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..
Bounds on the computational complexity of neurons due to dendritic morphology
Authors: Anamika Agrawal, Michael Buice
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For random Boolean functions, we show that there is a phase transition in learnability as a function of the input dimension, with most random functions below a certain critical dimension being learnable and those above not. This critical dimension is best predicted by the overall size of the dendritic arbor. This demonstrates that real neurons have a far higher computational complexity than is usually considered in neural models, whether in machine learning or computational neuroscience. Furthermore, using architectures that are, respectively, more apical" or basal" we show that there are non-trivially disjoint sets of learnable functions by each type of neuron. Importantly, these two types of architectures differ in the robustness and generality of the computations they can perform. The basal-like architecture shows a higher probability of function realization, while the apical-like architecture shows an advantage with fast retraining for different functions. |
| Researcher Affiliation | Collaboration | Anamika Agrawal Center for Data-Driven Discovery Allen Institute Seattle, WA 98109 Department of Neurobiology and Biophysics University of Washington Seattle, WA 98195 EMAIL |
| Pseudocode | No | The paper describes the model with mathematical equations and network schematics (Figure 1A) and details the training and optimization process in paragraph form, but it does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | Code associated with this work will be available at github.com/Anamika Ag/dend Comp Paper Code |
| Open Datasets | No | For dimension Ndim, there are x Ndim = 2Ndim possible inputs and nfunc = 22Ndim total Boolean functions. Because testing all Boolean functions in high dimensions (Ndim > 4 is infeasible, we define three representative families that span the space of task complexity. First, the class of typical functions: randomly sampled outputs where each row of the truth table returns 1 or 0 with equal probability (p = 0.5). |
| Dataset Splits | No | For each Ndim = 3-8, we trained all architectures on identical sampled sets (100-500 functions per class) to ensure a fair comparison. Distributions of entropy and sensitivity across function classes are shown in Supplementary Fig. 6. |
| Hardware Specification | No | The training time for a given architecture and function over one set of initial conditions (trial) took a maximum of 4 hrs (Ndim = 8). For multiple trials and functions, we use parallel computing over a cluster to run training on multiple functions and trials simultaneously. |
| Software Dependencies | No | While training using the gradient-descent based approach, we used Binary Cross-Entropy as the loss function between the predicted and the target outputs during the training. We used an Adam optimizer with a fixed learning rate, with learning rates and epochs tuned according to the input dimension to ensure convergence of loss (ranging from 6e4 epochs with lr = 5e-3 for Ndim = 3 to 6e6 epochs, lr = 1e-4 for Ndim = 8. We also ran instances of training with an Adam optimizer with a learning rate scheduler, but that did not impact realizability trends shown in the paper. Parameters (W b i , hb, θb, ) were initialized from uniform distributions within bounds chosen to match comparable dynamic ranges across architectures. |
| Experiment Setup | Yes | While training using the gradient-descent based approach, we used Binary Cross-Entropy as the loss function between the predicted and the target outputs during the training. We used an Adam optimizer with a fixed learning rate, with learning rates and epochs tuned according to the input dimension to ensure convergence of loss (ranging from 6e4 epochs with lr = 5e-3 for Ndim = 3 to 6e6 epochs, lr = 1e-4 for Ndim = 8. We also ran instances of training with an Adam optimizer with a learning rate scheduler, but that did not impact realizability trends shown in the paper. Parameters (W b i , hb, θb, ) were initialized from uniform distributions within bounds chosen to match comparable dynamic ranges across architectures. The parameter space was constrained with weights bounded between -20 and 20, branch gains (h) between 0 and 10, thresholds (θ) between -1 and 10, and output bias ( ) between -1 and 10. |