Across scales, without skipping the evidence

An activity trace, a molecular reporter, and a sequence of choices are observations obtained through particular measurement systems. Their units, resolution, and conditions belong in the description. The word “state” can refer to a classification of those observations or to an underlying process inferred from them.

A latent state is inferred rather than directly observed. Different models can sometimes account for the same measured pattern with different internal variables. Naming one variable “memory,” “arousal,” or “adaptation” does not validate that interpretation.

Ask what the inferred state predicts in observations not used to construct it, whether a simpler account performs comparably, and which intervention could separate competing accounts. Predictive usefulness and causal meaning are related questions that require different evidence.

Hodgkin and Huxley described axonal electrical responses using membrane voltage and time-dependent sodium and potassium conductances. In that framework, an ionic current depends on both conductance and the difference between membrane voltage and the ion’s equilibrium potential. A spike is therefore an outcome of interacting variables, not a direct readout of any one of them.

At a different scale, Hrvatin and colleagues measured light-related transcriptional responses across cell types in mouse visual cortex. Responses differed between cell populations, including non-neuronal cells. This makes a concrete point about measurement: an average tissue-level signal can conceal distinct cellular responses to the same experience.

These measurements answer different questions. A transcript count does not directly supply a membrane conductance, and a firing-rate change does not identify the transcriptional process responsible for it. A proposed bridge should specify the relevant cell type, temporal order, intermediate functional measurement, and an observation that would contradict the account.

Sources: Hodgkin, A. L., & Huxley, A. F.; Hrvatin, S., et al.

In cultured cortical neurons, Turrigiano and colleagues found that reducing inhibition initially increased firing. Over approximately 48 hours, firing returned near control levels while miniature excitatory synaptic responses became smaller. The recovered activity level coexisted with a change in synaptic strength.

This example motivates a distinction central to our research questions. An output can return because the original disturbance disappeared, because another process compensated for it, or because the population being measured changed. Those explanations can agree on the baseline and disagree about a later perturbation.

The next useful experiment would therefore test a response, not just repeat the baseline measurement. Specify an input whose effect is predicted to differ under the competing internal arrangements, while measuring enough of the intermediate process to constrain the explanation.

Source: Turrigiano, G. G., et al.

Churchland and colleagues found organized population dynamics in motor cortex during reaching, despite complex responses of individual neurons. Their analysis illustrates how a population description can expose structure that single-neuron summaries obscure. That result is tied to the recorded system, task, and analysis.

A low-dimensional trajectory is a representation of activity, not a complete inventory of circuit mechanisms. Its interpretation depends on the variables recorded, preprocessing, alignment, and model. Useful structure can motivate an explanation without uniquely determining it.

Musall and colleagues combined neural recordings with detailed measurements of mouse movements during a decision task. Movements that were not required by the task explained substantial neural variation, including signals aligned to trial events. This illustrates why a neural pattern labeled as a decision signal needs to be examined alongside what the subject was actually doing.

Jonas and Kording tested neuroscience-style analyses on a microprocessor whose organization was known. The analyses recovered patterns without reconstructing its information-processing hierarchy. This is a methodological test case, not evidence that brains work like microprocessors: it shows why recovering structure and recovering an explanation are different tasks.

Sources: Churchland, M. M., et al.; Jonas, E., & Kording, K. P.; Musall, S., et al.

A neural signal correlated with a behavioral change may be part of its cause, a consequence, or a response to a shared input. Timing narrows possibilities but does not by itself resolve them. A causal claim needs a design that distinguishes the relevant alternatives.

Interventions can help when their specificity and side effects are addressed. An intervention that changes choice by impairing movement does not isolate a learning computation. Demonstrating a role in a task also leaves open how that role is implemented and when the conclusion generalizes.

Methods contribute most when they resolve different uncertainties. Agreement between measurements can reflect a shared detection bias. A well-controlled intervention can address an alternative left open by correlation, provided its own measurement and specificity problems are examined. Describe what each line of evidence adds.

Species, developmental stage, preparation, and task also define the scope of a finding. Conservation of a molecule can motivate a cross-species hypothesis; it does not establish an identical behavioral function. The bridge itself needs testing.

Computational sleep research, adaptive-behavior experiments in design, and stress-recovery assay development approach this program from different scales. The common target is how experience changes a later response. Establishing a shared mechanism would require direct tests of the connections between them.

Conceptual illustrations can make these questions tangible. Their internal rules are chosen by the designer; biological interpretation requires measured data and validation. Keeping that distinction clear allows a model to be useful without asking it to stand in for an experiment.