Overview
Developed a sustainable AI solution for L'Oréal's skin condition classification challenge, focusing on model compression and Green AI principles to reduce computational footprint while maintaining accuracy.
Problem Statement
L'Oréal hosted a Skin condition classification hackathon to build a model that:
- •Accurately identifies multiple skin conditions from textual descriptions of their products
- •Minimizes carbon footprint (Green AI principles)
- •Maintains enterprise-grade accuracy
Data
- •Source: L'Oréal proprietary dataset
- •Size: 6,420 product descriptions
- •Classes: 33 skin conditions
Approach
Knowledge Distillation Pipeline
- •
Teacher Model
- •RoBERTa-large model
- •Fine-tuned to 78.17% validation accuracy
- •Served as knowledge source
- •
Student Model
- •DistilRoBERTa architecture
- •76.9% fewer parameters than teacher
- •Trained using soft labels + hard labels
- •
Optimization
- •Quantization-aware training
- •Pruning of redundant connections
Results & Impact
| Metric | Teacher | Student | Reduction |
|---|---|---|---|
| Accuracy | 78.17% | 76.26% | -2.4% |
| Model Size | 1360.4MB | 190.2MB | 86% |
| CO₂ per 1K inferences | 31.781g | 1.097g | 96.5% |
Key Learnings
- •Knowledge distillation effectively transfers complex patterns
- •Temperature parameter crucial for soft label quality
- •Green AI metrics should be standard in ML reporting
- •Edge deployment constraints drive creative optimization