Neurochaos Learning: Chaotic Neuron Models

• Neurochaos Learning is a brain-inspired computational framework that replaces conventional neurons with chaotic maps to encode inputs via dynamic trajectories.
• It extracts features like firing time, rate, energy, and symbolic entropy from chaotic iterations, enabling robust classification and regression even with limited data.
• NL architectures bypass backpropagation by using feature-based representations, achieving high accuracy and interpretability in diverse applications including image, biomedical, and sequence analysis.

Neurochaos Learning (NL) is a brain-inspired computational framework in which each artificial “neuron” in a learning architecture is replaced by a low-dimensional chaotic dynamical system—typically a discrete-time chaotic map such as the Generalized Lüroth Series (GLS) or the logistic map. Diverging from classical neural networks, NL architectures encode input features not through nonlinear static activation but via the transient trajectories of individually chaotic neurons, extracting timing- and entropy-based features to drive classification, regression, or other inference tasks. The resulting systems exhibit superior sample efficiency, robust performance in low-data regimes, causal structure preservation, and biologically plausible interpretable coding, providing an alternative path to deep statistical networks for a range of machine learning applications (B et al., 2019, S et al., 2024, Sethi et al., 2022, Narendra et al., 23 Jan 2025, Pant et al., 30 Oct 2025).

1. Mathematical Foundation and Chaotic Neuron Model

Neurochaos Learning replaces conventional artificial neurons by one-dimensional chaotic maps, most commonly the skew-tent GLS map defined as

$$C_b(x) = \begin{cases} \frac{x}{b}, & 0 \leq x < b \ \frac{1-x}{1-b}, & b \leq x < 1 \end{cases} \quad (b \in (0,1),\; x \in [0,1))$$

or, alternatively, the logistic map

$x_{t+1} = r x_t (1-x_t), \quad r \gtrsim 3.57$

Each neuron initializes at state $x_0 = q \in (0,1)$ and repeatedly updates $x_{t+1} = C_b(x_t)$ (or $x_{t+1} = r x_t (1-x_t)$) until the chaotic trajectory enters an $\epsilon$-neighborhood of the input stimulus $s \in [0,1)$: $N = \min \{ t \geq 0 : |x_t - s| < \epsilon \}$ Key properties leveraged are topological transitivity (eventual coverage of any open subset), ergodicity, sensitivity to initial conditions (positive Lyapunov exponent), and dense orbits. These guarantee that every stimulus can be “hit” by chaotic exploration from a single initial state (B et al., 2019, NB et al., 2020, S et al., 2024).

2. Neurochaos Feature Extraction Pipeline

The core NL pipeline consists of input normalization ($x_j \to [0,1]$), followed by independent per-feature chaotic embedding and extraction of summary statistics from the resulting “neural traces.” For each neuron and input coordinate, characteristic features are:

• Firing time $N$: steps to hit the input under chaotic iteration
• Firing rate $R$: fraction of iterates above the discrimination threshold $b$ ($R = (1/N) \sum_{k=1}^N \mathbb{I}[x_k \geq b]$)
• Energy $E = \sum_{k=1}^N (x_k)^2$
• Symbolic (Shannon) entropy of the binary sequence $s_k = \mathbb{I}[x_k \geq b]$, $H = -p_0\log_2 p_0 - p_1\log_2 p_1$, with $p_i$ the empirical fraction of zeros/ones.

These are concatenated for each coordinate to yield a feature vector of dimension $4n$, serving as input to a decision module (cosine similarity classifier, SVM, or any ML model) (Sethi et al., 2022, NB et al., 2020, S et al., 2024, Narendra et al., 23 Jan 2025).

3. Learning Algorithms, Architectures, and Variants

NL is not based on backpropagation but on feature-based representation. Core variants include:

• ChaosNet: simple centroid classifier in the chaotic-firing-time space; per-class mean vectors are computed and test samples classified via cosine similarity (Balakrishnan et al., 2019, B et al., 2019).
• ChaosFEX+ML: extracted chaotic features fed to standard classifiers (SVM, RF, kNN, AdaBoost) (S et al., 2024, Sethi et al., 2022).
• Random Heterogeneous Neurochaos Learning (RHNL): stochastic per-feature assignment of chaotic neuron types (GLS or logistic) to mimic randomness and heterogeneity in biological neural tissue, improving class-separability and robustness. Neuron type is specified by a probability vector and randomly permuted assignment (S et al., 2024, Pant et al., 30 Oct 2025).
• Hyperparameter-free Neurochaos Learning: AutochaosNet eliminates per-sample parameter tuning by replacing chaotic traces with segments of a universal chaotic sequence (e.g., from the Champernowne constant via the decimal shift map), and extracting trace statistics for forming input features (Henry et al., 2 Aug 2025).
• Semi-supervised NL (NL+ST): integration of chaos-based features with self-training (pseudo-labeling high-confidence unlabeled points), yielding large gains in low-label regimes (M et al., 3 Jan 2026).

The universal approximation theorem for single-layer Neurochaos architectures holds: a layer of $L$ chaotic neurons suffices to approximate any function $f:\{1, \ldots, L\} \to \mathbb{R}$ with arbitrary accuracy (NB et al., 2020). NL can also act as a feature augmenter (e.g., tracemean features) for regression models (linear, ridge, lasso, SVR), boosting performance and statistical efficiency (Henry et al., 19 May 2025).

4. Empirical Benchmarks and Performance Characterization

NL architectures demonstrate marked superiority, especially in few-shot, noisy, or imbalanced data regimes:

• On MNIST, with as few as 7 training samples per class (≈0.1% of the data), NL matches or outperforms shallow and classical ML baselines, over both accuracy and sample-efficiency metrics (B et al., 2019).
• On the Iris dataset, with only 2 training samples per class, NL achieves 95.8% accuracy—higher than SVM, kNN, or Decision Tree classifiers (B et al., 2019).
• On UCI and real-world biomedical datasets, NL-based feature transformations boost Macro-F1 by up to 25% (e.g., Statlog (Heart)), and up to 144% in the low-shot regime (Haberman’s Survival, FSDD digits) (Sethi et al., 2022).
• RHNL shows F1=1.0 on Wine, 0.99 on Breast Cancer and Banknote datasets, and superiority in low-sample regimes (MRI brain tumor classification) over all stand-alone methods (S et al., 2024).
• For classification of genome sequences (e.g., SARS-CoV-2 vs. SARS-CoV-1), ChaosFEX+SVM achieves macro F1>0.99 with only one training example per class over 1,000 random trials (NB et al., 2020).
• In forest fire detection, RHNL achieves F1=1.0 on Algerian Forest Fires dataset and high precision on Canadian and Portugal datasets, outperforming state-of-the-art ML baselines in most regimes (Pant et al., 30 Oct 2025).

5. Theoretical and Dynamical Properties: Chaos, Causality, and Adaptivity

NL’s effectiveness stems from deep dynamical-systems principles:

• Dynamical mixing and ergodicity: chaotic state evolution ensures trajectories cover the stimulus space and generate feature diversity (B et al., 2019, Sethi et al., 2022).
• Topological transitivity: guarantees any input can be mapped arbitrarily closely from any initial state, a property central to the expressivity of NL (B et al., 2019, NB et al., 2020).
• Stochastic resonance: at intermediate noise levels, classification or detection performance is maximized (peaked F1 as noise varies), reflecting resonance between intrinsic chaos and ambient noise (NB et al., 2021).
• Causality preservation: NL-based features preserve Granger and Compression-Complexity causality under transformations where deep nets destroy it, making NL robust to causal inference tasks (Narendra et al., 23 Jan 2025, B et al., 2022).
• Interpretability: Each chaotic feature is a transparent function of the input, enabling explicit attribution (Pant et al., 30 Oct 2025).
• Robustness: NL feature spaces are resistant to additive parameter noise and generalize under data and parameter drift (Balakrishnan et al., 2019).
• Exploration-exploitation balance: At the “edge of chaos” (maximum Lyapunov exponent ≈ 0), NL achieves optimal balance between sensitivity (adaptivity) and memory (stability), accelerating learning in nonstationary environments (Benati et al., 15 Jun 2025, Jiménez-González et al., 10 Jun 2025).

6. Extensions: Hybridization, Hardware, and Structured Data

Recent work extends NL in several directions:

• Graph neural chaos: Adapting Graph Neural Networks by replacing or augmenting aggregate/combine steps with chaotic updates, yielding NL-GNN hybrids for causal and relational data (Narendra et al., 23 Jan 2025).
• Neuromorphic NL: Implementation of NL on nanowire–memristive physical neural circuits, leveraging the intrinsic dynamics of nonlinear elements for energy-efficient real-time chaotic processing (Xu et al., 12 Jun 2025).
• Deepened and coupled chaos: Multi-layer and coupled RHNL stacks, with heterogeneous neuron types and sparse inter-layer projections, align more closely with biological circuit architectures and may further boost performance (S et al., 2024, Pant et al., 30 Oct 2025).
• Parameter-free variants: Universal sequence–based feature architectures (AutochaosNet) offer scalable, hyperparameter-free alternatives for real-world applications (Henry et al., 2 Aug 2025).
• Continual and regime-shift learning: Lyapunov-regularized training at marginal stability attains rapid adaptation and resilience under abrupt regime shifts (e.g., in chaotic time series prediction) (Benati et al., 15 Jun 2025).

7. Open Challenges and Future Research Directions

While Neurochaos Learning demonstrates strong performance and theoretical justification, several open questions and research directions remain:

• Systematic hyperparameter selection: Efficient, domain-agnostic procedures for choosing GLS parameters ($q, b, \epsilon$) or adapting them during training (B et al., 2019, Sethi et al., 2022).
• End-to-end differentiable chaotic modules: Backpropagation through differentiable chaos for deeper architectures (S et al., 2024).
• Hardware realization: Neuromorphic chips implementing piecewise-linear maps, enabling ultra-low-power edge computing (Xu et al., 12 Jun 2025, Pant et al., 30 Oct 2025).
• Universal orbits under other chaotic maps: Identifying and formalizing “normal” constants for various maps to further generalize hyperparameter-free NL (Henry et al., 2 Aug 2025).
• Structured data and causality: Formulating chaotic embedding modules for graph, sequence, and multimodal causal data; integrating NL with causal graph neural architectures (Narendra et al., 23 Jan 2025).
• Theoretical analysis: Rigorous bounds on the generalization error of NL, particularly under composition, stacking, and as data dimension grows (NB et al., 2020).

NL thus represents a deeply dynamical, brain-inspired learning paradigm, unifying symbolic, kinetic, and statistical processing—yielding neural systems robust to data scarcity, non-stationarity, and causal structure, and tractable for direct hardware deployment in neuromorphic systems.