TL;DR
This paper develops a theoretical framework to understand how unobserved ("hidden") neurons in a neural network can distort the measured synaptic interactions between the neurons we can observe. They show that these distortions can be significant, especially in "strongly coupled" networks, and provide a way to predict and potentially control these effective interactions.
What Problem Does This Paper Address?
Neuroscience experiments typically only record from a small fraction of neurons in a circuit. This means that the synaptic connections we measure between recorded neurons are not the "true" connections, but rather "effective" connections that are shaped by the activity of the unobserved neurons. This "hidden unit problem" makes it difficult to accurately interpret experimental data and understand how neural circuit structure gives rise to function. The authors aim to bridge this gap by providing a quantitative relationship between true and effective synaptic interactions, allowing us to understand when and how hidden neurons influence our measurements.
Methods
The authors model neural networks using a nonlinear Hawkes process, a common framework for fitting neural activity data. They then derive a theoretical relationship between the "true" synaptic interactions (what we'd measure in a fully observed network) and the "effective" synaptic interactions (what we measure in a partially observed network). This derivation involves a mean-field approximation to handle the complex interactions with hidden neurons. They also introduce a "Feynmanesque" graphical expansion to visualize and calculate these effective interactions, decomposing them into contributions from paths through the hidden network. For large networks, they analyze specific network architectures (Erdős-Réyni, Dale's law, Watts-Strogatz) and synaptic scaling regimes (weak vs. strong coupling).
Datasets
Not discussed in the paper. The authors use simulated networks to generate their results.
Key Insights and Ideas
- Hidden neurons significantly distort measured connectivity: The "effective" interactions between observed neurons are not simply a subset of the true interactions; they are fundamentally reshaped by the unobserved network.
- "Strongly coupled" networks are particularly susceptible: In networks where synaptic weights scale as 1/√N (a regime observed experimentally), hidden neurons have a much larger impact on effective interactions compared to "weakly coupled" (1/N scaling) networks.
- Effective interactions can have emergent properties: The effective interactions can exhibit different temporal dynamics (e.g., becoming inhibitory after being excitatory) and timescales than the true direct interactions.
- Degeneracy is a key challenge: Different underlying network structures can produce very similar effective interactions, making it difficult to infer the true network from observed data alone.
- A theoretical framework for prediction and control: The derived relationship provides a tool to predict how network properties shape effective interactions and opens avenues for potentially manipulating these interactions.
Key Results
- The authors derive a formula (Eq. 2) that decomposes effective interactions into the true interaction plus correction terms representing paths through hidden neurons.
- In strongly coupled (1/√N scaling) networks, the deviation of effective interactions from true interactions can be substantial (up to 50% relative change) and depends primarily on the fraction of observed neurons, not the absolute network size.
- In weakly coupled (1/N scaling) networks, the influence of hidden neurons is negligible for large networks.
- Simulations of small circuits (3- and 4-neuron networks) demonstrate how feedforward inhibition and network loops can shape effective interactions.
- Analysis of large random networks shows that hidden neurons can significantly alter the temporal dynamics of measured interactions, sometimes qualitatively changing their behavior.
Limitations and Future Work
- Approximations: The authors acknowledge that their mean-field approximation and linearization of hidden neuron activity are approximations. While they validate these approximations numerically, higher-order terms might become important in certain regimes (e.g., high firing rates, strong coupling).
- Inference challenges: The paper highlights the difficulty of inferring true network structure from effective interactions due to degeneracy.
- Future work: The authors suggest investigating higher-order nonlinear response functions, exploring conductance-based models, and developing more sophisticated methods for inferring network properties from experimental data. They also propose that their framework could be used to identify circuit architectures that perform specific computations, potentially aiding in the treatment of neurological diseases.
Summary updated