the requirements for gender–object association analysis. Results reveal stronger class-wise diagonal agreement for neutral objects (0.59) than for male-associated (0.27) and female-associated (0.21) objects, highlighting both the feasibility and current limitations of automated object-association bias assessment.
