Research map / Architectures and efficiency / RNNs and normalization
Structured flexibility in recurrent neural networks via neuromodulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the neuromodulated RNN (NM-RNN), where a small neuromodulatory RNN outputs a low-dimensional signal that scales the rank-1 components of a low-rank output-generating RNN's recurrent weights, and shows it trains and generalizes better than low-rank RNNs on canonical neuroscience tasks.
Problem
Task-optimized RNNs used in neuroscience have fixed recurrent weights, whereas synaptic strengths are dynamically modulated by neuromodulators; it is unclear what computational benefit such synaptic gain scaling provides.
Approach
Two linked networks: a small full-rank neuromodulatory RNN whose sigmoid readout scales each rank-1 component of the recurrent weight matrix of a larger low-rank RNN, yielding time-varying gain-based structured flexibility that resembles LSTM gating.
Key results
- On the Measure-Wait-Go task, NM-RNNs consistently achieve lower loss than LR-RNNs on extrapolated intervals across 10 parameter-matched runs.
- On the multitask setting, NM-RNN retraining of only the neuromodulatory subnetwork on MemoryAnti matches LSTM and higher-rank vanilla RNN and considerably outperforms the LR-RNN.
- On the Element Finder Task (T=25), NM-RNNs solve the task like LSTMs while low-rank and full-rank RNNs largely fail.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neuromodulated recurrent weights instead of fixed low-rank recurrent weights | NM-RNN vs. LR-RNN · Measure-Wait-Go extrapolation | L2 loss | Helped |
| Use neuromodulated recurrent weights instead of fixed low-rank recurrent weights | NM-RNN vs. LR-RNN · Measure-Wait-Go | output ramp accuracy | Helped |
| Use neuromodulated recurrent weights instead of fixed low-rank recurrent weights | NM-RNN vs. LR-RNN · DelayPro, DelayAnti, MemoryPro, and MemoryAnti | percent correct | Helped |
| Use an NM-RNN with feedback coupling from the output-generating subnetwork to the neuromodulatory subnetwork | NM-RNN · Element Finder Task | MSE loss | Helped |
| Add neuromodulatory component s1(t) (ablation) | Rank-3 NM-RNN · Measure-Wait-Go | task performance | Helped |
| Add neuromodulatory component s2(t) (ablation) | Rank-3 NM-RNN · Measure-Wait-Go | task performance | Helped |
| Add neuromodulatory component s3(t) (ablation) | Rank-3 NM-RNN · Measure-Wait-Go | output ramp slope accuracy | Helped |
Surprising findings
- The NM-RNN's gain advantage over LR-RNN in the multitask setting is not explained by retraining more parameters, since the comparison LR-RNN retrains a larger input weight matrix.
Limitations
- Networks are small (about N≈100 neurons), far fewer than biological or other related models.
- Results have not been compared to neural recording data, limiting biological conclusions.
- Only post-training dynamics were analyzed; effects during learning were not studied.
Experimental setup
Datasets: Measure-Wait-Go interval task DelayPro/DelayAnti/MemoryPro/MemoryAnti four-task set (Duncker et al.) Element Finder Task
Benchmarks: Measure-Wait-Go interval reproduction multitask DelayPro/DelayAnti/MemoryPro/MemoryAnti with MemoryAnti retraining Element Finder Task
Metrics: L2 loss MSE loss percent correct
Baselines: low-rank RNN (LR-RNN) vanilla RNN LSTM full-rank RNN
Builds on: low-rank RNN LSTM HyperRNN Beiran et al. low-rank RNN timing model Driscoll et al. multitask RNN
Closest papers by meaning
- Measuring and Controlling Solution Degeneracy across Task-Trained Recurrent Neural Networks (NeurIPS 2025 · similarity 0.87)
- How connectivity structure shapes rich and lazy learning in neural circuits (ICLR 2024 · similarity 0.87)
- Scaling Laws and Tradeoffs in Recurrent Networks of Expressive Neurons (NeurIPS 2026 · similarity 0.87)
- Flexible task abstractions emerge in linear networks with fast and bounded units (NeurIPS 2024 · similarity 0.86)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.86)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.