[论文解读] Motivation is Something You Need

Mehdi Acheli, Qi Yu|arXiv (Cornell University)|Feb 24, 2026

Neural and Behavioral Psychology Studies被引用 0

一句话总结

本论文提出一种受神经科学启发的双模态训练框架，在基础模型与在预设动机条件下激活的更大动机模型之间交替，以提高训练效率和模型性能。

ABSTRACT

This work introduces a novel training paradigm that draws from affective neuroscience. Inspired by the interplay of emotions and cognition in the human brain and more specifically the SEEKING motivational state, we design a dual-model framework where a smaller base model is trained continuously, while a larger motivated model is activated intermittently during predefined "motivation conditions". The framework mimics the emotional state of high curiosity and anticipation of reward in which broader brain regions are recruited to enhance cognitive performance. Exploiting scalable architectures where larger models extend smaller ones, our method enables shared weight updates and selective expansion of network capacity during noteworthy training steps. Empirical evaluation on the image classification task demonstrates that, not only does the alternating training scheme efficiently and effectively enhance the base model compared to a traditional scheme, in some cases, the motivational model also surpasses its standalone counterpart despite seeing less data per epoch. This opens the possibility of simultaneously training two models tailored to different deployment constraints with competitive or superior performance while keeping training cost lower than when training the larger model.

研究动机与目标

• Motivation: replicate SEEKING-like motivated states to enhance learning in neural networks.
• Propose a task-agnostic framework using a base model and a bigger motivated model within scalable architectures.
• Show that alternating training improves the base model and can outperform standalone larger models under certain conditions.
• Demonstrate efficiency gains and a train-once, deploy-twice paradigm for resource-constrained settings.

提出的方法

• Define four core elements: base model, motivated model, motivation condition, and weights map.
• Implement a weights map to align base and motivated models within scalable architectures (ResNet, ViT, EfficientNet).
• Trigger motivation when the training loss decreases for k consecutive batches, switching to training the motivated model.
• Copy weights and optimizer state when switching between states to maintain training continuity.
• Evaluate on image classification tasks across CIFAR, ImageNet, Flowers, Pets using ResNet, ViT, and EfficientNet architectures.
• Use ACC/FLOPs and ACC/F_Ratio as efficiency metrics to compare against baseline and next-level models.

实验结果

研究问题

• RQ1 Does motivation-inspired alternating training improve the base model performance compared to standard training across multiple architectures and datasets?
• RQ2 Can the motivated model achieve competitive or superior performance to standalone larger models while reducing training cost?
• RQ3 How does the choice of motivation condition and weight mapping affect learning and transfer performance?
• RQ4 Is there a practical train-once, deploy-twice workflow enabled by this approach for resource-constrained scenarios?

主要发现

• Motivation-inspired training improves base model accuracy across CIFAR-10, CIFAR-100, ImageNet, ViT, and EfficientNet variants.
• On CIFAR datasets, base models often achieve better accuracy per FLOP than their next-level counterparts, with instances of larger motivated models surpassing standalone counterparts.
• On ImageNet, the motivation-inspired scheme is up to 18x more efficient than the next-level model while delivering performance gains.
• Transfer learning with motivated weights yields 4% to 29% improvements on downstream tasks like CIFAR-10, CIFAR-100, Flowers, and Pets.
• EfficientNet experiments show the motivated model can outperform some classically trained larger models, with notable FLOP efficiency gains (up to 14x).
• Ablation studies confirm the importance of a well-timed motivation condition and show that random activation of the motivated model degrades or, for some architectures, marginally affects performance.