NeuNeu Models Overview

• NeuNeu models are a suite of advanced computational frameworks that span scaling law prediction, neural mass reduction, two-compartment spiking, and UAP-invariant feature learning.
• They employ modern architectures such as deep CNNs, Transformers, and reconfiguration networks to process complex neural signals and task trajectories.
• Empirical studies demonstrate that NeuNeu models achieve superior prediction accuracy, biophysical realism, and robustness compared to traditional methods.

The term "NeuNeu Model" applies to several distinct, technically advanced models in machine learning and computational neuroscience, each addressing different domains: scaling law extrapolation for deep learning (Neural Neural Scaling Laws), synaptic mean-field theory at the population level (next-generation neural mass models), and two-compartment neuron spiking models with brain-state specificity. Below is a comprehensive exposition of all principal NeuNeu models as developed in recent research.

1. Neural Neural Scaling Laws (NeuNeu): Time-Series Extrapolation for Task-Specific Scaling

Neural Neural Scaling Laws, abbreviated as "NeuNeu," reframes the prediction of downstream task performance from pretraining trajectories as a data-driven time-series extrapolation problem (Hu et al., 27 Jan 2026). Traditional scaling laws assume functional forms (e.g., power, logistic) relating pretraining validation loss to downstream performance, but these are inadequate to model the full diversity of observed behaviors, including plateaus and inverse scaling.

Architecture and Input Encoding

NeuNeu ingests two primary input streams per checkpoint:

• Token-level validation losses $l_t \in \mathbb{R}^n$, converted to probabilities $p_i = \exp(-\ell_i)$, maintaining the full distribution rather than reducing to mean.
• Histories of downstream task accuracies $y_t \in [0,1]^d$ and compute increments $g_t$.

The model architecture comprises:

• Loss Encoder: A deep CNN with group normalization transforms $p_{1:n}$ to a compact embedding.
• Context Encoder: Projects each historical $(y_i,g_i)$ to a combined context vector.
• Transformer Backbone: Six-layer pre-norm Transformer with self-attention processes the sequence $[\text{CLS};e;c_1;\ldots;c_t]$.
• Quantile Prediction Head: Linear mapping outputs predicted quantiles $(\hat q_{0.1}, ...,\hat q_{0.9})$; the median ($\hat q_{0.5}$) is the primary forecast.

Formal Problem and Training

Given historical checkpoints, NeuNeu predicts future accuracy $y_{t+k}$ for arbitrary horizons without fixed functional assumptions. Training uses quantile (pinball) loss across quantiles, regularized with AdamW optimizer and gradient clipping, and leverages validation trajectories from diverse open-source runs.

Performance and Generalization

NeuNeu achieves significantly lower mean absolute error (2.04%) in predicting future task accuracy across 66 classification tasks compared to logistic scaling laws (3.29% MAE), corresponding to a 38% reduction. NeuNeu generalizes zero-shot to unseen architectures, hyperparameter scales, and downstream tasks and demonstrates calibrated predictive intervals, with the 10–90% prediction interval containing 74.9% of held-out points (Hu et al., 27 Jan 2026).

2. Next-Generation Neural Mass (NeuNeu) Model: Synchrony-Aware Macroscopic Population Dynamics

The next-generation neural mass (NeuNeu) model provides an exact mean-field reduction of networks of $p_i = \exp(-\ell_i)$0-neurons with pulsatile, conductance-based synapses (Coombes et al., 2016). This approach overcomes the phenomenological limitations of classical neural-mass models by deriving macroscopic equations directly from the microscopic dynamics.

Exact Mean-Field Formulation

• Microscopic basis: Population of $p_i = \exp(-\ell_i)$1 $p_i = \exp(-\ell_i)$2-neurons:

$p_i = \exp(-\ell_i)$3

with conductance-based synaptic inputs and Lorentzian-distributed background drive.

• Mean-field reduction: Ott–Antonsen (OA) method yields an exact closure for the distribution, yielding a macroscopic complex order parameter $p_i = \exp(-\ell_i)$4 and the filtered synaptic variable $p_i = \exp(-\ell_i)$5.
• Macroscopic ODEs:

$p_i = \exp(-\ell_i)$6

$p_i = \exp(-\ell_i)$7 is the derived firing-rate function, not a heuristic sigmoid.

Distinguishing Properties

• Synchrony-awareness: $p_i = \exp(-\ell_i)$8 encodes global phase coherence, enabling explicit modeling of event-related synchronisation/desynchronisation—phenomena absent in classical neural-mass models.
• Biophysical validity: All parameters $p_i = \exp(-\ell_i)$9 are physically meaningful.
• Low dimensional: Two real ODEs (for $y_t \in [0,1]^d$0 and $y_t \in [0,1]^d$1) suffice for a single population, making the approach amenable to large-scale whole-brain simulations with interpretable dynamics.

Analytical and Numerical Results

• Precise bifurcation diagrams for Hopf and torus bifurcations, saddle-node-of-periodics, and interpretable transitions in oscillatory regimes.
• Simulation of event-related desynchronization/rebound matches experimental EEG/MEG observations (Coombes et al., 2016).

3. NeuNeu Two-Compartment Neuron Model: State-Specific Apical Amplification, Isolation, and Drive

The NeuNeu (Ca-AdEx) model introduces a biophysically-motivated, two-compartment spiking neuron that can express brain-state-specific regimes of apical-dendritic control: amplification, isolation, and drive (Pastorelli et al., 2023). This architecture models the functional separation of apical and basal dendritic integration in cortical pyramidal neurons.

Structural Composition

• Somatic compartment (basal): Adaptive Exponential Integrate-and-Fire (AdEx) neuron, receives bottom-up inputs.
• Apical compartment (Ca-hotzone): Single-lumped Ca$y_t \in [0,1]^d$2 spiking generator, receives top-down/contextual inputs.
• Inter-compartment coupling: Linear conductance $y_t \in [0,1]^d$3 models electrical transfer between compartments.

Spiking Dynamics and Transfer Functions

Differential equations for $y_t \in [0,1]^d$4 and $y_t \in [0,1]^d$5 (membrane potentials), adaptation variables, and Ca$y_t \in [0,1]^d$6 dynamics are specified for both compartments. The mean firing rate, as a function of input currents $y_t \in [0,1]^d$7, is accurately approximated by the "ThetaPlanes" piece-wise linear transfer function, segmenting the input space into silent, passive, and active (bursting) regions.

Brain-State Modulation

• Apical-amplification (wakefulness): Moderate adaptation, strong coupling, permissive apical leak.
• Apical-isolation (NREM sleep): Strong somatic adaptation, weak coupling, hyperpolarized leaks.
• Apical-drive (REM): Minimal adaptation, strong coupling, hyperpolarized somatic leak, allowing apical-initiated bursts.

Neuromodulatory proxies adjust $y_t \in [0,1]^d$8, $y_t \in [0,1]^d$9, and leak potentials to toggle network regimes, enabling context-dependent credit assignment and continual learning.

Integration and Optimization

• Parameter selection: Genome-level parameter tuning via L2L evolutionary meta-optimization with fitness defined by spiking and firing-rate benchmarks.
• Implications for AI systems: The ThetaPlanes abstraction enables scalable inclusion of state-specific, context-gated computation in spiking and rate-based neural networks, supporting biologically plausible learning protocols and robust credit assignment (Pastorelli et al., 2023).

4. The NEU Meta-Algorithm for Universal UAP-Invariant Feature Learning

The NEU (Non-Euclidean Upgrading) meta-procedure constitutes a universal upgrade to learned feature representations, ensuring that the model class retains the Universal Approximation Property (UAP) under injective, data-adaptive transformations (Kratsios et al., 2018).

Formulation

• UAP-invariance: For input space $g_t$0, a model class $g_t$1 is UAP on a compact set $g_t$2 if for every $g_t$3 and $g_t$4, some $g_t$5 approximates $g_t$6 within $g_t$7 in supremum norm.
• Feature map $g_t$8: Called UAP-invariant if for any model class $g_t$9 with UAP in $p_{1:n}$0, the pullback class $p_{1:n}$1 is UAP in $p_{1:n}$2; this is equivalent to injectivity of $p_{1:n}$3.
• Meta-algorithm:

1. Optimize $p_{1:n}$4 in a reconfiguration network class (injective, homeomorphic transformations) by minimizing a weighted loss over the data.
2. Fit the base model class $p_{1:n}$5 to the transformed data.

Reconfiguration Network

• Composed of "reconfiguration units": locally supported orientation-preserving homeomorphisms, each approximated by a compositional arrangement of neural modules with homeomorphic activations $p_{1:n}$6.
• Universality: Any orientation-preserving homeomorphism (ambient-isotopic to identity) can be approximated on compacts by the network to arbitrary precision.

Theoretical Properties

• NEU-transformed representations are always d-dimensional submanifolds of the ambient feature space.
• Exact memorization of finite datasets while fixing the function outside a small compact is provable, with explicit bounds relating support set measure and parameter count to network depth.

Empirical Results

Across regression and unsupervised learning benchmarks (including financial time series prediction and dimension reduction), NEU-augmented models outperform OLS, kernel ridge, random forests, and deep networks on both standard accuracy and robustness to pathological data structures (Kratsios et al., 2018).

5. Comparative Table of NeuNeu Models

Model	Core Domain	Distinguishing Feature
Neural Neural Scaling Laws	Scaling law prediction	Time-series, transformer-based learning
Next-generation neural mass	Macroscopic neural population	Exact mean-field closure, synchrony-aware
Two-compartment spiking (Ca-AdEx)	Biophysical neuron simulation	Brain-state regime switching; ThetaPlanes
NEU meta-algorithm	Universal feature learning	UAP-invariant reconfiguration networks

Collectively, the NeuNeu family spans neural learning dynamics, biophysical modeling, scaling law prediction, and representation learning, each contributing theoretically rigorous models that bridge empirical performance with structural guarantees.