TL;DR
A dynamic pruning method that uses reinforcement learning guided by a dynamic reward function to automatically explore the pruning rate of each layer, and uses the designed similarity-based connected graph (SCG) mechanism to perform channel removal operations during training.
Indicative in these results is that intelligent batch ordering can partially compensate for reduced architectural complexity, thus offering a computationally efficient alternative to deeper models.
Deep neural networks often contain substantial parameter redundancy, resulting in unnecessary computational cost and energy consumption. This work presents a dynamically adaptive gating mechanism for learning layer-wise sparsity through differentiable masking. In the proposed framework, each network parameter is associated with a self-learning gate that controls its contribution during training. The gating function follows a progressive soft-to-hard transition in which the slope is gradually annealed, enabling the model to move smoothly from continuous parameter weighting to near-binary pruning decisions while simultaneously learning an adaptive threshold parameter. The framework is evaluated across multilayer perceptrons (MLPs), deep neural networks (DNNs), Tabular Transformer models, and benchmarked against $\mathrm{L}_{0}$ regularization and Variational Dropout on convolutional neural networks. Experimental results demonstrate substantial model compression without degrading predictive performance, and in some cases improving it. On MNIST, the gated MLP retains only 14% of weights (86% pruning) while maintaining 98% accuracy. The gated DNN and Tab Transformer similarly outperform their dense counterparts while retaining only 46% and 46.22% of parameters, respectively. Benchmarking further shows that the proposed adaptive gating achieves competitive or superior accuracy-sparsity trade-offs compared with $\mathbf{L}_{0}$ regularization and Variational Dropout. These results demonstrate that the proposed dynamically adaptive gating framework provides an efficient and interpretable pathway for sparsity learning, enabling high-performing lightweight neural network deployment.
Convolutional neural networks (CNNs) are widely used for image segmentation and classification due to their strong ability to learn spatial representations. However, optimization stability and convergence efficiency remain challenging, particularly for tasks involving unclear or overlapping boundaries. Predefined learning rate schedules, such as cyclical learning rate (CLR), are commonly used in CNN training but offer limited adaptability to training dynamics. This study proposes the Cyclical Local Adaptive Learning Rate Strategy (CLARA), an enhanced CLR that introduces scheduler-level adaptive learning rate adjustment. CLARA dynamically adjusts the lower learning rate bound in response to training behaviour to improve convergence consistency during training. Integrated with U-Net, CLARA was evaluated on overlapping chromosome segmentation and achieved an Intersection over Union of 99.98% with 99.99% validation accuracy, while also producing more consistent segmentation outputs and improved separation in overlapping regions. Experiments on wafer defect classification using ResNet-50 achieved 99.39% validation accuracy, suggesting the applicability of the proposed scheduler across the evaluated segmentation and classification tasks.
Deep reinforcement learning (DRL) remains constrained by high computational costs, hindering its practical deployment. While neural network pruning offers a solution for model compression, most existing DRL approaches rely on unstructured pruning. This results in irregular sparse matrices that lack compatibility with standard hardware, offering no tangible real-time acceleration. Furthermore, pruning in DRL is notoriously challenging due to inherent training instability, which often leads to catastrophic performance degradation. To overcome these limitations, we propose a novel framework that integrates dynamic structured pruning with model merging. By periodically merging parallel network instances, our method effectively counteracts the instability triggered by aggressive structural changes, enabling the use of structured pruning directly compatible with general-purpose hardware. Experimental results on continuous control tasks demonstrate that our approach reduces cumulative training FLOPs by up to 72% while maintaining performance competitive with dense baselines across the majority of environments. Additionally, inference profiling confirms an average latency reduction of 16.2%, highlighting the framework's potential for accelerating both the training and deployment of DRL agents. Our code is available at https://github.com/hail-mary/neuron-pruning.
This work proposes LaRes, a novel hybrid framework that achieves efficient policy learning through reward function search by leveraging large language models to generate the reward function population, guiding RL in policy learning.
Goal-directed optimization is essential for steering molecular generators to propose candidates with desired properties. However, it is often implemented with policy-gradient reinforcement learning, which requires a generation-trajectory log-probability whose form depends on the model architecture and generation procedure. This makes an optimizer difficult to reuse across architectures and conditional generative designs. Supervised fine-tuning needs none of that machinery, but its update is driven by a fixed dataset, so the reward never enters the update. We introduce Elite-Weighted Supervised Fine-tuning (EW-SFT), which uses reward to guide elite selection of high-scoring molecules, and updates the model by its own pretraining loss on that set. Ablations show that reward information is passed primarily through elite selection, rather than through continuous weighting within the selected set. Because the update consumes only scored molecules and the model's native loss, the same rule applies across autoregressive, masked-diffusion, and discrete-flow generators, and across de novo, motif-extension, and linker-design tasks. Under a fixed budget of 3D shape alignment oracle calls on two kinase reference compounds, EW-SFT consistently outperforms the corresponding native optimizers. It further improves goal-directed optimization under a 2D similarity oracle on four held-out references and achieves comparable performance on a sample-efficiency benchmark without a trajectory-level RL formulation. These results demonstrate that EW-SFT is a unified and effective optimizer across molecular generators, design constraints, references, and oracles.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
