Temporal Geometry of Deep Networks: Hyperbolic Representations of Training Dynamics for Intrinsic Explainability

AI summary

Overview

Research area: Interpretable machine learning / meta-learning on neural network parameters, combining graph neural networks, hyperbolic geometry (Poincaré ball model), and temporal graph learning.

Technical level: Advanced. The paper assumes familiarity with Riemannian geometry (exponential and logarithmic maps, parallel transport, Einstein gyromidpoints, Klein and Poincaré models), permutation equivariance, and meta-networks over neural parameters.

Scope (one sentence): The paper builds and evaluates a hyperbolic, temporally evolving graph meta-network (GTH-GMN) that embeds sequences of MLP parameter graphs into the Poincaré ball to explain how a network's weighted topology self-organizes during training.

What This Paper Is About

Most existing methods that treat a neural network's weights as a data object look at a single checkpoint, ignoring the trajectory traced during training. The authors argue that information is encoded not only in the weights themselves but in how they evolve, and that a network's weighted topology is the structure through which generalization, compression, and transfer learning arise. Their goal is an intrinsic form of explainability: one grounded in the model's own training dynamics rather than in post-hoc probes, obtained by embedding sequences of parameter graphs in hyperbolic space.

Key Contributions

1. A temporal parameter-graph formulation of training. For an MLP with L layers and widths (n_0, …, n_L), each training epoch t produces a snapshot G_t = (X_t, E_t, W_t) where neurons and biases are nodes and signed weights are edges; zero weights register no adjacency. This yields a sequence {G_t} from t = 1 to T, rather than a single checkpoint.

2. The Geometric Temporal Hyperbolic Graph Meta-Network (GTH-GMN). A meta-network that embeds parameter graphs in the Poincaré ball using hyperbolic affine maps for queries, keys, and values, geodesic-distance attention logits, and aggregation via Einstein gyromidpoints computed in the Klein model. The architecture is equivariant to neuron permutations within each snapshot and invariant to permutations of past snapshots, so functionally equivalent networks share a representation.

3. Temporal kernel meta-evolution. Instead of storing the whole sequence of hyperbolic embeddings, a GRU maintains a recurrent state per layer and attention head, emitting fresh Q/K/V parameters over time (dimension 3d² + 3d). This preserves temporal smoothness and limits memory to compact recurrent states.

4. A signed weight regressor grounded in network-science priors. Magnitude is predicted as a heavy-tailed power law of hyperbolic distance, with a context-dependent exponent, while polarity is modeled in the tangent space via a hyperbolically consistent inner product. This decouples strength from sign and ties large weights to short geodesic distances.

Main Findings

• INR classification (MNIST and Fashion-MNIST): The authors' temporal model with K-trace pooling reached 95.6 ± 0.18 on MNIST INRs and 80.72 ± 0.29 on Fashion-MNIST INRs. For comparison: DWSNets 85.71 ± 0.57 / 67.06 ± 0.29; NFN_HNP 92.5 ± 0.07 / 72.7 ± 1.53; NFN_NP 92.9 ± 0.22 / 75.6 ± 1.07. Lower baselines included MLP (17.55 ± 0.01 / 19.91 ± 0.47), MLP + Perm. aug (29.26 ± 0.18 / 22.76 ± 0.13), MLP + Alignment (58.98 ± 0.52 / 47.79 ± 1.03), INR2Vec (Arch.) (23.69 ± 0.10 / 22.33 ± 0.41), and Transformer (26.57 ± 0.18 / 26.97 ± 0.33).

• CIFAR-10 generalization prediction (Kendall's τ): GTH-GMN obtained 0.846 ± 0.004, below NFN_HNP (0.934 ± 0.001, marked best), NFN_NP (0.922 ± 0.001), and StatNN (0.915 ± 0.002). The paper describes the temporal hyperbolic encoder as matching or heavily approaching strong Euclidean and tensor-based baselines rather than leading on every task.

• Sinusoid-MLP regression (test MSE of a downstream MLP regressor): GTH-GMN achieved the lowest reported error at 1.06 ± 0.24, compared with GMN 1.13 ± 0.08, DWSNets 1.39 ± 0.06, INR2Vec (Arch.) 3.86 ± 0.32, MLP + Alignment 4.47 ± 0.15, Transformer 5.11 ± 0.12, MLP + Perm. aug 5.65 ± 0.01, and MLP 7.39 ± 0.19.

• Geometric interpretability claims: The paper reports that radial and angular structure of the embeddings tracks the self-organization of the underlying networks: hubs drift outward, stronger edges cluster radially (shorter hyperbolic distances), and excitatory versus inhibitory effects appear as angular variation in the tangent space.

• Sine wave frequency experiment: An additional evaluation on the benchmark of Navon et al. (2024), where each input network is an MLP fitting x ↦ a sin(bx) with a, b ∈ ℝ, is introduced. The paper content provided is truncated at the start of this experiment, so its results are not reported here.

Methodology in Plain English

The researchers treat a training run as a movie instead of a photograph. An MLP trained for T epochs is sliced into snapshots (checkpoints), and each snapshot is turned into a graph: neurons and bias units become nodes, and weights become edges with signs. Nodes carry simple descriptors — which layer they belong to, whether they are a bias, and standardized z-scores summarizing the absolute and signed strength of their connections. Edge descriptions are deliberately minimal, encoding only whether an edge is a bias connection and whether it crosses layers, so that no functions of the raw weight leak into the model.

These descriptors are first learned in ordinary flat (Euclidean) space at the origin of the ball, which keeps normalization, dropout, and mixing stable. Curvature enters afterward: an attention layer performs its linear maps by projecting into the tangent space, lifting back onto the ball along geodesics, and using parallel transport to place the bias. Attention scores are the negative hyperbolic distance between query and key plus an edge-context bias term, passed through a softmax and then reweighted by a gate that lets structurally important edges (cross-layer, bias) exert more influence. Because vectors on a curved manifold cannot simply be added, aggregation uses Einstein gyromidpoints, computed in the Klein model where closed-form barycenters exist, before averaging heads in the tangent space. A safety projection keeps embeddings away from the ball's boundary for numerical stability.

To handle time, a GRU per layer and head updates the attention kernels based on mean-pooled node features from the previous snapshot, so the kernels evolve smoothly without storing all past embeddings. A separate signed regressor predicts each weight: magnitude follows a power law of hyperbolic distance with node-dependent scale, steepness, and a mildly context-varying exponent; polarity comes from the angle between the transported displacement vector and a conformally normalized source feature. A Fermi–Dirac decoder, with a learned radius R and temperature T, converts distances into link probabilities. Training combines binary cross-entropy with negative sampling, magnitude and polarity losses, and geometric regularizers (a slope prior, temporal smoothness, and a ranking loss tying stronger interactions to shorter distances), following a curriculum that anneals the ranking margin and progressively hardens negative sampling. Optimization runs in two phases: a standard Euclidean pass for the kernels, regressors, and decoder, then a Riemannian pass that refines node positions on the ball by rescaling gradients with the conformal factor, accumulating moments in a RAMSGrad-style scheme, applying updates through the exponential map, and parallel transporting the moments.

Experimentally: MNIST and Fashion-MNIST images were each fit by a shallow INR with traces of T ∈ [80–100] checkpoints, split at the image level into 45,000 training, 5,000 validation, and 10,000 test images, and the meta-network never sees the image itself. CIFAR-10 trials each involve an MLP trained on a subset, with per-epoch test accuracy collected and rank correlation measured. The sinusoid task uses MLPs fitted to x ↦ a sin(bx).

Why This Matters

The paper reframes explainability as a question about the geometry of a training trajectory rather than a property of a frozen model. If a parameter graph's evolving position in a negatively curved space genuinely reflects hierarchy, module consolidation, and the redistribution of influence across layers, then explanations can be read off the model's own dynamics instead of being produced by post-hoc probes. The paper also introduces a concrete architectural constraint that is often missing from meta-network research: functional symmetries (neuron-permutation equivariance, temporal permutation invariance) are preserved while a latent geometry is recovered.

Potential applications:

• Model diagnostics and auditing. Tracking radial and angular structure across snapshots could give engineers a way to see when a network consolidates functional modules or redistributes influence across layers during re-training.
• Re-training pipelines in optimization environments. The paper specifically targets MLPs that require dynamic re-training, where checkpoint-only analyses are least informative.
• Meta-learning and transfer. The framework is motivated by generating weights for transfer learning across domains and by generalizing across architectures of varying depth, width, and skip connections.
• Model selection and generalization estimation. Predicting accuracy directly from weights, as in the CIFAR-10 experiment, is relevant to deciding which trained model to deploy without a full evaluation.

Industry relevance: Meta-networks over parameters are relevant to model zoos, hyperparameter and architecture search, federated learning, and parameter sharing — directions the related work cites as using graph-level abstractions. Practitioners who must repeatedly re-train models and want interpretable signals about how the model changed, and not only what it scores, are the closest industrial audience. The paper does not report deployment, latency, or cost measurements, so production feasibility is not established in the content provided.

Future Directions

• Scaling hyperbolic temporal models. The related work section explicitly names scaling these models and adapting curvature over long sequences as open difficulties, which the GRU-based kernel evolution only partially addresses.
• Closing the CIFAR-10 gap. On Kendall's τ the method reached 0.846 versus 0.934 for NFN_HNP, so improving rank fidelity on large-image generalization prediction remains an open question.
• Beyond MLPs. The current temporal graph construction is defined for MLPs with L layers and widths (n_0, …, n_L); extending the temporal hyperbolic treatment to CNNs and Transformers, which the cited graph meta-network work covers, is not demonstrated.
• Completing the sine-wave frequency experiment and interpreting geometry quantitatively. The paper introduces the x ↦ a sin(bx) benchmark and states that an ablation study appears in Appendix J, suggesting further work on which geometric regularizers and loss terms drive the reported gains.

Target Audience

Researchers and graduate students working on meta-learning over neural parameters, permutation-equivariant and graph-based meta-networks, and geometric deep learning, particularly those already comfortable with the Poincaré ball model and Riemannian optimization. It will also interest interpretability researchers looking for intrinsic, dynamics-based explanations rather than post-hoc attribution, and network-science researchers exploring whether parameter graphs exhibit small-world, modular, and heavy-tailed structure. Readers without a differential-geometry background will find the method sections demanding, although the paper points to an appendix on hyperbolic geometry for background.

Authors’ abstract

Intrinsic explainability remains a challenging problem, particularly in contexts where multilayer perceptrons (MLPs) require dynamic re-training within an optimization environment. This paper investigates how MLPs and their training dynamics can be represented and studied in non-Euclidean spaces; our representation features the Poincaré model of hyperbolic geometry. We aim to capture the geometric evolution of their weighted topology and self-organization over time. Instead of restricting the analysis to single checkpoints---as per established measure-based explainability methods---we construct temporal \textit{parameter graphs}, i.e., snapshots over time $T$ steps of the optimization/training process for MLPs. This reflects the view that neural networks encode information not only in their weights but also in the trajectory traced during training. Drawing on the idea that many complex networks admit embeddings in hidden metric spaces where distances correspond to connection likelihood, we present a geometric and temporal graph-based metalearning framework for obtaining dynamic hyperbolic representations of the underlying neural parameter graphs. Our model embeds temporal parameter graphs in the Poincaré model ball, and learns from them while maintaining equivariance to within-snapshot neuron permutations and invariance to permutations of past snapshots. In doing so, the approach preserves functional equivalence over time and recovers the latent evolving geometry of the network. Experiments on regression and classification tasks with trained MLPs show strong meta-network performance, accompanied by hyperbolic temporal representations. This reveals how the network structure emerges over time under specific training environments, thus providing insights into the network's self-organization.