Disentangling 3D Animal Pose Dynamics with Scrubbed Conditional Latent Variables
Authors: Joshua Wu, Hari Koneru, James Ravenel, Anshuman Sabath, James Roach, Shaun Lim, MICHAEL TADROSS, Alex Williams, Timothy W. Dunn
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the utility of this approach in downstream tasks such as clustering, decodability, and motion synthesis. Further, we apply our technique to improve disease detection in a Parkinsonian mouse model. For all results, we trained separate models to isolate the effects of each scrubber on each nuisance variable. We quantified this using an effect size metric, d, calculated as the MMD between a subject s healthy and disease sessions relative to the MMD between healthy sessions across animals (Table 2 left, B.6). We also used a k-NN classifier to predict the type of session (healthy vs. disease) given a behavioral representation (Table 2 middle). SC-VAE-QD outperformed other models on these metrics. |
| Researcher Affiliation | Academia | Joshua H Wu1, Hari Koneru1 J Russell Ravenel1 Anshuman Sabath1 James M Roach1 Shaun SX Lim1 Michael R Tadross1 Alex H Williams2,3, Timothy W Dunn1 1Duke University 2New York University 3Flatiron Institute |
| Pseudocode | Yes | Algorithm 1 SC-VAE-MALS; Algorithm 2 SC-VAE-QD; Algorithm 3 Automatically tuning forgetting factor (λ) |
| Open Source Code | Yes | Our datasets, train-validation-test splits, Python code, and documentation for running all benchmarks can be found on our Github repository. See https://github.com/tdunnlab/scrubvae. |
| Open Datasets | Yes | We used data from n = 4 C57BL/6 mice freely exploring a 30 cm2 open field (324k frames per mouse, 1 hour at 90 fps). For applications, we used n = 36 C57BL/6 mice recorded in 1-hour sessions (90 fps) in this same open field before and after Parkinson s disease (PD) induction (6-OHDA unilateral injection model). Our datasets, train-validation-test splits, Python code, and documentation for running all benchmarks can be found on our Github repository. See https://github.com/tdunnlab/scrubvae. |
| Dataset Splits | Yes | We used a 50-25-25 train-validation-test split for our datasets. All hour-long videos were split into 1-minute long sections and randomized across each dataset while ensuring equal distribution of animal identity. |
| Hardware Specification | Yes | Models were trained on NVIDIA A100 and V100 GPUs. |
| Software Dependencies | No | The paper mentions 'Adam W optimizer (Loshchilov & Hutter, 2017)' and 'Python code' on Github, but does not specify versions for Python or any other key software libraries like PyTorch or TensorFlow, which are necessary for replication. |
| Experiment Setup | Yes | Models were trained using Adam W optimizer (Loshchilov & Hutter, 2017) with a learning rate of 0.0001 and gradient clipping at 10,000. Losses for models had the same weights for joint position error and the VAE ELBO loss. Models were trained for 400 iterations over the training dataset with batch sizes of 2048. The model architecture described in Section C.1 was not changed across tasks. We used a pose sequence window size of 51 frames. The size of the latent dimension zt was 64 for the results in Sections 4.4-4.2 and 32 for the results in Section 4.5. Table 9: Hyperparameter search lists specific lambda, alpha, n_iter, and beta values for different models. |