Resolving Interclass Similarities in Sign Language Using Adaptive Loss Function

Co-authored and presented peer-reviewed research at the Young Researchers Conference on Artificial Intelligence & Big Data, hosted by ADA University. The study introduces an 'Adaptive Loss Function' to mitigate interclass similarities in Sign Language Recognition. By dynamically penalizing confusing sign pairs, the model achieved a 23.5% reduction in Word Error Rate (WER) and increased accuracy for similar signs by 18%. This teacher-student collaboration, supported by UFAZ and sponsored by Neurotime, focuses on advancing digital accessibility in Azerbaijan through deep learning and multi-modal computer vision.