Real-time variational method for learning neural trajectory and its dynamics

Authors: Matthew Dowling, Yuan Zhao, Il Memming Park

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our method on synthetic and real-world data, and, notably, show that it achieves competitive performance. 5 EXPERIMENTS We first evaluate and compare e VKF to other online variational methods as well as classic filtering methods using synthetic data.
Researcher Affiliation Academia Matthew Dowling Stony Brook University, New York, USA matthew.dowling@stonybrook.edu Yuan Zhao National Institute of Mental Health, USA yuan.zhao@nih.gov Il Memming Park Champalimaud Research, Champalimaud Foundation, Portugal memming.park@research.fchampalimaud.org
Pseudocode Yes Algorithm 1 eVKF
Open Source Code No The paper does not explicitly state that open-source code for the described methodology is provided or include a link to a code repository.
Open Datasets Yes To evaluate e VKF with real-world neural data, we considered electrophysiological recordings taken from monkey motor cortex during a reaching task (Churchland et al., 2012).
Dataset Splits No The paper mentions training data points ('We train each method for 3500 data points') and subsequent inference but does not explicitly specify train/validation/test dataset splits needed for reproduction of data partitioning.
Hardware Specification Yes For measuring the time per step as in Table 2 the experiments were run on a computer with an Intel Xeon E5-2690 CPU at 2.60 GHz.
Software Dependencies No The paper mentions using 'Adam' and 'Si LU nonlinearity' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes To parameterize the dynamics, pθ(zt | zt 1), we use a single layer MLP with 32 hidden units and Si LU (Elfwing et al., 2018) nonlinearity. During training we use Adam (Kingma & Ba, 2014), and update the dynamics every 150 time steps. In total we use 3500 time points for training the dynamics model for all methods.