Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Dynamical phases of short-term memory mechanisms in RNNs
Authors: Bariscan Kurtkaya, Fatih Dinc, Mert Yuksekgonul, Marta Blanco-Pozo, Ege Cirakman, Mark Schnitzer, Yucel Yemez, Hidenori Tanaka, Peng Yuan, Nina Miolane
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically verify these results by training and evaluating approximately 80,000 recurrent neural networks (RNNs), which are publicly available for further analysis1. Overall, our work provides new insights into short-term memory mechanisms and proposes experimentally testable predictions for systems neuroscience. |
| Researcher Affiliation | Collaboration | 1Koc University, Turkey 2CNC Program, Stanford University, Stanford, USA 3Geometric Intelligence Lab, UC Santa Barbara, Santa Barbara, USA 4Kavli Institute for Theoretical Physics, UC Santa Barbara, Santa Barbara, USA 5Computer Science, Stanford University, Stanford, USA 6James H. Clark Center for Biomedical Engineering & Sciences, Stanford University, Stanford, USA 7Howard Hughes Medical Institute, Stanford University, Stanford, USA 8Phi Lab, NTT Research, Sunnyvale, USA 9Center for Brain Science, Harvard University, Cambridge, USA 10Institute for Translational Brain Research, Fudan University, Shanghai, China. |
| Pseudocode | No | The paper describes methods using mathematical equations and textual explanations, but does not include any clearly labeled pseudocode or algorithm blocks. For example, it provides the RNN update rule in Equation (1) and (S31), but this is not formatted as pseudocode. |
| Open Source Code | Yes | We empirically verify these results by training and evaluating approximately 80,000 recurrent neural networks (RNNs), which are publicly available for further analysis1. Overall, our work provides new insights into short-term memory mechanisms and proposes experimentally testable predictions for systems neuroscience. 1https://github.com/fatihdinc/dynamical-phases-stm |
| Open Datasets | Yes | 2The dataset is publicly available through https://doi.org/10.5281/zenodo.15529757. To obtain this dataset, we used several computers with NVIDA RTX 3090 GPUs or equivalent, which roughly amounts to 230 kg CO2 emission, 930 km driven by an ICE car, 115 kg coal burned, 4 tree seedlings sequesting carbon for 10 year as computed via https://mlco2.github.io/impact/. |
| Dataset Splits | No | The paper describes training recurrent neural networks (RNNs) on specific tasks and evaluating their performance, but it does not mention or specify any explicit training, validation, or test dataset splits in the conventional sense (e.g., percentages or counts of samples partitioned from a dataset). |
| Hardware Specification | Yes | To obtain this dataset, we used several computers with NVIDA RTX 3090 GPUs or equivalent, which roughly amounts to 230 kg CO2 emission, 930 km driven by an ICE car, 115 kg coal burned, 4 tree seedlings sequesting carbon for 10 year as computed via https://mlco2.github.io/impact/. Most of the experiments were conducted on computing systems equipped with an Intel i9-10900X CPU and an Apple M2 CPU. |
| Software Dependencies | No | All recurrent neural network (RNN) architectures were implemented using the Py Torch framework. Unless stated otherwise, models were trained using the stochastic gradient descent (SGD) with the momentum value of 0.9 and weight decay value of 10 7. The critical exponent β was estimated using scipy.stats.linregress and validated by an additional log-log fit via curve fit. The paper mentions PyTorch and Scipy but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | The learning rate was initialized at α = 10 2. Additionally, we set the time decay constant to τ = 10 ms, the time step to t = 5 ms, the delay interval to Tdelay = 150 ms, and the response interval to Tresp = 50 ms. The architecture was trained for 150, 000 epochs with a total of N = 100 neurons. All weights and biases, as described in Eq. (2), were optimized throughout training. For this experiment, noise was set to ϵ = 0, and firing rates were initialized from a multivariate Gaussian distribution with mean µr = 0 and standard deviation σr = 0.1. Training was performed using stochastic gradient descent (SGD) with Py Torch s default hyperparameters. |