Sample-Specific Output Constraints for Neural Networks
Authors: Mathis Brosowsky, Florian Keck, Olaf Dünkel, Marius Zöllner6812-6821
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate Constraint Net on two regression tasks: First, we modify a CNN and construct several constraints for facial landmark detection tasks. Second, we demonstrate the application to a follow object controller for vehicles and accomplish safe reinforcement learning in this case. In both experiments, Constraint Net improves performance and we conclude that our approach is promising for applying neural networks in safety-critical environments. |
| Researcher Affiliation | Academia | Mathis Brosowsky 1,2, Florian Keck 2, Olaf Dünkel 2, Marius Zöllner 1,2 1FZI Research Center for Information Technology, 2Karlsruhe Institute of Technology, {brosowsky, zoellner}@fzi.de, {florian.keck, olaf.duenkel}@student.kit.edu |
| Pseudocode | Yes | Algorithm 1 Training algorithm for Constraint Net. The constraint parameter si for a data point (xi, yi) is sampled from a set of valid parameters Syi ={s S : yi C(s)}. |
| Open Source Code | Yes | 1https://github.com/mbroso/constraintnet_facial_detect, 3https://github.com/mbroso/constraintnet_foc |
| Open Datasets | Yes | In this experiment, we consider a facial landmark detection on images of the Large-scale Celeb Faces Attributes (Celeb A) dataset (Liu et al. 2015) and evaluate Constraint Net for different classes of constraints. |
| Dataset Splits | No | The paper does not provide specific dataset split information (exact percentages, sample counts, or detailed splitting methodology) for training, validation, and test sets. It mentions 'test set' but not the overall partitioning. |
| Hardware Specification | Yes | The times are measured on a Lambda Quad (CPU: 12x Intel Core i7-6850K 3.6 GHz, GPU: 4x Nvidia 12 GB Titan V, RAM: 128 GB) using one of the four dedicated graphics cards. |
| Software Dependencies | No | The paper mentions algorithms (e.g., 'Twin Delayed DDPG (TD3) algorithm') but does not specify versions for key software components or libraries (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | No | The paper describes the model architectures and training algorithms used, but does not provide specific hyperparameter values (e.g., learning rate, batch size, number of epochs, optimizer settings) or detailed system-level training configurations in the main text. |