- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper proposes a new model for a linear dynamical system which, combined with a neural network predicting the initial state of the system from initial observations, is applied to the prediction of future states of the modelled time series. The model itself is a mixture of handcrafted choices, forming an inductive bias, and learned parameters, which I describe below to the best of my understanding.
Up to normalisation and centring, the model is given by and , where is the evolution matrix, the output matrix, the state-transition matrix, the state, the output, and is a graph neural network mapping the initial state (position and velocity) to the state.
The key innovation compared to just learning and is to give them a semi-analytical structure. So where is a learned orthogonal matrix and is a block-diagonal matrix where each block is either capturing a damped oscillatory mode (with learnable frequency and decay parameters) or is a fixed "Jordan block" capturing non-oscillatory dynamics. These are mixed by the learned to let the modes interact (without changing the eigenvalues of the systems, and hence the type of dynamics that they capture). A subset of the oscillatory modes are undamped and mapped to physical rotations of the object modelled.
is structured in a similar manner. For example, in the experiments is a collection of 3D points (atoms in a molecule or joints in a human body). In this case, for the rotational mode above, the corresponding components of are set analytically using Rodrigues formulas, corresponding to an exact rotation of the 3D object. Other non-oscillatory components are mapped to exact , , and translations of the overall object. Other modes (such as the residual collective motion) are subspaces of the PCA space of the data, of dimension one or two, depending on whether they are associated to an oscillatory or non-oscillatory mode. These modes are defined globally and locally for groups of nearby points in , given a mix of oscillatory and non-oscillatory modes, some global and some local, partially analytically defined and partially learned.
The learnable parts (, , and frequencies and damping parameters) are fitted to data end-to-end on a training set of sequences. The model is then applied to a test set of new sequences, evaluated on predicting future from the initial . Experiments on small datasets of molecules and human poses show competitive performance compared to prior work in the area, ablations show the importance of the different modes considered in the model, and further analysis explores their interpretability.
Critical strengths
-
The paper appears to be technically sound.
-
In a world where inductive biases are all but removed from machine learning, this paper tries to introduce a set of sophisticated inductive biases to model complex 3D object dynamics.
-
The biases are interpretable, which allows the authors to diagnose, isolate, and correct specific problems with the model (e.g., in predicting the rotation component of the rotation-dominated motion of certain molecules).
Critical weaknesses
-
On a general level, it is very unclear that working on inductive biases like the ones proposed in the paper is a fruitful endeavour when data-driven general-purpose models have been shown over and over again to ultimately win out (e.g., famously, Sutton's "bitter lesson").
-
As it stands, I don't think the paper convincingly shows a specific advantage compared to more generic models. The proposed model is, in fact, quite complex, which has also the disadvantage of making it less practical for others to use. I would expect a model like this, which proposes a manually-defined system of interacting dynamics, to show a clear advantage over more general purpose solutions at least in some cases, and not just in terms of interpretability. The obvious example is when training data is very limited, where inductive biases could achieve better generalisation of approaches that do not have such biases, but this does not seem to be shown very directly in the experiment (see the questions below for more details).
-
The experimental data is quite limited. For example, in the human body pose sequences, the model is tasked with predicting just five (future time steps. While these are subsampled from a longer sequence of 30 frames, nevertheless this is not pushing the model capabilities much.
-
Results are also a little mixed. Despite focusing on small datasets and exploring complex hand-crafted inductive biases, the model does not uniformly wins out compare to prior work. Specifically, it loses to EGNO on the run sequences (i.e., 50% of the mocap data). This is imputed to a weakness of the proposed model which cannot adapt well to non-linear dynamics where, for instance, the rotation rate of limbs can change during the gait. This seems to be a significant limitation of this approach to modelling dynamic, particularly when more expressive models are available (and are often reliable).
-
The "interaction" model consists of rotating the state space. While this preserves the different dynamical modes, which is what the authors want, it also significantly restricts the "interactions" between these modes which, as far as I understand, are essentially additive in the output.
Questions for authors
-
Can the author provide experiments comparing their CDM with less-structured baselines (e.g,. just PCA + linear dynamics), showing a clear advantage when little training data is available and, presumably, less of an advantage when more training data is included?
-
Can the authors provide some evidence that the model has good generalisation capabilities? For example, if the model is trained on sequences where humans run slowly, would the model generalise to humans that run faster?
-
Can the authors show that the model has an advantage in long-horizon prediction? The model cam predict arbitrarily in the future quite easily and cheaply (just compute , so why not increasing ? Likewise, why subsampling time at all (except for being directly comparable to prior work)? Why not testing hundreds of time steps?
-
What kind of restrictions does using rotations of the state space to model "interactions" pose in practice?
10 comments
Response to Reviewer hzLz (1/7)
1. Predictive Benefits of Structured Dynamics (W1, W2, Q1)
1.1. Accuracy Across Training Data Sizes (W1, W2, Q1).
We vary the number of training samples over 100 / 250 / 500 / 1000; 500 is the setting of the paper. All models use the same data splits and evaluation protocol, and the neural models share the same epoch budget. PCA + linear is the baseline suggested in Q1: Cartesian PCA coefficients evolved by a single linear operator, without a neural network, with rank selected on validation. EGNO is retrained on the same splits. CMD-lite reduces our model's parameter count by 6-9$\times$, bringing it within 2.8-4.2$\times$ of EGNO.
Table 1. Test MSE () versus the number of training samples.
Naphthalene
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 0.27 | 0.26 | 0.26 | 0.26 |
| EGNO | 0.40 | 0.37 | 0.35 | 0.34 |
| CMD | 0.31 | 0.26 | 0.22 | 0.20 |
| CMD-lite | 0.32 | 0.26 | 0.23 | 0.20 |
Uracil
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 0.48 | 0.48 | 0.48 | 0.48 |
| EGNO | 0.61 | 0.57 | 0.54 | 0.50 |
| CMD | 0.44 | 0.39 | 0.34 | 0.30 |
| CMD-lite | 0.44 | 0.39 | 0.34 | 0.30 |
Salicylic acid
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 0.57 | 0.57 | 0.56 | 0.56 |
| EGNO | 1.32 | 1.11 | 0.74 | 0.70 |
| CMD | 0.63 | 0.56 | 0.48 | 0.43 |
| CMD-lite | 0.64 | 0.56 | 0.49 | 0.43 |
Toluene
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 4.62 | 4.51 | 4.47 | 4.46 |
| EGNO | 4.88 | 4.88 | 4.86 | 4.84 |
| CMD | 4.50 | 4.06 | 2.42 | 2.48 |
| CMD-lite | 4.39 | 4.09 | 2.70 | 2.00 |
Ethanol
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 3.54 | 3.50 | 3.43 | 3.38 |
| EGNO | 3.32 | 3.30 | 3.27 | 3.26 |
| CMD | 3.39 | 3.12 | 2.69 | 2.41 |
| CMD-lite | 3.40 | 3.09 | 2.71 | 2.56 |
Malonaldehyde
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 11.98 | 11.87 | 11.73 | 11.71 |
| EGNO | 11.01 | 10.90 | 10.83 | 10.77 |
| CMD | 12.06 | 11.39 | 10.70 | 10.25 |
| CMD-lite | 11.91 | 11.37 | 10.81 | 10.15 |
Aspirin
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 7.54 | 7.33 | 7.26 | 7.24 |
| EGNO | 7.73 | 7.37 | 7.03 | 6.91 |
| CMD | 6.69 | 5.68 | 4.36 | 3.24 |
| CMD-lite | 6.77 | 5.87 | 4.35 | 3.14 |
Benzene
| Model | 100 | 250 | 500 | 1000 |
|---|---|---|---|---|
| PCA + linear | 3.14 | 2.97 | 2.88 | 2.84 |
| EGNO | 30.91 | 30.91 | 30.80 | 30.80 |
| CMD | 3.16 | 2.51 | 1.70 | 1.33 |
| CMD-lite | 2.69 | 2.54 | 2.22 | 1.41 |
The predictive advantage of CMD increases with training data. From 100 to 1000 samples, the geometric-mean MSE improves by 1.63$\times$ for CMD, compared with 1.16$\times$ for EGNO and 1.04$\times$ for PCA + linear. At 1000 samples, CMD attains 1.57$\times$ lower MSE than PCA + linear and 2.24$\times$ lower than EGNO. Both baselines show smaller improvements as data is added.
The same trend persists in CMD-lite. Despite a 6--9$\times$ reduction in parameters, it closely follows the full model's learning curve: from 100 to 1000 samples its geometric-mean MSE improves by 1.63$\times$, the same factor as the full model. The data-scaling behaviour is therefore not exclusive to the original high-capacity configuration.
At 100 samples CMD is within 1% of PCA + linear, so we do not claim a particular advantage in the lowest-data regime. Sections 1.2 and 1.3 examine the contribution of the structured dictionary and of the individual mode families.
Response to Reviewer hzLz (2/7)
1.2. Structured versus Free Dictionary (W2).
To evaluate the contribution of the structured dictionary, we train a free-dictionary variant whose columns are learned without predefined mode families or per-node organization. We retain the same encoder, coefficient-evolution formulation, auxiliary rigid branch, and training budget, with a matched dictionary-parameter budget. We first compare both models at CMD's learning rate and additionally tune the free dictionary's learning rate on Mocap.
Table 2. Structured versus free dictionary. Test MSE (; MD17 horizon 8, Mocap horizon 5).
| System | CMD | Free (CMD LR) | Free (tuned LR) |
|---|---|---|---|
| Naphthalene | 0.22 | 0.25 | --- |
| Uracil | 0.34 | 0.35 | --- |
| Salicylic acid | 0.48 | 0.53 | --- |
| Toluene | 2.42 | 2.47 | --- |
| Ethanol | 2.69 | 2.83 | --- |
| Malonaldehyde | 10.70 | 11.37 | --- |
| Aspirin | 4.36 | 4.37 | --- |
| Benzene | 1.70 | 4.06 | --- |
| Mocap-Walk | 2.89 | 50.20 | 5.85 |
| Mocap-Run | 17.03 | 136.62 | 22.13 |
The free dictionary underperforms CMD on all ten systems. Across MD17, it has 1.17$\times$ higher geometric-mean MSE. On Mocap, the free dictionary exhibits training instability at CMD's learning rate. Even after tuning its learning rate, CMD maintains lower errors by 2.02$\times$ on Walk and 1.30$\times$ on Run. Using these tuned Mocap results, the free dictionary has 1.25$\times$ higher geometric-mean MSE across all ten systems.
These results demonstrate a consistent predictive advantage of the proposed structured representation beyond interpretability.
1.3. Per-Family Ablation (W2).
To assess the predictive contribution of each mode family, we remove one family at a time and retrain the model from scratch under the same training budget.
Table 3. Per-family ablation. Test MSE (; MD17 horizon 8, Mocap horizon 5).
| System | CMD | Rigid | Collective | Local |
|---|---|---|---|---|
| Naphthalene | 0.22 | 0.23 | 0.52 | 0.29 |
| Uracil | 0.34 | 0.37 | 0.78 | 0.40 |
| Salicylic acid | 0.48 | 0.51 | 1.12 | 0.57 |
| Toluene | 2.42 | 4.91 | 2.98 | 2.63 |
| Ethanol | 2.69 | 3.33 | 2.69 | 2.66 |
| Malonaldehyde | 10.70 | 11.00 | 10.72 | 11.02 |
| Aspirin | 4.36 | 5.28 | 7.33 | 4.32 |
| Benzene | 1.70 | 24.85 | 1.76 | 1.56 |
| Mocap-Walk | 2.89 | 459.54 | 115.10 | 3.58 |
| Mocap-Run | 17.03 | 3541.04 | 465.09 | 21.79 |
Removing the Rigid family increases prediction error in all ten systems, while removing Collective degrades performance in nine of ten. The effects are particularly pronounced on Mocap: removing Rigid increases MSE from 2.89 to 459.54 on Walk and from 17.03 to 3541.04 on Run, while removing Collective yields errors of 115.10 and 465.09, respectively.
Removing Local degrades performance in seven systems, with smaller and system-dependent effects. These results demonstrate the distinct predictive contributions of the three mode families beyond interpretability.
Response to Reviewer hzLz — Overview
We thank the reviewer for the detailed feedback. The central question is whether CMD’s structured inductive biases provide predictive benefits beyond interpretability (W1, W2, Q1).
Predictive benefits. Our additional experiments show that they do. CMD outperforms a free-dictionary variant across all ten systems (Section 1.2, W2), and the individual mode families each contribute to prediction accuracy (Section 1.3, W2). The advantage over less-structured baselines grows as training data increases (Section 1.1, W1, W2, Q1), and this trend persists in the substantially smaller CMD-lite (Section 1.1).
Temporal modelling. Beyond predictive accuracy at supervised time steps, CMD supports direct prediction at intermediate times without retraining (Section 2.1, Q3) and extrapolation beyond the training horizon without autoregressive rollout (Section 2.2, W3, Q3).
Limitations and expressivity. We also report where the structure does not help. A 1.53× shift in running speed remains challenging (Section 3, Q2), and the systems where CMD is not best are those dominated by rigid motion, where accuracy is limited by how the initial coefficients are estimated (Section 4, W4). Finally, we quantify both the benefits and the expressivity restrictions of cross-family coefficient coupling (Section 5, W5, Q4).