The Emergence Machine: Hellinger-Based Drift Monitoring and Input Characterization

A Regulatory Enactive Architecture for Regime Detection and Adaptation in Non-Stationary Streaming Data

-Nicholas Davis, PhD


Abstract

Contemporary machine learning systems are predominantly designed under assumptions of stationarity, optimizing predictive accuracy within stable data regimes. However, real-world environments—such as financial markets, physiological signals, and human–AI interaction—are inherently non-stationary, exhibiting continuous structural drift. In such contexts, prediction-centric approaches degrade due to their inability to detect and respond to regime shifts in real time.

We introduce the Emergence Machine (EM), a regulatory enactive architecture for adaptive sense-making in streaming environments. Central to EM is a mechanism we term Hellinger-based Drift Monitoring and Input Characterization (HDMIC), which operates not as a clustering algorithm but as a distribution-sensitive drift detection system over sliding temporal windows. HDMIC treats drift as a first-class informational signal, enabling the system to differentiate regimes and regulate its adaptive behavior.

Rather than minimizing prediction error, EM performs a continuous perception–interpretation–regulation–adaptation loop. This allows the system to maintain behavioral coherence under structural drift by dynamically adjusting sensitivity and attractor commitment. We demonstrate how EM reframes learning as regulated interaction with a changing environment, offering a new paradigm for adaptive intelligence in non-stationary domains.


1. Introduction

Most machine learning systems are developed under the assumption that data distributions are stationary or slowly varying. Under this assumption, learning is framed as the optimization of predictive models over a fixed or gradually evolving distribution, typically through error minimization or likelihood maximization (Bishop, 2006; Goodfellow et al., 2016). While effective in controlled settings, this paradigm becomes fragile in real-world environments where the statistical properties of data shift over time.

Such environments—including financial markets, physiological signals such as EEG, and interactive human–AI systems—are inherently non-stationary. They exhibit structural drift, wherein the underlying generative processes evolve, sometimes gradually and sometimes abruptly, leading to changes in distributional form, variance structure, or temporal dependencies (Gama et al., 2014; Webb et al., 2016). In these contexts, models trained under stationarity assumptions degrade, not necessarily because of noise, but because their foundational assumptions no longer hold.

Traditional approaches to non-stationarity typically treat drift as a problem to be mitigated. Techniques such as concept drift detection, adaptive retraining, and ensemble updating aim to restore predictive accuracy by minimizing the impact of distributional change (Lu et al., 2018). However, these approaches retain a fundamentally prediction-centric view: drift is treated as deviation from a target distribution, and the goal remains convergence toward accurate prediction.

In contrast, we argue that this framing is insufficient for continuously evolving environments. We propose a conceptual shift:

Drift is not error—it is an informational signal indicating structural change in the environment.

This perspective aligns with enactive approaches to cognition, which emphasize that intelligent systems do not passively model a static world, but actively regulate their coupling with a changing environment through ongoing interaction (Varela et al., 1991; Thompson, 2007). From this standpoint, the central challenge is not prediction under stability, but maintaining coherence under change.

To operationalize this perspective, we introduce the Emergence Machine (EM), a regulatory architecture grounded in Enactive Regulation Theory (ERT). EM reframes learning as a continuous process of perception, interpretation, regulation, and adaptation in response to structural drift. At the core of this architecture is Hellinger-based Drift Monitoring and Input Characterization (HDMIC), a mechanism that computes distributional divergence over sliding temporal windows using Hellinger distance (Hellinger, 1909; Le Cam, 1986). Rather than minimizing this divergence, EM treats it as a first-class signal that informs system behavior.

This leads to a key departure from classical approaches. EM does not perform clustering in the traditional sense, nor does it optimize a predictive model over a fixed hypothesis space. Instead, it performs online regime differentiation and regulatory adaptation, dynamically adjusting its behavior in response to detected structural changes. In doing so, EM enables adaptive sense-making in non-stationary streaming data, supporting sustained interaction with environments that cannot be reduced to stable distributions.

This paper makes three contributions: (1) a reframing of drift as an informational substrate, (2) the HDMIC mechanism for distributional drift monitoring, and (3) the Emergence Machine architecture as a regulatory enactive system.

2. Background and Motivation

2.1 Limitations of Prediction-Centric Systems

Contemporary machine learning systems are predominantly organized around a prediction-centric paradigm, in which learning is framed as the optimization of predictive accuracy under assumed distributional stability. This paradigm underlies a wide range of approaches, including supervised learning, probabilistic modeling, and deep neural networks (Bishop, 2006; Goodfellow et al., 2016).

Within this framework, systems typically:

While effective in controlled or well-behaved environments, these assumptions break down in non-stationary settings. In particular, prediction-centric systems exhibit systematic limitations when faced with abrupt or structural regime changes, where the underlying data-generating process shifts in ways not captured by prior training.

A substantial body of work in concept drift has documented these challenges, showing that model performance degrades when distributional assumptions are violated, often requiring retraining, ensemble adaptation, or explicit drift detection mechanisms (Gama et al., 2014; Lu et al., 2018). However, these approaches typically remain within the prediction-centric paradigm: drift is treated as a perturbation to be corrected, rather than a fundamental property of the environment.

As a result, such systems often lack mechanisms to:

This leads to a core limitation: prediction-centric systems are designed to converge under stability, but struggle to remain coherent under ongoing change.

2.2 Drift as Structural Signal

To address these limitations, we introduce a reframing of drift as a first-class property of dynamic environments.

We define structural drift as:

A change in the underlying distributional properties of the data-generating process over time.

This definition encompasses:

In contrast to traditional approaches, which treat drift as noise or error, the Emergence Machine (EM) treats drift as:

This perspective aligns with work in adaptive systems and non-stationary learning that emphasizes the importance of detecting and characterizing distributional change (Widmer & Kubat, 1996; Žliobaitė, 2010). However, EM extends this view by assigning drift a functional role in system organization, rather than treating it solely as a condition to be mitigated.

In this sense, drift becomes an informational substrate: it provides ongoing feedback about how the environment is evolving, enabling the system to adjust its behavior in relation to that change.

2.3 Enactive Regulation Perspective

The conceptual foundation of EM is grounded in an enactive approach to cognition, which emphasizes that intelligence arises through active engagement with the environment rather than passive representation (Varela et al., 1991; Thompson, 2007).

From this perspective:

Enactive theories propose that agents maintain viability by continuously adjusting their behavior in response to changing conditions, rather than by optimizing a static internal representation. This stands in contrast to predictive processing frameworks, which prioritize error minimization as the central organizing principle (Friston, 2010).

Within this framework, the key challenge is not to eliminate discrepancy, but to:

maintain coherent interaction under conditions of continuous change.

The Emergence Machine operationalizes this perspective computationally. By treating drift as an informational signal and embedding it within a regulatory loop, EM shifts the role of learning from:

This reframing enables a system that does not merely adapt its parameters, but actively reorganizes its behavior in response to evolving environmental structure.

3. Hellinger-Based Drift Monitoring and Input Characterization (HDMIC)

3.1 Overview

Hellinger-Based Drift Monitoring and Input Characterization (HDMIC) is the core perceptual mechanism of the Emergence Machine (EM). Despite its name, HDMIC does not perform clustering in the conventional sense. It does not partition the input space into fixed groups, nor does it minimize intra-cluster variance or optimize a clustering objective.

Instead, HDMIC functions as:

a streaming, distribution-sensitive drift detection mechanism that differentiates regimes and modulates adaptive behavior in real time.

This distinction is critical. While prior work in data stream mining has explored distributional change detection and concept drift (Gama et al., 2014; Webb et al., 2016), such approaches are typically used to trigger model updates or retraining. In contrast, HDMIC embeds drift detection directly within a regulatory loop, where distributional divergence continuously informs system behavior.

Thus, HDMIC is not a preprocessing step or auxiliary detector, but a primary perceptual interface through which the system senses and interprets environmental change.

3.2 Windowed Distribution Comparison

At each time step ttt, HDMIC compares two temporally localized distributions:

Probability distributions are estimated over these windows using empirical histograms or kernel density approximations, depending on the data modality.

The divergence between these distributions is computed using the Hellinger distance.

Hellinger distance is a symmetric, bounded measure of divergence between probability distributions, with desirable properties for streaming contexts, including robustness to support mismatch and stability under small sample variations (Hellinger, 1909; Le Cam, 1986).

This computation yields a scalar signal:

a measure of how different the current distribution (“now”) is from the reference distribution (“before”).

Unlike many divergence-based methods that use such measures for hypothesis testing or model selection, HDMIC treats this value as a continuous perceptual signal rather than a binary decision criterion.

3.3 Drift as a First-Class Signal

The computed Hellinger distance is interpreted as a measure of structural drift and categorized along a continuum:

Crucially, this signal is not minimized. In contrast to optimization-based frameworks, where divergence is treated as error to be reduced, HDMIC treats divergence as:

informational evidence of environmental change.

This represents a fundamental shift in how distributional distance is used. Rather than driving convergence, the drift signal informs the system about:

In this sense, HDMIC transforms a statistical distance metric into a functional signal for adaptive sense-making.

3.4 Regime Differentiation (Not Clustering)

HDMIC does not assign data points to clusters or maintain explicit cluster memberships. Instead, it enables implicit regime differentiation through continuous monitoring of distributional divergence.

Specifically, the system distinguishes between:

This process can be understood as:

online regime boundary detection via distributional divergence

Unlike traditional clustering, which partitions data into static groups, HDMIC tracks the temporal evolution of distributions, allowing regimes to emerge, dissolve, and reconfigure over time.

This aligns with prior work on concept drift characterization (Webb et al., 2016), but extends it by embedding regime differentiation within an ongoing regulatory process rather than treating it as a discrete event.

3.5 Regulatory Triggering

The most critical function of HDMIC is its role in driving system regulation. The drift signal modulates multiple aspects of system behavior, including:

Through these mechanisms, HDMIC acts as a control signal that dynamically reshapes the system’s response to incoming data.

Thus:

HDMIC functions as the perception layer for regulation, transforming distributional divergence into actionable control signals.

This distinguishes EM from both:

Instead, EM uses HDMIC to maintain coherent adaptive behavior under continuous structural change.

4. The Emergence Machine Architecture

4.1 Overview

The Emergence Machine (EM) is organized as a continuous, closed-loop process that integrates perception, interpretation, regulation, and adaptation:

Perception → Interpretation → Regulation → Adaptation

This loop defines EM as a dynamical system rather than a static predictive model. Unlike conventional machine learning pipelines—where perception (feature extraction), learning, and inference are often separated—EM maintains an ongoing coupling between incoming data and system behavior.

This architecture is inspired by enactive and dynamical systems approaches to cognition, in which intelligent behavior arises from continuous interaction between an agent and its environment, rather than from internal model optimization alone (Varela et al., 1991; Beer, 2000). Similar closed-loop formulations have been explored in adaptive control and embodied AI, where perception and action are co-constitutive (Sutton & Barto, 2018; Clark, 2016). However, EM differs in that drift detection is not auxiliary but central, driving the entire loop.

The result is a system that does not converge to a fixed model, but instead maintains coherence through ongoing regulatory adjustment under structural drift.

4.2 Perception

The perceptual component of EM is implemented through HDMIC, which computes distributional divergence between recent and reference windows (Section 3).

At each time step, HDMIC produces:

This signal serves as the system’s primary interface with the environment. Unlike traditional feature extraction, which aims to represent input data, perception in EM is oriented toward:

detecting how the environment is changing over time.

This aligns with perspectives in adaptive systems and signal processing that treat change detection as a fundamental perceptual task (Basseville & Nikiforov, 1993). However, EM extends this by embedding drift perception directly into a regulatory architecture, rather than using it solely for anomaly detection or alerting.

4.3 Interpretation

The interpretation stage transforms raw drift signals into contextualized assessments of environmental change. Rather than treating all divergence uniformly, EM differentiates between qualitatively distinct types of drift:

This categorization reflects distinctions established in concept drift literature, where drift can be abrupt, incremental, or recurring (Gama et al., 2014; Webb et al., 2016). However, in EM, interpretation is not used to select a corrective action (e.g., retraining), but to modulate regulatory dynamics.

Thus, interpretation serves to:

This step is critical for preventing overreaction to noise while enabling rapid response to genuine regime shifts.

4.4 Regulation

Regulation is the core organizing principle of EM. Based on interpreted drift signals, the system dynamically adjusts its internal parameters and behavioral stance in relation to the environment.

Key regulatory mechanisms include:

These mechanisms collectively implement drift-aware control, in which the system continuously balances:

4.5 Adaptation

The adaptation stage implements the consequences of regulatory adjustments, updating the system’s internal organization and external behavior in response to structural drift. Rather than performing parameter updates within a fixed model, adaptation in EM involves the reorganization of system dynamics to maintain coherence with evolving environmental structure.

Specifically, EM updates regime representations, attractor structures, and behavioral output, as described below.

Regime Representations

Regime representations are internal encodings of the current environmental condition, constructed from recent distributional structure and its relation to prior states. These representations are not static models or discrete labels, but temporally grounded summaries of system–environment coupling.

A regime representation captures:

Importantly, regimes are not predefined categories. They are emergent and continuously updated, allowing the system to track:

Thus, regime representations function as:

dynamic situational awareness structures that anchor interpretation and guide regulation.

Attractor Structures

Attractor structures are dynamically stable patterns of system behavior that emerge under particular regime conditions. Rather than being explicitly stored representations, attractors are recurrent configurations of interaction toward which the system tends to stabilize.

An attractor in EM is characterized by:

Attractors can be understood as:

Crucially:

Thus, attractors provide:

a mechanism for continuity without rigidity, enabling the system to remain stable without becoming fixed.

Behavioral Outputs

Behavioral outputs are the externally observable actions or decisions generated by the system in response to the current regime and its associated attractor dynamics.

These outputs are not produced by a static policy or predictive model. Instead, they are the result of:

Examples include:

Behavioral outputs are therefore:

situated actions arising from the system’s ongoing coupling with the environment, rather than predictions generated from a fixed model.

Importantly, prediction—while still present—is secondary to maintaining coherent adaptation. Rather than optimizing predictive accuracy, EM prioritizes:

sustained alignment between system behavior and evolving environmental structure.

This distinction is critical. In predictive and reinforcement learning systems, performance is typically evaluated in terms of error minimization or reward maximization (Sutton & Barto, 2018). In contrast, EM evaluates success in terms of:

Thus, the primary objective of EM is not optimality under fixed conditions, but:

the maintenance of viable, adaptive behavior under continuous structural drift.

5. Conceptual Distinction from Classical Approaches

The Emergence Machine (EM) departs fundamentally from dominant paradigms in machine learning and data stream analysis. While it shares surface-level similarities with clustering, prediction, and drift detection methods, its underlying principles and operational logic are distinct. This section clarifies these differences by situating EM relative to classical approaches and articulating its regulatory paradigm.

5.1 Not Clustering

At a superficial level, EM may appear related to clustering methods, as it differentiates regimes and identifies structural variation in data. However, EM does not perform clustering in the conventional sense.

Classical clustering algorithms—such as k-means, Gaussian mixture models, and density-based methods—aim to:

These approaches assume that the underlying structure of the data can be represented as a set of relatively stable groupings.

In contrast, EM:

Instead, EM performs continuous regime differentiation based on distributional divergence over time. Regimes are not static groupings, but temporally evolving structures that emerge, transform, and dissolve as the environment changes.

Thus, rather than clustering data, EM:

tracks the evolution of distributions and detects boundaries between regimes as they unfold in time.

This distinction is critical. EM replaces the notion of static structure with dynamic, temporally situated organization, aligning more closely with streaming and non-stationary perspectives than with traditional clustering frameworks.

5.2 Not Prediction-Centric

EM also departs from the dominant prediction-centric paradigm in machine learning. Most contemporary systems are designed to:

Even in non-stationary settings, approaches such as online learning and concept drift adaptation typically retain prediction as the primary objective, updating models to maintain accuracy over time (Gama et al., 2014; Lu et al., 2018).

In contrast, EM:

While EM may produce predictions or decisions as part of its behavior, these outputs are secondary to its primary function: maintaining coherent interaction with a changing environment.

This positions EM outside both:

Instead, EM operates according to a different organizing principle:

adaptation through regulation under drift, rather than optimization under stability.

5.3 Regulatory Paradigm

The core contribution of EM is its formulation of a regulatory paradigm for adaptive systems. Rather than minimizing error or fitting models to data, EM performs a continuous loop:

Detection of divergence → interpretation of structural change → regulation of behavior → maintenance of coherence

This loop redefines the role of learning in non-stationary environments. Key characteristics of this paradigm include:

This perspective aligns with enactive and dynamical systems approaches to cognition, which emphasize ongoing coupling and adaptive regulation (Varela et al., 1991; Beer, 2000). It also resonates with work in adaptive control, where stability is achieved through continuous feedback rather than static optimization (Åström & Murray, 2008).

However, EM extends these traditions by grounding regulation explicitly in distributional drift, using statistical divergence as a primary perceptual signal.

In this sense, EM represents a shift from:

6. Applications

The Emergence Machine (EM) is designed for environments characterized by continuous structural drift, where the underlying data-generating processes evolve over time. Its regulatory architecture enables adaptive behavior in domains where traditional prediction-centric systems struggle. Below, we outline key application areas where EM’s approach is particularly well-suited.

6.1 Financial Markets

Financial markets are prototypical non-stationary systems, exhibiting regime shifts driven by macroeconomic conditions, liquidity changes, and behavioral dynamics. Price processes are influenced by evolving participant strategies, structural breaks, and feedback effects, leading to time-varying distributions and volatility clustering (Cont, 2001).

Traditional predictive models often degrade in such environments due to their reliance on stationarity assumptions or fixed feature representations. While adaptive methods exist, they typically respond reactively to performance degradation.

In contrast, EM:

This enables a shift from:

predicting price movements
 to
 maintaining coherent trading behavior under evolving market conditions

Such an approach aligns with the need for online adaptation in algorithmic trading and financial signal processing, where robustness to regime change is critical.

6.2 EEG and Brain–Computer Interfaces (BCI)

EEG and BCI systems operate on highly dynamic neural signals that vary across time, tasks, and individuals. These signals are non-stationary due to factors such as cognitive state changes, fatigue, electrode drift, and environmental noise (Makeig et al., 2004).

Standard approaches often require:

EM offers an alternative by:

This enables:

adaptive interpretation of neural signals in real time,
 supporting more robust and flexible BCI systems.

Such capabilities are particularly important for long-duration or real-world BCI deployments, where static models are insufficient.

6.3 Human–AI Interaction and Co-Creative Systems

In interactive and co-creative contexts, the “data distribution” is not external but co-constructed through interaction. Human–AI systems engaged in creative tasks—such as drawing, music, or collaborative design—exhibit dynamic trajectories shaped by both participants.

These environments are characterized by:

Traditional AI systems, which generate outputs independently or respond reactively, often fail to sustain meaningful interaction over time.

EM enables:

This supports:

participatory sense-making, where AI systems do not merely generate outputs, but actively maintain coherence within an evolving interaction (Davis et al., 2017).

As such, EM provides a computational foundation for co-creative AI systems that adapt to human partners in real time.

6.4 Complex Adaptive Systems

Beyond specific domains, EM is applicable to a broad class of complex adaptive systems characterized by:

Examples include:

In these contexts, the ability to:

is essential for maintaining system viability.

EM’s architecture provides a general-purpose framework for:

adaptive sense-making in environments where structure is not fixed, but continuously emerging.

7. Discussion

7.1 Drift as Informational Substrate

A central contribution of the Emergence Machine (EM) is the reframing of drift from a source of error to an informational substrate for adaptive systems. In conventional machine learning, distributional divergence is typically interpreted as degradation—an indication that a model is no longer well-calibrated to its data. This framing leads to reactive strategies such as retraining, model replacement, or ensemble updating (Gama et al., 2014; Lu et al., 2018).

In contrast, EM treats drift as a continuous signal of environmental change, enabling the system to remain responsive to evolving structure rather than attempting to suppress or eliminate it.

This reframing has several important implications:

This perspective aligns with broader trends in adaptive systems and online learning that emphasize the importance of change detection (Basseville & Nikiforov, 1993), but extends them by assigning drift a functional role in system organization. Rather than being an external condition to manage, drift becomes an internal driver of system dynamics.

7.2 Regulation Over Optimization

EM advances a shift from optimization-based to regulation-based approaches to intelligent systems.

Traditional machine learning systems are organized around optimization:

While effective in stable environments, optimization inherently assumes that a fixed objective can be defined and pursued. In non-stationary contexts, however, such objectives become unstable or ill-defined, as the environment itself is changing.

EM instead embodies the principle:

Optimization converges. Regulation sustains.

In this framework:

Regulation operates through continuous feedback, adjusting system behavior in response to environmental signals rather than pursuing a fixed target. This aligns with principles from control theory, where stability is achieved through feedback loops rather than static solutions (Åström & Murray, 2008).

Importantly, EM does not eliminate optimization entirely. Instead, it subordinates optimization to regulation. Predictive accuracy and performance remain relevant, but they are not the primary organizing principles. Instead, they are embedded within a broader process of maintaining coherent interaction with the environment.

This shift has implications for the design of adaptive systems:


7.3 Toward Enactive AI Systems

EM represents a step toward a broader class of enactive AI systems, which are defined by their capacity to actively participate in and co-evolve with their environments.

In traditional AI paradigms, systems are typically designed to:

In contrast, enactive approaches emphasize that cognition arises through interaction, not representation alone (Varela et al., 1991; Thompson, 2007). From this perspective, intelligent systems are not detached observers, but participants in dynamic environments.

EM operationalizes this perspective by:

This results in systems that:

Such systems are particularly relevant for domains involving sustained interaction, including human–AI collaboration, creative systems, and real-time decision-making.

More broadly, EM contributes to a shift in how AI systems are conceptualized:

from model-based intelligence → to → interaction-based intelligence

This shift opens new directions for research, including:

8. Conclusion

We introduced the Emergence Machine (EM), a regulatory enactive architecture for adaptive behavior in non-stationary environments. Central to EM is Hellinger-Based Drift Monitoring and Input Characterization (HDMIC), which transforms distributional divergence into a continuous perceptual signal. This signal enables the system to differentiate regimes and dynamically regulate its behavior in response to structural change.

Unlike traditional machine learning systems that assume stability and optimize predictive accuracy, EM is designed for environments in which change is persistent and fundamental. By embedding drift detection within a closed-loop process of perception, interpretation, regulation, and adaptation, EM maintains coherent interaction with evolving environments without relying on retraining or static model assumptions.

This work advances a conceptual shift in how learning systems are understood. Rather than treating learning as the optimization of predictive models under assumed stability, we propose:

learning as a process of maintaining coherence under structural drift.

This reframing has implications beyond the specific architecture presented here. It suggests that intelligent systems operating in real-world environments should be designed not for convergence to fixed solutions, but for ongoing viability under continuous change.

More broadly, EM contributes to the development of enactive AI systems that participate in and co-adapt with their environments. Future work may extend this framework by incorporating richer forms of interaction, developing new evaluation metrics for adaptive coherence, and exploring applications in domains where non-stationarity is intrinsic.


References

Åström, K. J., & Murray, R. M. (2008). Feedback Systems: An Introduction for Scientists and Engineers. Princeton University Press.

Basseville, M., & Nikiforov, I. V. (1993). Detection of Abrupt Changes: Theory and Application. Prentice Hall.

Beer, R. D. (2000). Dynamical approaches to cognitive science. Trends in Cognitive Sciences, 4(3), 91–99.

Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.

Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.

Cont, R. (2001). Empirical properties of asset returns: Stylized facts and statistical issues. Quantitative Finance, 1(2), 223–236.

Davis, N., Hsiao, C.-P., Singh, K., Li, L., & Magerko, B. (2017). Empirically studying participatory sense-making in abstract drawing with a co-creative cognitive agent. Proceedings of the 8th International Conference on Computational Creativity (ICCC).

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M., & Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4), 44.

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.

Hellinger, E. (1909). Neue Begründung der Theorie quadratischer Formen von unendlich vielen Veränderlichen. Journal für die reine und angewandte Mathematik.

Jain, A. K. (2010). Data clustering: 50 years beyond k-means. Pattern Recognition Letters, 31(8), 651–666.

Le Cam, L. (1986). Asymptotic Methods in Statistical Decision Theory. Springer.

Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., & Zhang, G. (2018). Learning under concept drift: A review. IEEE Transactions on Knowledge and Data Engineering, 31(12), 2346–2363.

Makeig, S., Debener, S., Onton, J., & Delorme, A. (2004). Mining event-related brain dynamics. Trends in Cognitive Sciences, 8(5), 204–210.

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press.

Thompson, E. (2007). Mind in Life: Biology, Phenomenology, and the Sciences of Mind. Harvard University Press.

Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.

Webb, G. I., Hyde, R., Cao, H., Nguyen, H. L., & Petitjean, F. (2016). Characterizing concept drift. Data Mining and Knowledge Discovery, 30(4), 964–994.

Widmer, G., & Kubat, M. (1996). Learning in the presence of concept drift and hidden contexts. Machine Learning, 23(1), 69–101.

Žliobaitė, I. (2010). Learning under concept drift: An overview. arXiv preprint arXiv:1010.4784.