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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Efficient Training of Minimal and Maximal Low-Rank Recurrent Neural Networks

Authors: Anushri Arora, Jonathan Pillow

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that this perspective leads to an online training method that achieves higher accuracy with smaller networks than previous methods such as FORCE, and outperform backprop-trained networks of similar size while requiring substantially less training time. We first evaluate on the classical sine wave generation task, originally used to benchmark FORCE [9]. Our rank-2 networks... consistently outperform FORCE-trained full-rank networks... We next evaluate on a binary decision-making task modeled by a bistable attractor ODE. Both our method and networks trained with BP, learn from teacher trajectories that start at random initial conditions and converge to one of two fixed points (training details: SI E). As shown in Figure 3.B, our approach achieves lower test error across comparable network sizes than both full-rank and low-rank BP-trained networks.
Researcher Affiliation Academia Anushri Arora Department of Computer Science Princeton University EMAIL Jonathan W. Pillow Princeton Neuroscience Institute Princeton University EMAIL
Pseudocode Yes We note, our NEF approach can be modified for such a setting via online recursive updates to the weight vector n. This translates to a recursive least-squares (RLS) optimization scheme (SIAlgorithm 1 summarizes our implementation). A more detailed description of this process is provided in SIAlgorithm 2.
Open Source Code Yes 1Code: https://github.com/anushri10/Efficient-Training-of-Minimal-and-Maximal-Low-Rank-RNNs.git
Open Datasets No We first evaluate on the classical sine wave generation task, originally used to benchmark FORCE [9]. Our rank-2 networks, with targets comprising both the sine signal and its cumulative integral... We next evaluate on a binary decision-making task modeled by a bistable attractor ODE. Both our method and networks trained with BP, learn from teacher trajectories that start at random initial conditions and converge to one of two fixed points (training details: SI E).
Dataset Splits Yes To generate each trajectory, we used Euler integration with a time step of dt = 0.01 over a duration of 4 seconds, yielding 400 time steps per trajectory. A total of 160 unique initial conditions were uniformly sampled along the z-axis. Of these, 150 were used for training and the remaining 10 for testing.
Hardware Specification No The paper does not specify the hardware used for running the experiments. It provides training times in Table 1 but without any hardware context.
Software Dependencies No No explicit mention of specific software dependencies with version numbers (e.g., library names with versions like PyTorch 1.9 or TensorFlow 2.x) is found in the paper.
Experiment Setup Yes To generate each trajectory, we used Euler integration with a time step of dt = 0.01 over a duration of 4 seconds, yielding 400 time steps per trajectory... Of these, 150 were used for training and the remaining 10 for testing. Each of the networks trained via BP in Fig 3 were trained for 15 epochs, with a batch size of 10 per epoch. Thus, a total of 150 gradient steps were performed (performance plateaued at 100 gradient steps). Additionally, each of the networks were trained over three random seeds of initializations, where parameters were initialized from standard normal distributions.