Flow Equivariant RNNs (FERNNs)

• Flow Equivariant Recurrent Neural Networks (FERNNs) are sequence models that embed one-parameter Lie group flows to enforce continuous symmetry in temporal data.
• They elevate the hidden state to a joint V×G tensor and use flow convolution and flow shift for maintaining geometric consistency and effective weight sharing.
• Empirical benchmarks show FERNN variants achieve lower error rates and faster training while maintaining stable long-horizon predictions compared to traditional G-RNNs.

Flow Equivariant Recurrent Neural Networks (FERNNs) are a class of sequence models that embed continuous symmetry constraints—specifically one-parameter Lie group flows—directly into recurrent neural network (RNN) architectures. By enforcing equivariance not only to static group actions but also to temporally parameterized flows, FERNNs achieve principled inductive bias for data with smoothly evolving geometric structure, offering significant improvements in generalization, temporal extrapolation, and invariance to underlying motion parameters. FERNNs have been rigorously formalized and empirically validated across multiple domains, including video sequence prediction and world modeling in partially observed environments (Keller, 20 Jul 2025, Bernardo et al., 6 Nov 2025, Lillemark et al., 3 Jan 2026).

1. Mathematical Foundations of Flow Equivariance

FERNNs rest on group representation theory and the formalism of flows as one-parameter subgroups of Lie groups. A flow is parameterized by a Lie algebra element ν of a group G and is given by

$\psi_t(\nu) = \exp_G(t \nu) \in G$

satisfying $\psi_0(\nu) = e$ and $\psi_s \cdot \psi_t = \psi_{s+t}$. Typical choices for G include Euclidean translations $\mathbb{R}^n$, planar rotations $SO(2)$, or their products, directly modeling real-world image transformations such as translation or rotation.

The key equivariance requirement is that neural sequence maps preserve the action of these flows:

$\Phi[\psi(\nu) \cdot f] = \psi(\nu) \cdot \Phi[f],$

where $$(\psi(\nu) \cdot f)_t(x) = f_t( \psi_t(\nu)^{-1} \cdot x )$$. This property must hold for all flow generators $\nu \in V \subset \mathfrak{g}$, where $\mathfrak{g}$ is the Lie algebra of G (Keller, 20 Jul 2025, Lillemark et al., 3 Jan 2026).

Static equivariant models, such as G-equivariant RNNs, enforce equivariance only under time-invariant group actions and fail for time-varying (flow) symmetry. Theorem 3.1 in (Keller, 20 Jul 2025) establishes that the standard G-RNN recurrence inherently breaks flow symmetry except in degenerate settings.

2. Architecture and Implementation

FERNNs elevate the hidden state from its conventional representation (e.g., feature maps over G) to a joint $V \times G$ tensor, where V indexes flow generators ("velocity channels", Editor's term) and G indexes spatial positions. At each time step, the FERNN applies a structured recurrence composed of:

• Input Lifting: The input $\psi_0(\nu) = e$0 is lifted to $\psi_0(\nu) = e$1 via a trivial copy, without additional mixing across V.
• Flow Convolution: The hidden state $\psi_0(\nu) = e$2 undergoes group convolution that shares weights across ν, reinforcing equivariance.
• Flow Shift: Each velocity channel ν is shifted along G in accordance with $\psi_0(\nu) = e$3, synchronizing the hidden state with the respective flow generator.
• Recurrence: The update is given by:

$\psi_0(\nu) = e$4

where $\psi_0(\nu) = e$5 is a pointwise nonlinearity, and $\psi_0(\nu) = e$6 are group convolution kernels on $\psi_0(\nu) = e$7 and G, respectively. Typically, $\psi_0(\nu) = e$8 is trivial in V to avoid error propagation due to boundary artifacts.

The same principle extends to continuous-time (flow) cases, yielding the ODE:

$\psi_0(\nu) = e$9

with a built-in G-equivariant structure, block-diagonalized via the group Fourier transform (Bernardo et al., 6 Nov 2025).

3. Empirical Performance and Benchmarks

FERNNs have been evaluated extensively on synthetic and real video sequence benchmarks characterized by known geometric flows:

Task & Dataset	G-RNN	FERNN-V₁	FERNN-V₂
Flowing MNIST (V₂ᵀ, MSE)	8.1e-3	5.3e-4	1.5e-4
Rotating MNIST (V₄ᴿ, MSE)	4.0e-3	1.3e-3	6.1e-4
KTH Action (acc., V₂ᵀ)	66.5%	69.8%	71.6%

FERNNs train 3–5× faster, exhibit invariant generalization to unseen velocities, and maintain low MSE in sequence rollouts up to 70 steps—contrast with G-RNNs, where prediction error explodes beyond the training horizon. Even partial or coarse flow coverage in V substantially outperforms non-equivariant models. Ablations reveal that learned G-RNN⁺ variants, which attempt to discover flow symmetry directly through learned kernels, fail to match equivariant FERNN accuracy (Keller, 20 Jul 2025).

Flow Equivariant World Models (FloWM), which extend FERNNs to memory-augmented partially observed control settings, demonstrate similar superiority. For example, on 2D MNIST world and 3D block-world rollouts, FloWM achieves MSE <0.002 after 150–210 prediction steps, maintaining accurate memory of object position, while diffusion and SSM baselines drift or hallucinate (Lillemark et al., 3 Jan 2026).

4. Link to Continuous Attractors and Manifolds

FERNNs admit an analytic connection between symmetry structure and the emergent low-dimensional manifolds in their dynamics. For discrete-time or continuous-time equivariant RNNs,

• Any subgroup $\psi_s \cdot \psi_t = \psi_{s+t}$0 defines a fixed-point manifold $\psi_s \cdot \psi_t = \psi_{s+t}$1 isomorphic to the coset space $\psi_s \cdot \psi_t = \psi_{s+t}$2, of dimension $\psi_s \cdot \psi_t = \psi_{s+t}$3.
• On $\psi_s \cdot \psi_t = \psi_{s+t}$4 the network implements a continuous attractor: each point $\psi_s \cdot \psi_t = \psi_{s+t}$5 is a fixed point.
• Stability is determined by the spectral components $\psi_s \cdot \psi_t = \psi_{s+t}$6 under the group Fourier decomposition: tangential modes along $\psi_s \cdot \psi_t = \psi_{s+t}$7 are neutrally stable, while radial modes must have negative real part in their Jacobian eigenvalues for stability,

$\psi_s \cdot \psi_t = \psi_{s+t}$8

For compact Lie groups, Fourier-mode ODEs decouple into independent blocks for each irreducible representation (Bernardo et al., 6 Nov 2025). This formalism makes FERNNs analytically tractable and links neural manifold structure to symmetry constraints.

5. Applications and Practical Implementation

FERNNs are instantiated across domains requiring temporal geometric consistency. Major applications include:

• Sequential image/video modeling with moving components (e.g., Flowing MNIST, Moving KTH actions) (Keller, 20 Jul 2025).
• Memory-augmented world modeling in embodied agents, where partial observability and viewpoint changes necessitate simultaneous equivariance to both internal (object) and external (agent self-motion) flows. Group-structured latent memory and flow alignment mechanisms allow FERNN-based world models (FloWM) to preserve object trajectories and identities over hundreds of agent steps. (Lillemark et al., 3 Jan 2026)

Implementation leverages:

• Discretized grids of flow generators (V), e.g., for translation: $\psi_s \cdot \psi_t = \psi_{s+t}$9.
• Steerable group convolutions (E(2)-steerable CNNs) for rotational symmetry.
• Shared kernel parameters across V, yielding no increase in parameter count compared to G-RNNs.
• Max-pooling across V at readout to guarantee flow-invariant predictions when required.

FERNNs are typically trained using Adam optimizers, with gradient clipping for stability. Datasets are designed to probe both in-distribution and out-of-distribution (velocity, length) generalization.

6. Limitations and Ongoing Directions

While FERNNs guarantee exact equivariance for prescribed V, their hidden state scales linearly with $\mathbb{R}^n$0, leading to computational overhead for high-dimensional flow sets. Boundary truncation in V introduces minor errors; all experimental mixing in V is disabled to avoid error propagation, which may limit expressivity.

Ongoing research directions include:

• Steerable FERNNs: Learning steerable representations over G to reduce hidden state size and improve flexibility.
• Learning the flow set V: Allowing the model to discover underlying flow symmetries directly from data, endowing it with latent geometric structure.
• Extension to advanced architectures: Generalizing the lifting+flow shift construct to GRUs, LSTMs, Neural ODEs, SSMs, or Transformer-based models.
• Biological modeling: Relating FERNN dynamics to neural manifold geometry, traveling waves, and continuous attractors in neuroscience.
• Computational tradeoffs: While FERNN and FloWM incur moderate increases in inference/training FLOPs, they remain unique in achieving long-horizon memory fidelity in dynamic, partially observed worlds (Lillemark et al., 3 Jan 2026).

7. Relationship to Broader Equivariant Modeling

FERNNs extend the scope of geometric deep learning by formalizing and enforcing symmetry constraints in the temporal (flow) domain—not just for static group actions. Through this extension, models align more closely with the algebraic structure of physical, biological, and engineered systems, in which dynamics and perception are governed by continuous symmetries. FERNNs unify concepts from group convolutional RNNs, memory-augmented models, and control-theoretic observers, charting a mathematically principled and empirically validated path for symmetry-guided sequential learning (Keller, 20 Jul 2025, Bernardo et al., 6 Nov 2025, Lillemark et al., 3 Jan 2026).