CTDS: Centralized Teacher, Decentralized Student

• CTDS is a framework that decouples centralized teaching from decentralized execution, where a high-capacity teacher guides local student models with partial information.
• It employs distillation techniques like Q-value regression, feature-matching, and pseudo-labeling across domains such as MARL, federated learning, and neural compression.
• The framework enhances sample efficiency and performance, achieving significant gains in win-rate, accuracy, and parameter reduction while maintaining privacy and modularity.

A Centralized Teacher with Decentralized Student (CTDS) framework refers to a family of training and model compression paradigms in which a single high-capacity, centralized teacher network coordinates, guides, or distills its knowledge into one or more decentralized student networks. These students possess no access to the full training context or global observations during inference, receiving distilled supervision or structure from the teacher either during training or via carefully designed communication protocols. CTDS instances span multi-agent reinforcement learning, federated semi-supervised learning, neural network compression, and collaborative or mutual knowledge distillation. This modular design enables the exploitation of global information and powerful representations during training, while maintaining efficiency, modularity, and privacy during decentralized inference or deployment.

1. Core Principles and Architectures

CTDS segregates the roles of teacher and student(s) across training and inference. The central teacher, often with privileged or centralized access (e.g., to global states in MARL (Zhao et al., 2022), all client updates in federated learning (Zhao et al., 2023), or deep internal representations in compression (Malik et al., 2020)), learns richer representations and provides distilled guidance. The decentralized students interact with their own partial information sources or data shards, and are optimized to match, approximate, or reconstruct the teacher’s outputs or features. The main architectural motifs include:

• Centralized Value or Representation Learning: Teacher models leverage holistic state, full dataset, or aggregate information inaccessible to any decentralized agent/student.
• Decentralized Student Learning and Execution: Each student operates with strictly local, partial, or restricted context, reflecting real-world limitations in observability, communication, or data locality.
• Distillation and Knowledge Transfer: Supervision is delivered via regression to Q-values (Zhao et al., 2022), feature-matching (Malik et al., 2020), soft-label projections, or pseudo-labeling under various augmentation/pseudo-label regimes (Zhao et al., 2023).
• Plugin Compatibility: CTDS structures are typically generic and compatible with existing frameworks, e.g., value-decomposition in MARL (VDN, QMIX) or any classification head in compression.

2. Mathematical Formulations

The mechanics and loss terms of CTDS instantiate across domains:

• Reinforcement Learning (MARL): In the Dec-POMDP setting, the teacher's Q-networks have global access ($\hat Q_i(s_t, a)$), while each student learns $Q_i(\tau_i, a)$ with only local (history, partial observation) inputs. The student minimizes an MSE loss to the teacher's Q-values:

$$L_{\rm distill} = \sum_{i=1}^n\sum_{a\in\mathcal A}[\hat Q_i(\hat\tau_i, a) - Q_i(\tau_i, a)]^2$$

(Zhao et al., 2022)

• Neural Compression: The teacher’s dense feature vector $d \in \mathbb{R}^p$ is partitioned into $n$ non-overlapping chunks. Each student $k$ learns $f_{S_k}: \mathcal{X} \to \mathbb{R}^{p_k}$ to reconstruct $d_k$ using mean-squared error or adversarial (GAN) loss. Aggregation for downstream tasks is through simple concatenation and an output head $g$:

$$L_{k}^{\rm MSE} = \frac{1}{|\mathcal B|} \sum_{x \in \mathcal B} \|d_k(x) - f_{S_k}(x)\|_2^2$$

(Malik et al., 2020)

• Federated Semi-Supervised Learning: A server maintains a global teacher (as EMA of the student model), with student models on devices. Knowledge is distilled by pseudo-labels, using an adaptive protocol based on KL divergence matching between teacher and student prediction distributions to a uniform prior:

$Q_i(\tau_i, a)$0

(Zhao et al., 2023)

• Collaborative Distillation: Students also interact via mutual knowledge transfer (peer-to-peer), as in (Sun et al., 2021), enabling online and self-distillation beyond offline teacher-student transfer.

3. Methodological Variants

CTDS supports multiple methodological realizations:

Domain	Teacher Role	Student(s) Role
Multi-Agent RL (Zhao et al., 2022)	Credit assignment w/global Q	Learn Q-values w/local observations
Federated SSL (Zhao et al., 2023)	EMA pseudo-label generator	On-device SSL + local EMA adaptation
Neural Compression (Malik et al., 2020)	Feature producer	Independent chunk-wise feature distillations
Collaborative/Mutual (Sun et al., 2021)	Pre-trained feature/logit provider	Online distillation (mutual, self, offline)

• MARL: Teacher-student decoupling balances centralized learning (credit assignment, state representation) with decentralized deployment (local Qs, test-time autonomy), systematically closing the information gap of pure CTDE (Zhao et al., 2022).
• Federated Learning: CTDS ensures privacy and communication efficiency by pushing all learning to the student side, sending the teacher model only when pseudo-labeling quality warrants, yielding high accuracy even under severe non-IIDness (Zhao et al., 2023).
• Model Compression: Multiple student models, each assigned to a partitioned feature subspace, collectively recover much of the teacher’s expressivity with 10–30× parameter reductions (Malik et al., 2020).
• Mutual/Collaborative Learning: Decentralized students learn from the teacher, peers, and themselves, transferring both response- and relation-based knowledge, and optimizing a composition of distillation, mutual, and self-supervised losses (Sun et al., 2021).

4. Training Protocols and Algorithms

Training CTDS systems typically involves two synergistic phases:

1. Teacher Learning or Fixation:
   • Teachers either train on global state (RL), aggregate the most recent global student (federated SSL), or are pre-trained on full data (compression, mutual KD).
   • For EMA teachers, parameter updates are performed by a slow moving average of student weights.
2. Student Distillation/Adaptation:
   • Students receive distilled supervision (Q-values, features, pseudo-labels) and optimize accordingly.
   • In some cases, post-distillation fine-tuning of classification heads or aggregation layers is performed to recover full-task performance (compression).
   • Federated setups incorporate communication-efficient update protocols (e.g., FedSwitch adaptively transmits teacher model parameters to select clients only when needed) (Zhao et al., 2023).
   • Online or collaborative settings involve additional peer-to-peer and self-regularization loops (Sun et al., 2021).

All variants enforce decentralized execution or inference: after training, only the student (not the teacher) is used for decision-making, typically relying strictly on local observations or on-device data.

5. Theoretical Motivation and Distillation Analysis

CTDS is motivated by the observation that direct local learning in decentralized settings (e.g., MARL) constrains the agent to statistically insufficient partial information, whereas a global teacher with centralized access can model joint states, rich correlations, or data distributions more effectively. The distillation process enables each student to learn an implicit marginalization or expectation over latent/unobserved variables, as formulated in (Zhao et al., 2022):

$Q_i(\tau_i, a)$1

This theoretical insight extends to other regimes: partitioning dense representations (compression), aggregating global statistical knowledge (federated), and leveraging peer diversity (collaborative KD), thereby improving both the sample efficiency and generalization of decentralized agents or models.

6. Empirical Results and Comparative Performance

Experimental results consistently demonstrate substantial improvements of CTDS-based approaches over standard baselines:

• Multi-Agent RL (StarCraft II SMAC): CTDS yields 15–20% absolute win-rate gains (e.g., QMIX+CTDS outperforms vanilla QMIX by 16.7%) and better robustness under constrained observability (Zhao et al., 2022).
• Federated SSL: FedSwitch (CTDS-EMA) achieves ≈90.2% accuracy on CIFAR-10, outperforming both TS-Server EMA and TS-Client EMA, along with significant communication savings and privacy retention (Zhao et al., 2023).
• Model Compression: Teacher-class network approach achieves up to 26.1× reduction in parameters with minimal drop or even improvement in accuracy versus single-student KD (e.g., MNIST: CTDS $Q_i(\tau_i, a)$2 GAN 99.21% vs. KD 93.34%) (Malik et al., 2020).
• Collaborative Learning: Full CTDS mutual/collaborative distillation achieves up to 2–4% accuracy improvement over classical mutual learning and self-KD on CIFAR-100 and Market-1501 (Sun et al., 2021).

Ablations confirm the importance of decentralized student structure (more students, chunked features, or multiple knowledge types), with specific gains from adaptive switching (federated), local teacher adaptation, and combining mutual, response, and relation-based losses (collaborative learning).

7. Significance, Applications, and Limitations

CTDS reconciles the strengths of centralized learning (maximal knowledge extraction, accurate credit assignment, data efficiency) with the imperatives of decentralized execution (privacy, autonomy, parallelism, communication efficiency):

• Domains: MARL (credit assignment under partial observability), federated learning (privacy-centric SSL), neural network compression (parameter reduction), mutual knowledge distillation (collaborative intelligence).
• Advantages: Genericity (plug-in to standard value-based MARL, compression, or KD frameworks), superior accuracy/efficiency trade-offs, robustness to non-IIDness, and operational gains (lower latency, scalable on-device inference).
• Limitations: Chunk partitioning and capacity balancing in compression may require extensive tuning; error accumulation across students can require fine-tuning; GAN-augmented variants introduce training instability/complexity (Malik et al., 2020); performance may degrade as the information gap between student and teacher increases without adequate distillation or peer feedback.

Prominent applications include SMAC micromanagement, federated semi-supervised computer vision, scalable model deployment on IoT, and collaborative or distributed learning scenarios (Zhao et al., 2022, Zhao et al., 2023, Malik et al., 2020, Sun et al., 2021).