Page 693 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Use Case - 5: AI-Powered Video Ringback Tone cities 4.8: Smart home/

Organization: iMUSIC Culture & Technology Co.,Ltd.

Country: China

Contact Persons:

An Lin, lina.ua@ chinatelecom .cn
Guanzheng Xiao, xiaogz.ua@ chinatelecom .cn

1 Use Case Summary Table

Item Details
Category Smart home/cities

Pre-call information and transmission issues, users’ self-expression
Problem Addressed
needs, merchants’ product promotion needs
Model training and fine-tuning constitute core processes in AI system
development. Supervised learning typically drives model training,
involving data preprocessing (cleaning, normalization, augmentation),
architecture selection (e.g., Transformer), loss functions (cross-entropy),
and optimizers. Regularization techniques (Dropout, L2) combined with
early stopping and cross-validation prevent overfitting. Distributed
training and mixed-precision computing accelerate large-scale model
development.
Key Aspects of Solu- Fine-tuning leverages pretrained models (e.g., Bidirectional Encoder
tion Representations from Transformers(BERT), Residual Network(ResNet))
through transfer learning: freezing lower-layer parameters while adapt-
ing top layers for new tasks. It employs reduced learning rates (typically
1/10 of initial training) with category-specific data. Advanced techniques
include layer-wise learning rate adjustment (lower rates for base layers),
knowledge distillation, and adapter module insertion. Data augmenta-
tion and domain adaptation methods (e.g., adversarial training) enhance
cross-domain generalization. Performance validation utilizes hold-out
test sets and metrics (accuracy, F1-score).
Diffusion, transformer, Generative Adversarial Networks(GAN), Retriev-
Technology Keywords
al-Augmented Generation(RAG), Diffusion Transformer(DiT)
Data Availability Private

Metadata (Type of Text, audio, image
Data)

657