Robust Hybrid Learning With Expert Augmentation
Authors: Antoine Wehenkel, Jens Behrmann, Hsiang Hsu, Guillermo Sapiro, Gilles Louppe, Joern-Henrik Jacobsen
TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically validate the expert augmentation on three controlled experiments modelling dynamical systems with ordinary and partial differential equations. Finally, we assess the potential real-world applicability of expert augmentation on a dataset of a real double pendulum. Our experiments on various controlled problems demonstrate that AHMs improve the generalization capabilities of state-of-the-art hybrid learning algorithms on synthetic and real-world data in the amortize setting. Section 4 is titled 'Experiments', and it includes subsections like 'Synthetic experiments' and 'A real world dataset the double pendulum', along with 'Results' which presents 'average log-MSEs over 10 runs' (Figure 5) and 'mean relative precision (in %, indicates one standard deviation) over 10 runs' (Figure 6). |
| Researcher Affiliation | Collaboration | Antoine Wehenkel EMAIL Apple Jens Behrmann EMAIL Apple Hsiang Hsu EMAIL Harvard Guillermo Sapiro EMAIL Apple Gilles Louppe EMAIL University of Liège Jörn-Henrik Jacobsen EMAIL Apple |
| Pseudocode | Yes | Algorithm 1 Expert augmented hybrid learning |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code, nor does it provide links to a code repository or indicate code availability in supplementary materials. The Open Review link is for peer review, not code. |
| Open Datasets | Yes | The dataset of a double pendulum introduced by Asseman et al. (2018) contains 21 videos of the pendulum shown in Figure 4a. |
| Dataset Splits | Yes | We create a dataset with many initial conditions by splitting the videos into consecutive chunks of 20 frames sub-sampled at 100Hz, i.e., 200ms of video. We construct a distribution shift, as shown in Figure 11 from Appendix C.5, over the expert variables ze by splitting each 40 seconds sequence into three parts. The training set only contains chunks from the last 16 seconds of each run. It corresponds to configurations with smaller energy and, thus, slower angular speeds than the test set, which only contains frames from the first 12 seconds. The validation set contains the remaining 12 seconds of frames in the middle. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It mentions training models but no hardware environment. |
| Software Dependencies | No | The paper mentions methods like 'Neural ODEs' and algorithms like 'APHYNITY' and 'Hybrid-VAE' but does not specify any software libraries, frameworks, or solvers with version numbers (e.g., PyTorch 1.9, Python 3.8). |
| Experiment Setup | No | The paper states: 'For all experiments we train the models to maximize p ,Â(y = y1:t1|x = y0) on the training data. We validate and test the models on the predictive distribution p(y = y1:t2|x = y0, xo = y0, yo = y1:t1), where t2 > t1 assesses the generalization over time. The best models are always selected based on validation performance, that is with samples from Ω. In our experiments we use a Gaussian distribution for the posterior, which is equivalent to a mean squared error (MSE) loss on the physical parameters.' However, it lacks specific hyperparameters such as learning rate, batch size, optimizer type, or number of training epochs, which are crucial for reproducibility. |