Energy Manifold Natural Gradient Descent: From Riemannian Optimization to Modern Neuroscience, NeuroAI and Quantum Physics

When Geometry, Energy, Artificial Intelligence and Neuroscience Converge

Modern Artificial Intelligence is rapidly moving beyond the idea that learning simply means minimizing an error function. Increasingly, researchers are asking a deeper question: what is the structure of the space in which learning takes place?

This question becomes particularly important when the system being modelled is constrained, nonlinear, dynamic, or governed by physical principles.

A recent work titled “Energy Manifold Natural Gradient Descent: Riemannian Optimization for Neural PDE Solvers”, by Zhangyong Liang and Huanhuan Gao, introduces Energy Manifold Natural Gradient Descent (EMNGD), a mathematical framework that extends energy-based natural-gradient optimization from unconstrained Euclidean parameter spaces to constrained Riemannian parameter manifolds.

At its core, the framework proposes a simple but powerful principle:

An optimization algorithm should not only determine how to reduce error; it should account for the geometry, constraints, and energy structure of the system being optimized.

The paper develops this principle for neural PDE solvers and establishes theoretical properties involving constrained tangent-space optimization, energy-induced geometry, convergence, coordinate invariance, Woodbury reformulations and scalable Nyström-based approximations.
This mathematical perspective also creates an interesting conceptual pathway toward Computational Neuroscience, NeuroAI, Brain–Computer Interfaces, neural manifolds, neurorehabilitation and physics-inspired modelling.

The connection with Quantum Physics is particularly interesting—but must be interpreted carefully. EMNGD does not demonstrate that the brain performs quantum computation. Instead, the connection exists primarily at the level of mathematical concepts such as geometry, energy, state spaces, constraints and variational principles.

1. Why Do We Need Something Beyond Ordinary Gradient Descent?

Gradient descent is one of the fundamental mechanisms underlying modern machine learning.

In its simplest mathematical form,

$$\theta_{k+1} = \theta_k-\alpha_k\nabla L(\theta_k),$$

where $\theta$ represents the parameters of a model, $L(\theta)$ represents the objective or loss function, and $\alpha_k$ determines the magnitude of the update.

The method is extremely effective because it provides a straightforward way to move toward a lower-loss configuration.

However, this formulation implicitly assumes that the parameter space can be treated as an ordinary Euclidean space.

Many scientific problems do not satisfy this assumption.

The parameters of a model may be subject to normalization conditions, physical restrictions, symmetry constraints, conservation laws, positivity conditions, or other mathematical relationships.

In these situations, the model cannot move arbitrarily in every direction.

The feasible parameter configurations may instead form a curved mathematical space, commonly represented as a manifold.

This creates an important distinction:

• Unconstrained optimization: move in any direction.

• Constrained optimization: move only in physically or mathematically permissible directions.

• Manifold optimization: account for the intrinsic geometry of those permissible directions.

The EMNGD framework addresses precisely this problem.

2. Neural PDE Solvers: Where Machine Learning Meets Physics

Partial differential equations are among the most important mathematical tools used to describe physical systems.

They can represent processes involving space, time, diffusion, waves, fluid dynamics, electromagnetic phenomena and many other physical processes.

A simplified PDE can be represented as

$$Lu=f,$$

together with an appropriate boundary condition,

$$Bu=g.$$

Here, $L$ represents a differential operator, $u$ is the unknown solution, and $f$ represents the governing source or forcing term.

Neural PDE solvers use a neural network to approximate the unknown function:

$$u_\theta(x).$$

Instead of explicitly solving the differential equation using only conventional numerical techniques, the neural network parameters are optimized so that the network output satisfies the governing equation and its boundary conditions.

The optimization may therefore involve an energy functional such as

$$E(u) = \int_\Omega (Lu-f)^2dx + \tau\int_{\partial\Omega}(Bu-g)^2ds.$$

The neural network is trained to reduce the residual associated with the physical problem.

This is one of the important ideas behind Physics-Informed Machine Learning.

The neural network is no longer learning solely from empirical labels.

It is being constrained by mathematical knowledge about the system.

3. From an Error Function to an Energy Landscape

One of the most interesting conceptual aspects of this approach is the interpretation of the objective as an energy landscape.

Instead of thinking only in terms of:

“Which parameter value gives the smallest loss?”

we can ask:

“How does a change in the parameters alter the underlying function and its energy?”

A neural network creates a mapping between parameter space and function space:

$$\theta \longrightarrow u_\theta.$$

Therefore, a parameter update is not merely a numerical movement.

It changes the function represented by the network.

The paper constructs an energy-induced metric based on the curvature of the function-space energy. In simplified form,

$$G_E(\theta)_{ij} = D^2E(u_\theta) \left( \partial_{\theta_i}u_\theta, \partial_{\theta_j}u_\theta \right).$$

This means that the optimization geometry is determined by how parameter perturbations affect the underlying function-space energy.

This produces an important conceptual shift:

Parameter-space movement and function-space movement are not necessarily the same thing.

A very large numerical parameter change may produce only a small functional change, while a small parameter change may produce a large functional consequence.

4. Understanding Riemannian Geometry

To understand EMNGD, it is useful to understand the concept of a Riemannian manifold.

Imagine the surface of a sphere.

A point located on the surface cannot move arbitrarily through three-dimensional space if it is required to remain on the sphere.

At every point, however, there is a local collection of allowable directions.

This collection is called the tangent space:

$$T_xM.$$

The tangent space provides the locally feasible directions for movement.

In a machine-learning problem, the parameter manifold can similarly describe the set of valid parameter configurations.

The model must then optimize within the corresponding tangent space.

The paper formulates EMNGD on a smooth Riemannian manifold $(M,g_0)$, where the Riemannian metric defines the geometry of the feasible parameter space.

The important point is that EMNGD operates with two interacting geometries:

• the geometry of the parameter manifold;

• the geometry induced by the underlying function-space energy.

This separation is central to the framework.

5. What Is Energy Manifold Natural Gradient Descent?

Energy Manifold Natural Gradient Descent, or EMNGD, extends Energy Natural Gradient Descent to situations where the parameter space itself is a constrained Riemannian manifold.

The optimization process can be conceptually represented as:

$$\boxed{ \text{Parameter Manifold} \rightarrow \text{Tangent Space} \rightarrow \text{Energy Geometry} \rightarrow \text{Optimal Feasible Direction} \rightarrow \text{Retraction} \rightarrow \text{Updated State} }$$

Instead of calculating an unconstrained update and then simply attempting to repair it, EMNGD incorporates the feasible geometry directly into the optimization direction.

This is mathematically significant because the paper shows that an ambient inverse-curvature operation followed by projection does not generally produce the same result as solving the curvature problem directly on the tangent space.

In simple terms:

If a system is constrained, the constraint should influence the direction of learning itself.

6. Why Tangent-Space Optimization Matters

Suppose the parameter space is represented by a manifold $M$.

At a particular point $x$, the system is allowed to move only along directions contained within

$$T_xM.$$

Therefore, the optimization problem must respect this local geometry.

This becomes particularly important when the curvature of the objective varies strongly between feasible and infeasible directions.

An unconstrained Newton-like update may identify a mathematically attractive direction that cannot actually be realized by the model.

EMNGD instead searches for the best direction that remains feasible.

This leads to one of the most important theoretical interpretations of the framework.

Under the stated coercivity conditions, the push-forward of the undamped EMNGD direction corresponds to the best feasible approximation to the function-space Newton vector under the energy metric.

Therefore, EMNGD can be understood as a bridge between:

Newton optimization → function-space geometry → constrained parameter geometry.

7. Retraction: Preserving the Manifold

After calculating a tangent-space direction, the model must return to the actual manifold.

This is achieved using a retraction.

A simplified update can be written as

$$x_{k+1} = R_{x_k}(-\alpha_k\eta_{x_k}),$$

where $R$ maps the tangent-space update back onto the manifold.

This prevents the optimization procedure from leaving the feasible parameter space.

The concept is important because curved spaces do not behave like flat Euclidean spaces.

A straight-line update in the surrounding space does not necessarily remain on the manifold.

Thus:

The optimizer calculates locally feasible movement, and the retraction converts that local movement into a valid global parameter update.

8. Why Does the Metric Matter?

Natural-gradient methods are based on a deeper question:

What should distance mean in the space of the model?

Ordinary gradient descent implicitly assumes Euclidean geometry.

Natural-gradient methods replace this with a more meaningful metric.

Different scientific problems may require different geometries.

For example, machine learning has explored metrics associated with:

• Fisher information;

• Wasserstein geometry;

• Sobolev structures;

• information geometry;

• function-space representations;

• and energy-induced geometries.

EMNGD uses the curvature of the underlying function-space energy to construct its metric.

This has an important implication for neuroscience.

The numerical distance between two neural states does not necessarily correspond to their functional distance.

Two neural configurations may be numerically different but functionally equivalent.

Conversely, a small change in a critical neural population may produce a substantial behavioural effect.

This motivates a broader question:

Should computational models of the brain use geometries that reflect neural function rather than merely numerical coordinates?

9. Neural Manifolds and Modern Neuroscience

Modern neuroscience increasingly studies neural activity as a high-dimensional dynamical system.

If $N$ neurons are recorded simultaneously, their activity at a particular time can be represented as

$$\mathbf{x}(t) = [x_1(t),x_2(t),...,x_N(t)].$$

This creates an $N$-dimensional observation space.

However, meaningful neural activity does not necessarily explore every possible point within this enormous space.

Behaviourally relevant activity may occupy a lower-dimensional, structured region.

This is commonly discussed using the concept of neural manifolds.

A neural manifold can therefore be viewed conceptually as a structured space containing the states or trajectories relevant to a particular neural process.

For example, during movement, neural population activity may follow trajectories corresponding to different motor states.

This creates an intriguing parallel with manifold-based optimization.

The connection is not that the brain has been proven to implement EMNGD.

Rather, EMNGD raises a methodological possibility:

If neural activity occupies structured manifolds, perhaps some computational models of neural systems should explicitly respect that geometry.

This is a hypothesis that can be tested experimentally.

10. From Neural Signals to Neural State Spaces

Brain–Computer Interfaces provide a particularly interesting example.

A BCI generally records physiological signals such as EEG, MEG, fNIRS or other modalities and attempts to infer information about the underlying neural state.

The process can be simplified as:

$$\text{Neural Signal} \rightarrow \text{Feature Space} \rightarrow \text{Latent Representation} \rightarrow \text{Decoder} \rightarrow \text{Output}.$$

Traditional machine-learning methods frequently focus on maximizing classification or regression performance.

But neural activity is not static.

It changes with:

• attention;

• fatigue;

• learning;

• task demands;

• sensory feedback;

• disease;

• rehabilitation;

• and individual variability.

Consequently, the latent representation itself may evolve.

A future BCI architecture could potentially represent this evolution using a subject-specific neural manifold:

$$z(t)\in\mathcal{M}_{subject}.$$

The decoder could then operate on the geometry of that latent state space rather than treating every observation as an independent Euclidean feature vector.

This is an important potential research direction for adaptive and personalized BCI.

11. From SIGNAL → CLASSIFIER to Neural State Modelling

A conventional BCI pipeline may approximately resemble:

$$\boxed{ \text{SIGNAL} \rightarrow \text{CLASSIFIER} \rightarrow \text{COMMAND} }$$

A more advanced NeuroAI framework could potentially evolve toward:

$$\boxed{ \text{NEURAL STATE} \rightarrow \text{GEOMETRIC MODEL} \rightarrow \text{PREDICTION} \rightarrow \text{PERTURBATION} \rightarrow \text{MEASUREMENT} \rightarrow \text{ADAPTATION} }$$

This represents a conceptual shift from simply identifying a neural state to modelling how that state evolves.

The objective becomes more ambitious.

Instead of asking:

“What does the current EEG signal represent?”

we begin asking:

“Where is the neural system currently located in its state space, where can it move next, and how can an intervention change that trajectory?”

This is where NeuroAI, dynamical systems and geometric modelling can potentially converge.

12. Potential Relevance to Neurorehabilitation

Neurorehabilitation provides another compelling setting for this type of thinking.

After neurological injury, the relationship between neural activity, motor intention, sensory feedback and movement can change.

A rehabilitation system therefore needs to respond to a dynamic and individualized biological system.

Imagine combining:

• EEG;

• EMG;

• motion data;

• robotic rehabilitation;

• stimulation;

• behavioural measurements;

• and clinical outcome measures.

These observations could potentially be integrated into a latent neural-behavioural state space.

The patient's state could be represented as:

$$S_t\in\mathcal{M}_{patient}.$$

Rehabilitation could then be conceptualized as a controlled transition:

$$S_0 \rightarrow S_1 \rightarrow S_2 \rightarrow ... \rightarrow S_{functional}.$$

The objective would not simply be to minimize a machine-learning loss.

Instead, the computational system could potentially attempt to identify interventions that move the individual toward a desired functional state while respecting physiological and behavioural constraints.

This provides an interesting conceptual foundation for future precision neurorehabilitation and closed-loop neurotechnology.

Importantly, this remains a prospective research direction. EMNGD itself has not been clinically validated as a neurorehabilitation algorithm.

13. The Physics-Informed Perspective

The connection between EMNGD and physics is direct because the framework is developed for neural PDE solvers.

Physics frequently describes complex systems using:

• differential equations;

• energy functions;

• variational principles;

• constraints;

• state spaces;

• curvature;

• and dynamical evolution.

Physics-Informed Machine Learning attempts to incorporate some of this prior mathematical knowledge directly into AI models.

Instead of asking a neural network to learn every relationship solely from data, we can impose mathematical structure on the learning process.

This can potentially improve:

• physical consistency;

• data efficiency;

• interpretability;

• stability;

• and generalization.

The EMNGD framework takes this philosophy one step further by asking how the geometry of the optimization itself should reflect the underlying energy structure.

14. The Connection with Quantum Physics

The relationship with Quantum Physics requires particular scientific caution.

EMNGD is not a quantum algorithm, and the paper does not provide evidence that biological brains perform quantum computation.

The meaningful connection is mathematical.

Modern quantum theory uses highly structured state spaces and mathematical concepts involving:

• vectors and inner products;

• operators;

• state transformations;

• constrained states;

• variational formulations;

• geometry;

• and information-theoretic structures.

Similarly, geometric optimization considers:

$$\text{State} + \text{Metric} + \text{Constraint} + \text{Energy} + \text{Evolution}.$$

Therefore, there is a legitimate conceptual intersection between the mathematical language used in geometric optimization and mathematical structures found in physics.

But there is a crucial distinction:

$$\boxed{ \text{Mathematical Similarity} \neq \text{Physical Equivalence} }$$

The fact that neuroscience, machine learning and quantum physics can use related mathematical concepts does not establish that the same physical mechanism operates in all three systems.

This distinction is essential for scientifically responsible interdisciplinary research.

15. Energy Is Not Always Physical Energy

The word energy can also create confusion when discussing neuroscience and machine learning.

An energy function in an optimization algorithm is not automatically equivalent to physical energy.

We should distinguish between:

Mathematical energy: an objective function used to quantify the state of a model.

Physical energy: a measurable physical quantity governed by physical laws.

Neural or metabolic energy: biological energetic processes associated with neural activity.

These concepts may interact in certain models, but they should not be treated as interchangeable.

Therefore, when discussing EMNGD, “energy” primarily refers to the mathematical function-space objective from which the optimization geometry is derived.

This distinction prevents a common mistake in interdisciplinary discussions.

16. The Woodbury Identity: Making Large Problems More Manageable

One of the technically important components of the EMNGD framework is its connection to the Woodbury identity.

Modern neural networks can contain enormous numbers of parameters.

Suppose:

$$p=\text{number of parameters}$$

and

$$N=\text{number of residual samples}.$$

If

$$N\ll p,$$

then directly solving a large parameter-space system may be computationally expensive.

For quadratic residual energies and generalized Gauss–Newton structures, the paper demonstrates that the damped tangent system can be transferred from parameter space into sample space using Woodbury duality.

Conceptually:

$$\boxed{ \text{Large Parameter-Space System} \rightarrow \text{Sample-Space System} }$$

This does not fundamentally change the underlying optimization geometry.

It changes the computational representation used to solve the system.

That distinction is extremely important.

17. Nyström Approximation and Computational Scalability

The paper also investigates Nyström approximation as a mechanism for reducing the computational burden associated with large positive-semidefinite matrices.

The basic idea is to approximate a large matrix using a lower-rank representation.

This can make large-scale computations more manageable.

However, approximation introduces a trade-off between computational efficiency and directional accuracy.

The paper distinguishes between using a Nyström approximation directly to calculate an approximate direction and using the approximation as a preconditioner for an iterative solver.

The second approach can, under appropriate conditions, recover the exact damped direction as the iterative process converges.

This illustrates a broader computational principle:

An approximation can be valuable not only because it gives a cheaper answer, but because it can make the exact answer computationally accessible.

18. Damping and Numerical Stability

Curvature-based optimization can become difficult when the underlying matrices are poorly conditioned.

A damped system can be represented as

$$(G_E+\lambda I)d=g.$$

Here, $\lambda$ acts as a regularization parameter.

Damping can improve numerical stability and conditioning, although excessive damping can move the optimization direction closer to an ordinary Riemannian gradient direction.

This demonstrates another important principle:

The mathematically sophisticated direction is not necessarily the numerically easiest direction to compute.

Scientific machine learning must therefore balance:

• geometric accuracy;

• computational efficiency;

• numerical stability;

• and convergence.

19. Convergence and Theoretical Guarantees

The theoretical framework developed in the paper includes several important properties.

These include:

• coordinate invariance;

• reduction to ENGD in the Euclidean case;

• global first-order convergence under Armijo backtracking;

• robustness to controlled inexact tangent solves;

• and well-posedness under damping.

Coordinate invariance is particularly interesting from a geometric perspective.

If the same mathematical manifold is described using different coordinates, the underlying geometric optimization direction should not depend artificially on the coordinate representation.

This is one reason geometry-based methods are attractive for complex scientific systems.

20. What Do the Experiments Actually Demonstrate?

The experimental results reported in the paper provide evidence for the framework on the evaluated neural PDE benchmarks.

The reported studies include:

• verification of the Euclidean reduction;

• agreement between primal and Woodbury formulations;

• neural PDE optimization experiments;

• comparisons with baseline optimizers;

• and computational investigations of Nyström-based approaches.

The paper reports higher accuracy and faster convergence than the compared baselines under the benchmark settings.

However, an important limitation must be emphasized.

Much of the empirical evaluation focuses on the Euclidean specialization of the framework.

The authors identify architecture-matched experiments involving genuinely constrained neural PDE models as an important future direction for determining the practical benefit of the manifold component itself.

This distinction between theoretical generality and empirical validation is scientifically important.

21. What EMNGD Does Not Prove

Interdisciplinary research becomes stronger when the boundaries of evidence are clearly stated.

The EMNGD paper does not demonstrate that:

• the human brain implements EMNGD;

• biological neural populations literally perform Riemannian optimization;

• consciousness is quantum mechanical;

• the brain is a quantum computer;

• EMNGD automatically improves EEG decoding;

• EMNGD has already been clinically validated for neurorehabilitation;

• or quantum mechanics explains BCI performance.

These statements would go beyond the evidence.

What the paper provides is a rigorous mathematical framework for optimization under constraints using energy-induced geometry.

That framework can inspire neuroscience hypotheses, but those hypotheses must be independently tested.

22. Why This Framework Is Interesting for NeuroAI

NeuroAI attempts to understand biological intelligence while simultaneously using neuroscience to inspire better artificial systems.

This requires more than simply applying deep learning to neural datasets.

The brain is:

• dynamic;

• adaptive;

• constrained;

• nonlinear;

• context-dependent;

• highly interconnected;

• and individual-specific.

Therefore, a useful NeuroAI model may eventually need to represent not only neural states but also the structure of transitions between those states.

This leads to a broader computational architecture:

$$\boxed{ \text{Neural Data} \rightarrow \text{Latent State} \rightarrow \text{Geometry} \rightarrow \text{Dynamics} \rightarrow \text{Prediction} \rightarrow \text{Intervention} }$$

This is where concepts from EMNGD become conceptually interesting.

23. Toward Geometry-Aware BCI

A future geometry-aware BCI could potentially combine signal processing with manifold learning and dynamical modelling.

A conceptual pipeline might look like:

$$\text{EEG/MEG/fNIRS} \rightarrow \text{Representation Learning} \rightarrow \text{Neural Manifold} \rightarrow \text{State Estimation} \rightarrow \text{Geometric Decoder} \rightarrow \text{Adaptive Output}.$$

Such a system could potentially account for changes in neural representation over time.

This may be particularly relevant to:

• motor-imagery BCI;

• speech-related BCI;

• adaptive decoding;

• neuroprosthetic control;

• longitudinal BCI;

• and rehabilitation-oriented BCI.

The major research challenge would be determining whether explicit manifold geometry provides measurable advantages over conventional Euclidean machine-learning approaches.

That is an experimentally testable question.

24. Toward Precision Neurorehabilitation

The same principle can be extended conceptually to neurorehabilitation.

Instead of treating rehabilitation as a sequence of independent therapy sessions, we could model it as a dynamic trajectory through a patient-specific state space.

For example:

$$\mathcal{M}_{patient} = \{ \text{neural state}, \text{motor state}, \text{sensory state}, \text{behavioural state} \}.$$

Each therapy session could provide new observations.

The computational system could continuously update its model:

$$S_t \rightarrow S_{t+1}.$$

The ultimate objective could be to estimate which intervention produces the most desirable transition.

This could potentially combine:

neural signals + robotics + neuromodulation + behavioural assessment + AI + dynamical systems.

Such an approach could contribute to the development of more personalized closed-loop rehabilitation technologies.

25. From Prediction to Perturbation

A major transition in future neuroscience may be from observation toward prediction and intervention.

Traditional analytical neuroscience often follows:

$$\text{Observe} \rightarrow \text{Analyse}.$$

A more advanced computational framework could follow:

$$\boxed{ \text{Observe} \rightarrow \text{Model} \rightarrow \text{Predict} \rightarrow \text{Perturb} \rightarrow \text{Measure} \rightarrow \text{Adapt} }$$

This is especially relevant for closed-loop neurotechnology.

The system continuously measures the neural state, predicts its trajectory, applies an intervention and then measures the response.

The geometry of the state space could potentially help determine which transitions are feasible or meaningful.

26. A Broader Connection: Dynamical Systems

The human brain is fundamentally dynamic.

Neural activity evolves continuously over time.

A static classifier can describe what a signal represents at a particular moment, but it does not necessarily explain how the neural system moves between states.

A dynamical representation can instead be expressed as:

$$z_{t+1}=f(z_t,u_t),$$

where:

• $z_t$ is the latent neural state;

• $u_t$ represents an external input or intervention;

• and $f$ describes the underlying dynamics.

The question then becomes:

What constraints define the possible trajectories of $z_t$?

This is where manifold geometry becomes potentially valuable.

Rather than modelling arbitrary transitions between arbitrary points, a geometric dynamical system attempts to understand the structured space through which the system evolves.

27. The Mathematical Architecture

At a high level, EMNGD can be understood through three mathematical components.

First, there is the parameter manifold:

$$\mathcal{M}.$$

This defines the admissible parameter configurations.

Second, there is a mapping from parameters to functions:

$$P:\mathcal{M}\rightarrow\mathcal{X}.$$

Third, there is an energy functional:

$$E:\mathcal{X}\rightarrow\mathbb{R}.$$

Together, these form:

$$\boxed{ \mathcal{M} \xrightarrow{P} \mathcal{X} \xrightarrow{E} \mathbb{R} }$$

The complete optimization objective is therefore:

$$F=E\circ P.$$

The differential

$$J_x=dP_x$$

maps tangent directions in parameter space into function space.

This mathematical structure explains why EMNGD is fundamentally different from simply minimizing a scalar loss in an unconstrained parameter vector.

28. A Potential Universal Principle for Scientific AI

The most transferable lesson from this framework may not be the specific EMNGD algorithm.

It may be the following principle:

Optimization should respect the structure of the system being optimized.

In physics, that structure may be defined by physical laws.

In biology, it may be defined by physiological constraints.

In neuroscience, it may involve neural population dynamics and connectivity.

In robotics, it may involve mechanical constraints.

In quantum systems, it may involve mathematically structured state spaces.

In machine learning, it may involve parameter symmetries and representation geometry.

The specific mathematics differs across disciplines.

But the general principle remains:

$$\boxed{ \text{SYSTEM STRUCTURE} \rightarrow \text{GEOMETRY} \rightarrow \text{OPTIMIZATION} \rightarrow \text{PREDICTION} \rightarrow \text{CONTROL} }$$

29. Why “Better Geometry” May Matter More Than “Bigger Models”

Artificial intelligence has achieved remarkable progress through increasingly large models, larger datasets and greater computational resources.

But scaling alone may not solve every scientific problem.

A sufficiently large model can still struggle if it ignores the structure of the system it is trying to model.

This suggests another direction for AI research:

Not only larger models.

But:

better representations, better constraints, better geometries and better dynamical models.

In scientific AI, this may become particularly important because the objective is not simply to generate plausible outputs.

The objective is often to model something that must obey rules.

30. The Quantum–Neuroscience Boundary

The intersection of Quantum Physics and Neuroscience is scientifically fascinating, but it is also an area where careful language is essential.

There is a legitimate mathematical relationship between ideas such as:

$$\text{state spaces}, \quad \text{geometry}, \quad \text{energy}, \quad \text{variational principles}, \quad \text{constraints}.$$

These concepts occur across multiple areas of mathematics and physics.

However:

$$\boxed{ \text{Shared Mathematics} \not\Rightarrow \text{Shared Biological Mechanism} }$$

Therefore, geometric similarities between optimization and quantum theory should not be presented as evidence that the brain is quantum computational.

A scientifically meaningful Quantum Neuroscience programme requires experimentally testable physical mechanisms, not only mathematical analogies.

This distinction is particularly important when discussing consciousness.

31. What Could Future Research Test?

The most valuable outcome of interdisciplinary theory is not simply conceptual similarity.

It is the generation of testable hypotheses.

Future studies could investigate whether:

• neural population activity possesses stable subject-specific manifold geometry;

• manifold-aware representations improve BCI decoding;

• geometric optimization improves adaptation to non-stationary EEG;

• neural manifolds predict rehabilitation outcomes;

• multimodal neural signals can be represented within a common latent geometry;

• energy-based models can capture meaningful neural attractor dynamics;

• geometry-aware control can improve closed-loop neurotechnology;

• and physics-informed neural models can reproduce biological dynamics more accurately than purely data-driven models.

These questions can be experimentally evaluated.

That is where the conceptual bridge between EMNGD and neuroscience becomes scientifically useful.

32. Toward a Future NeuroAI Architecture

A possible future architecture could combine:

$$\boxed{ \text{EEG} + \text{EMG} + \text{fNIRS/fMRI} + \text{Motion} + \text{Behaviour} + \text{Clinical Data} }$$

with:

$$\boxed{ \text{Deep Learning} + \text{Manifold Learning} + \text{Dynamical Systems} + \text{Geometric Optimization} }$$

The resulting model could estimate an individualized latent state:

$$z_t\in\mathcal{M}_{subject}.$$

A dynamical model could then describe:

$$z_{t+1}=f(z_t,u_t).$$

The ultimate objective would be to connect:

$$\boxed{ \text{Neural State} + \text{Dynamics} + \text{Constraints} + \text{Intervention} \rightarrow \text{Predicted Outcome} }$$

This could represent an important direction for precision neurotechnology.

33. From Artificial Neural Networks to Artificial Dynamical Systems

Traditional machine learning often focuses on static mappings:

$$x\rightarrow y.$$

Biological intelligence is considerably more dynamic.

A neural system continuously changes its internal state as a function of:

• sensory information;

• internal activity;

• memory;

• environmental context;

• behavioural goals;

• and previous states.

Therefore, future NeuroAI systems may increasingly need to model:

$$\text{state} + \text{trajectory} + \text{constraints} + \text{transition dynamics}.$$

The important question becomes not simply:

“What is the current state?”

but:

“How can the system evolve from this state?”

And further:

“Which transitions are feasible?”

Geometry provides a natural mathematical language for investigating these questions.

34. The Emerging Convergence

We can now visualize a broader interdisciplinary pathway:

$$\boxed{ \text{Computational Neuroscience} }$$

↓

$$\text{Neural Dynamics}$$

↓

$$\text{Neural Manifolds}$$

↓

$$\text{Riemannian Geometry}$$

↓

$$\text{Geometric Deep Learning}$$

↓

$$\text{Physics-Informed AI}$$

↓

$$\text{Energy-Based Modelling}$$

↓

$$\text{Adaptive NeuroAI}$$

↓

$$\boxed{ \text{Closed-Loop Neurotechnology} }$$

This should not be interpreted as a single established scientific pipeline.

Rather, it represents a potential interdisciplinary research landscape where mathematical tools developed in optimization and physics may be investigated within computational neuroscience.

35. The Central Scientific Question

The deeper lesson from EMNGD is therefore not simply about one optimization algorithm.

It is about how we think about learning.

If a system is constrained, then its constraints matter.

If a system has geometry, that geometry matters.

If a system evolves dynamically, its trajectories matter.

If a system is governed by physical laws, those laws matter.

And if a system is biological, its physiological structure matters.

This leads to a broader principle:

$$\boxed{ \textbf{Learning should respect the structure of the system being learned.} }$$

36. Final Perspective: From Learning to Understanding

The development of increasingly sophisticated AI systems has largely focused on improving prediction.

But scientific intelligence requires something deeper.

A scientific model should ideally represent:

• what states are possible;

• how those states are related;

• how the system evolves;

• what constraints govern its evolution;

• and how interventions alter its trajectory.

EMNGD provides one mathematical example of this philosophy.

It combines energy geometry, manifold constraints and natural-gradient optimization to construct feasible learning directions for neural PDE problems.

Its potential relevance to neuroscience lies not in claiming that biological brains literally execute EMNGD, but in the possibility that similar mathematical principles may help us construct better computational descriptions of constrained neural systems.

For BCI research, this raises the possibility of moving beyond:

$$\text{Signal} \rightarrow \text{Classification}$$

toward:

$$\text{Neural State} \rightarrow \text{Geometry} \rightarrow \text{Dynamics} \rightarrow \text{Prediction} \rightarrow \text{Intervention}.$$

For neurorehabilitation, it raises the possibility of representing recovery as an individualized trajectory through a neural-behavioural state space.

For NeuroAI, it raises a more fundamental question:

Can artificial systems learn not merely from the data generated by a complex system, but according to the structural geometry of that system?

And for the intersection of Neuroscience and Quantum Physics, the scientifically responsible question is not whether the brain is “quantum,” but whether mathematical frameworks involving state spaces, energy, geometry, constraints and variational principles can provide useful new ways of understanding complex biological systems.

The future of interdisciplinary science will depend on maintaining this distinction between analogy, mathematical formulation, computational evidence and physical evidence.

Perhaps the next generation of intelligent systems will therefore not be defined solely by:

$$\text{MORE DATA} \rightarrow \text{MORE PARAMETERS} \rightarrow \text{MORE COMPUTE}$$

but increasingly by:

$$\boxed{ \text{BETTER REPRESENTATION} \rightarrow \text{BETTER GEOMETRY} \rightarrow \text{BETTER DYNAMICS} \rightarrow \text{BETTER PREDICTION} \rightarrow \text{BETTER INTERVENTION} }$$

The question is no longer only:

Can machines learn?

The deeper question may be:

Can machines learn according to the structure of the world they are trying to understand?

That question sits at a fascinating intersection of Artificial Intelligence, Computational Neuroscience, Physics, Mathematics, Brain–Computer Interfaces and Neurotechnology.

And perhaps, ultimately:

The next frontier of intelligent systems may not be larger models—but better geometry.

Key Takeaways

• EMNGD extends Energy Natural Gradient Descent to constrained Riemannian parameter manifolds.

• It incorporates feasible tangent directions directly into the optimization problem.

• The optimization geometry is induced by the underlying function-space energy.

• Under the paper's assumptions, EMNGD provides a best-feasible-Newton-correction interpretation.

• Woodbury duality can transform certain tangent systems into sample-space formulations.

• Nyström approximation can improve computational scalability, with an associated approximation–accuracy trade-off.

• The framework has theoretical guarantees involving convergence, coordinate invariance and robustness to controlled inexact solves.

• The reported experiments support the approach on the evaluated neural PDE benchmarks.

• Genuine constrained-manifold experiments remain important for assessing the empirical value of the manifold component.

• Connections to neural manifolds, BCI, NeuroAI and neurorehabilitation are promising research hypotheses, not demonstrated clinical applications.

• The connection to Quantum Physics is primarily mathematical and conceptual and should not be interpreted as evidence for quantum brain computation.

Scientific Boundary: What We Know vs. What We Can Explore

It is useful to separate three levels of interpretation.

Established by the EMNGD framework

Energy geometry + Riemannian constraints + tangent-space optimization + retraction + scalable linear algebra.

Reasonable interdisciplinary research direction

Neural manifolds + geometric optimization + adaptive BCI + NeuroAI + precision neurorehabilitation.

Not established by this work

Quantum consciousness, quantum brain computation, or biological implementation of EMNGD.

Maintaining this distinction is essential for rigorous interdisciplinary neuroscience.

Conclusion

The most transferable idea may ultimately be expressed in one sentence:

$$\boxed{ \textbf{Do not optimize only according to where you can move.} }$$

Instead:

$$\boxed{ \textbf{Understand the geometry of where you are allowed to move.} }$$

And perhaps, for the future of NeuroAI:

$$\boxed{ \textbf{Do not merely model neural activity.} }$$ $$\boxed{ \textbf{Model the geometry of possible neural states.} }$$

This perspective could open new computational questions at the intersection of Riemannian Geometry, Artificial Intelligence, Computational Neuroscience, Brain–Computer Interfaces, Neurorehabilitation, Physics-Informed Machine Learning and theoretical physics.

The future of intelligent systems may depend not only on how much they learn, but on whether they understand the structure within which learning occurs.