A theory of neural emulators

TL;DR

The work introduces emulator theory (ET) and neural emulators as a scale- and mechanism-agnostic predictive framework for modeling neural dynamics and behavior, potentially extending to conscious states. It formalizes predictive emulators via joint time-series modeling with $E_t^b(w_{j,t,g})$ and $E_t^c(w_{j,t,g})$, where neural states evolve as $w_{j,t,g}=R(w_{j,t-1,g}, s_{t-1})$, and posits that sufficiently accurate emulators can reproduce source capacities without explicit mechanistic explanations. The authors propose the path-independent neural causality (PINC) framework, including ISI, PIC, and MSD, to justify exogenous generation of neural states and indistinguishability from ground-truth models, while distinguishing ideal versus scalable emulators. Limitations are discussed, notably the lack of content specification for conscious states and the need for empirical validation, data requirements, and architectural choices. Overall, ET offers a principled, prediction-centered paradigm for neuroscience that may unify cross-scale modeling and motivate new experiments on neural causality and consciousness in both biological and artificial systems.

Abstract

A central goal in neuroscience is to provide explanations for how animal nervous systems can generate actions and cognitive states such as consciousness while artificial intelligence (AI) and machine learning (ML) seek to provide models that are increasingly better at prediction. Despite many decades of research we have made limited progress on providing neuroscience explanations yet there is an increased use of AI and ML methods in neuroscience for prediction of behavior and even cognitive states. Here we propose emulator theory (ET) and neural emulators as circuit- and scale-independent predictive models of biological brain activity and emulator theory (ET) as an alternative research paradigm in neuroscience. ET proposes that predictive models trained solely on neural dynamics and behaviors can generate functionally indistinguishable systems from their sources. That is, compared to the biological organisms which they model, emulators may achieve indistinguishable behavior and cognitive states - including consciousness - without any mechanistic explanations. We posit ET via several conjectures, discuss the nature of endogenous and exogenous activation of neural circuits, and discuss neural causality of phenomenal states. ET provides the conceptual and empirical framework for prediction-based models of neural dynamics and behavior without explicit representations of idiosyncratically evolved nervous systems.

Figures (6)

• Figure 1: Circuit- and scale-agnostic neural emulators. (a). Recording from a rodent brain (light-blue) based on parcelation (cubes) of increasing granularity (red-hue diamonds). (b). Neural time series from parcellation in (a) (light blue) and behavior time series (green). (c). Emulators learn joint probability of time-series in (b). (d). Behavior output of emulators is increasingly similar to biological behavior as parcellation granularity increases.
• Figure 2: Granularity-based neural emulators. Relationship between the capacity (or accuracy) of a behavior (blue curve) and neural (red curve) emulator vs the granularity of the neural data used for training with hypothesized requirements for perfect behavior models (dashed blue line) and conscious states (dashed red line). Proposed experiment of recording sparsely sampled brain-wide LFP (magenta arrow) and putative emulator capacity from such datasets (magenta dots).
• Figure 3: Input source indistinguishability (ISI). (a) A two neuron local network receiving "native" sensory and internal state input and generating downstream output. (b) Spike rasters for network in (a). (c) Same as (a) but inputs to network are simulated by an external system. (d) Same as (a) but for simulated system as in (c) results in identical spike rasters. (e) Left: single neuron receiving axonal inputs (red) from various sources and outputting a spike pattern (green); Right: proposed spike rasters for (e). (f) Left: same as (e) but axonal inputs are simulated by external system; Right: identical raster to (e) achieved by simulated system in (f). (g) Left: a self-driving recurrent neural network; Right: proposed spike rasters for (g). (h) Left: same as (g) but with recurrent inputs removed and simulated inputs; Right: identical spike raster as in (g) achieved by simulated system in (h).
• Figure 4: Path-independent causality conjecture (PIC). (a) Causal graph of two necessary and sufficient biophysical neural states, S$_1$ and S$_2$, required for phenomenal conscious state C. In the native state, neural states S$_1$ and S$_2$ are caused by endogenous prior states P$_1$ and P$_2$ (e.g. sensory and internal state processing). (b) Same as (a) but neural states S$_1$ and S$_2$ are generated via exogenous (e.g. artificially generated) inputs resulting in the same phenomenal cognitive state C irrespective of the causal path of activation of necessary and sufficient biophysical neural states.
• Figure 5: Different depths of models. Relative to David Marr's marr1982vision three levels of modeling (green rectangle), emulators model all levels of a biophysical system, while simulators model only behavior.