knowledge distillation advanced

**Advanced Knowledge Distillation** encompasses **sophisticated techniques for transferring knowledge from a large teacher model to a smaller student model beyond basic soft-label matching** — including intermediate feature distillation, attention transfer, relational knowledge distillation, and task-specific distillation strategies that enable students to capture structural knowledge the teacher has learned, not just its output predictions. **Beyond Basic Logit Distillation** Basic KD (Hinton 2015) matches the student's softmax output to the teacher's soft labels. Advanced methods distill knowledge from multiple levels: ```svg Teacher Model Student Model┌─────────────┐ ┌──────────────┐ Input Layer Input Layer ├─────────────┤ ├──────────────┤ Hidden 1 │──── Feature ────→│ Hidden 1 Hidden 2 Distillation Hidden 3 │──── Attention ──→│ Hidden 2 Hidden 4 Transfer ├─────────────┤ ├──────────────┤ Logits │──── Logit KD ──→│ Logits └─────────────┘ └──────────────┘ ``` **Feature/Intermediate Layer Distillation** | Method | What is Distilled | Loss | |--------|------------------|------| | FitNets | Hidden layer activations | MSE(student_feat, teacher_feat) via adapter | | Attention Transfer (AT) | Attention maps (spatial) | MSE on attention map norm | | PKT (Probabilistic KT) | Feature distribution in embedding space | KL divergence | | NST (Neuron Selectivity) | Neuron activation distributions | MMD (Maximum Mean Discrepancy) | | CRD (Contrastive Rep. Dist.) | Representation structure | Contrastive loss | Adapter layers (1×1 conv or linear projection) bridge dimension mismatches between teacher and student hidden layers. **Relational Knowledge Distillation (RKD)** Instead of matching individual outputs, RKD transfers the **relationships** between samples: ```python # Distance-wise RKD: preserve pairwise distance structure teacher_dist = pairwise_distance(teacher_embeddings) # NxN matrix student_dist = pairwise_distance(student_embeddings) loss_rkd = huber_loss(student_dist / student_dist.mean(), teacher_dist / teacher_dist.mean()) # Angle-wise RKD: preserve angular relationships among triplets # Captures higher-order structural information ``` **LLM-Specific Distillation** Distilling large language models has unique considerations: - **Black-box distillation**: When teacher weights are inaccessible (GPT-4 → open model), use only generated outputs. Techniques: instruction following data generation, chain-of-thought distillation (distill reasoning traces, not just answers). - **White-box distillation**: With teacher weight access, match logit distributions over the full vocabulary (50K+ dimension KL divergence) and intermediate transformer layer representations. - **Progressive distillation**: Gradually reduce model size through multiple distillation stages rather than one large compression step. - **Distillation for specific capabilities**: Selectively distill math reasoning, code generation, or instruction following by curating task-specific transfer sets. **Multi-Teacher and Self-Distillation** - **Multi-teacher**: Ensemble of specialists, each contributing expertise. Student learns from domain-weighted combination of teacher outputs. - **Self-distillation**: Model distills knowledge from its own deeper layers to shallower layers, or from a previous training epoch to the current one. - **Born-Again Networks**: Iteratively distill — student becomes the new teacher for the next round, often surpassing the original teacher. **Advanced knowledge distillation is the primary model compression technique enabling deployment of LLM-class intelligence on resource-constrained devices** — by transferring not just predictions but structural, relational, and intermediate representations, modern distillation achieves compression ratios of 10-100× while retaining 90-98% of teacher performance.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account