Multi-Task Green Learning (MTGL)

• Multi-Task Green Learning (MTGL) is a paradigm that enables simultaneous learning of multiple tasks while reducing energy, memory, and computational demands.
• It employs feed-forward unsupervised subspace learning and tree-based classifiers to achieve state-of-the-art segmentation and classification with exponential resource savings.
• MTGL’s energy-aware task scheduling and explicit interpretability build trust in high-stakes applications such as medical imaging and robotics.

Multi-Task Green Learning (MTGL) is a paradigm that targets simultaneous, resource-efficient learning of multiple tasks through algorithmic and architectural innovations that minimize computational, energy, and memory demands, while ensuring interpretability and transferability. MTGL frameworks are distinguished by their explicit trade-off between predictive accuracy and energy consumption, and by their avoidance of unnecessary parameter growth or opaque representation learning. This approach is motivated by the limitations of conventional deep learning systems, which are often characterized by high parameter counts, significant energy and carbon footprints, and “black-box” behaviors that impede clinical trust and adoption in sensitive domains such as medical imaging (Kao et al., 27 Jan 2026, Say et al., 1 Apr 2025).

1. Foundational Principles and Motivation

MTGL inherits its motivation from the drive for sustainable, interpretable, and practical artificial intelligence. Conventional multi-task learning architectures, especially those relying on deep neural networks trained via backpropagation, commonly suffer from:

• Large parameter counts: Models like 3D V-Net may contain tens of millions of parameters, which inflates memory and computational resource requirements.
• High energy costs: Prolonged training translates to higher power consumption and CO₂ emissions.
• Data-hungriness: Such models often require thousands of labeled data points per task.
• Interpretability limitations: End-to-end learned representations typically lack transparency, which is particularly problematic in high-stakes domains (e.g., medical diagnostics).

Green Learning (GL) concepts address these limitations with feed-forward, modular, backpropagation-free systems; unsupervised subspace learning replaces gradient descent, while closed-form regression and tree-based classifiers support supervised modules. These approaches deliver a dramatic reduction in parameter count, FLOPs, and memory requirements, with intrinsic interpretability due to fully linear or tree-based operations (Kao et al., 27 Jan 2026).

2. Architectures and Algorithms

MTGL instantiates diverse architectural designs, but a unifying theme is the reduction or explicit management of computational resources across multi-task pipelines.

2.1 Backpropagation-Free Feed-Forward MTGL

Kao et al. (Kao et al., 27 Jan 2026) propose a fully backpropagation-free system for simultaneous left ventricle segmentation and ejection fraction classification in echocardiography, structured as follows:

• Hierarchical, unsupervised VoxelHop encoder: Employing the Successive Subspace Learning (Saab transform), each layer extracts principal components (DC plus dominant AC filters) from local spatio-temporal patches, propagating only the most energetic AC channels at each resolution.
  • For layer ℓ, the feature vector $\mathbf{x} \in \mathbb{R}^{d}$ is projected via DC and $K_\ell$ principal AC directions selected to explain at least $99\%$ of input variance.
  • Aggressive channel pruning retains a tiny fraction of AC filters at higher hops, yielding exponential resource savings (e.g., $643$ of $11{,}102$ ACs at the 4th hop, $5.8\%$).
• Coarse-to-fine multi-level regression decoder: Segmentation is performed by a series of XGBoost regressors at increasingly finer scales, each refining the upsampled mask from the previous level by predicting local residuals, especially near boundaries.
• XGBoost classification head: Features from all encoder hops are aggregated via spatial pyramid pooling and global/max pooling, concatenated, and fed to an XGBoost classifier for ejection fraction stratification.

2.2 Dynamic, Energy-Conscious Task Scheduling

Najarro et al. (Say et al., 1 Apr 2025) introduce MTGL in the online continual/robotic learning context, emphasizing:

• Interleaved, progress-guided task selection: At each training step, task selection is determined by a score combining recent learning progress (slope of error decrease) and per-task energy consumption (measured by normalized neuron activations or FLOPs).
  • The score $$s^{(t)}(i) = \exp(k \cdot LP^{(t)}(i)) / EC^{(t)}(i)$$, where $LP^{(t)}$ is learning progress and $EC^{(t)}$ is energy cost, enables tunable prioritization of either metric.
  • An exploration factor $\varepsilon$ injects occasional randomness.
• Lightweight, shared encoder–attention–decoder architecture: Each task has its own projection and decoding modules, with shared state encoding, attention, and task flags to facilitate transfer and tracking.

3. Resource Efficiency and Environmental Impact

MTGL achieves marked superiority in computational, energy, and memory efficiency versus conventional deep neural multi-task baselines.

Model	Parameters (M)	Energy (kWh)	CO₂ Emissions (kg)
3D V-Net	45.61 (40.4×)	7.74	3.25
3D UNETR	10.47 (9.5×)	5.04	2.12
3D U-Net	16.32 (14.4×)	5.50	2.31
3D nnU-Net	49.08 (43.4×)	9.22	3.87
MTGL	1.13	4.97	2.09

The feed-forward VoxelHop+XGBoost MTGL system reaches state-of-the-art segmentation (DSC = $0.912$, IoU = $0.838$) and classification (accuracy $0.943$) on EchoNet-Dynamic at a fraction (1/10–1/40) of the parameter and energy budget required by conventional 3D segmentation networks (Kao et al., 27 Jan 2026). Similarly, energy-weighted interleaved learning (INTER-LPE) can reduce energy use by $20$–$30\%$ for effect-prediction robotics without substantial loss in predictive accuracy (Say et al., 1 Apr 2025).

4. Interpretability and Clinical/Operational Transparency

Interpretability is central in MTGL. Feed-forward architectures eschew millions of hidden parameters in favor of operations that are linear (Saab filters) or tree-structured (XGBoost), where feature importances and splitting criteria are accessible and visualizable. For instance:

• VoxelHop filters at the first hop correspond to spatial edges, demarcating anatomical borders. Deeper hops (e.g., hop 4) relate to temporal motion features such as contraction patterns.
• Multi-level regression exposes which features correct which errors (e.g., edge voxels in segmentation), and residual analyses can localize failure points.
• Feature importances from XGBoost classifiers and regressors provide direct quantification of which inputs drive decisions, facilitating clinical validation and auditability (Kao et al., 27 Jan 2026).

A plausible implication is that interpretability at all model stages aids regulatory review and clinical adoption, as well as failure diagnosis in robotics and other engineered settings.

5. Quantitative Evaluation and Knowledge Transfer

MTGL has demonstrated superior quantitative performance in highly competitive, real-world settings and transfer learning scenarios.

• In echocardiographic imaging, MTGL delivers the narrowest interquartile range in Dice and IoU, indicating robust consistency (paired $t$-test $p<10^{-4}$).
• Full utilization of all encoder hops is critical; ablating shallow or deep hops degrades performance (segmentation fidelity and classification accuracy drop to $91.4\%$ vs. $94.3\%$).
• In online robotic effect prediction, interleaved learning schedules outperform sequential or randomly mixed task ordering by enabling faster convergence, higher sample efficiency, and strong resistance to catastrophic forgetting. The energy-weighted trade-off curve provided by INTER-LPE allows practitioners to tune between least-energy and fastest-learning regimes as application constraints demand (Say et al., 1 Apr 2025).

6. Limitations and Prospects

MTGL’s "green" credentials depend on accurate energy computation, which is currently approximated by activation counts or FLOPs rather than hardware-level measurements; extending this with real power monitoring is a prospective area.

Demonstrations of MTGL have focused on a limited range of tasks: for example, two-task medical imaging (segmentation/classification) and three-task robotic effect prediction (Push, Hit, Stack). Scaling MTGL to continual, lifelong learning with real-world robotics, expanding the repertoire of clinical tasks, and integrating reinforcement learning objectives remain open challenges (Say et al., 1 Apr 2025). Nonetheless, the architectural principles—feed-forward unsupervised subspace learning, interpretable regressors/classifiers, and energy/learning progress-aware scheduling—generalize to a wide variety of multi-task settings.

7. Broader Implications and Generalizability

MTGL offers a blueprint for sustainable, transparent, and resource-aware multi-task learning applicable beyond immediate case studies. The VoxelHop plus coarse-to-fine regression and tree-based classifier recipe is adaptable to complex imaging (e.g., tumor segmentation and grade classification in MRI), robotics (effect prediction, lifelong manipulation), and energy- or privacy-constrained edge deployment scenarios. By challenging the “bigger-is-better” assumption, MTGL demonstrates that rigorous control over computational complexity and interpretability can yield systems with competitive or superior empirical performance, efficiency, and real-world deployability (Kao et al., 27 Jan 2026, Say et al., 1 Apr 2025).