My concerns of this paper focus around the weak experimentation validating the claims present within the paper, as well as the lack of proof, methodology or experimental, that the core disentanglement claims are supported.
- No ablation study
There is no ablation study within the paper, which is a major weakness. The architecture is fairly involved and there seems to be 5 different loss functions (reconstruction, future clean prediction, subject invariance, temporal invariance, and adversarial subject cross-entropy) and yet there is no ablation used to demonstrate the importance of them, and the writing makes this fairly unclear that there are that many loss functions. On line 292, it is stated that “Hence, the spectral autoencoder is trained using a weighted combination of clean-prediction loss and reconstruction loss” but the weights are never explained or ablated. In addition to this, the ablations would have been necessary to understand the effects of the other components, such as latent factorization, dilated convolution blocks, the adversarial classifier, setting z^N = 0 during prediction, etc etc…
- No or Unclear validation of the actual factorization
The paper proposes to factorize the latent space into subject-stable characteristics, temporally evolving biological state, and measurement-specific nuisance variation. It is argued that disentanglement was measured with MIG, SAP, DCI were metrics used to calculate the disengagement. However, they are all concatenated together for the calculation…. So it is unclear how they are disentangled across groups. In addition to this, Table 1 and 2 do not demonstrate which subspace the coordinate corresponds to. To validate the paper's core claim, the authors must report a full 33 subspace probing matrix showing the predictability of (Subject, Time, Measurement) from zS alone, zT alone, and zN alone.
- The methodology does not seem to enforce information leakage across specific latent states.
The methodology does not prevent information leakage across the specific latent subspaces. While the paper claims to factorize spectral variation into three distinct semantic subspaces—subject-stable (z^S), temporally evolving (z^T), and measurement nuisance (z^N)—the training losses in Sections 3.3–3.6 only enforce one-way invariance and leave multiple pathways for cross-subspace leakage completely unconstrained. First, nothing prevents static subject identity from leaking into the temporal subspace z^T: because the temporal invariance loss L_T = ‖z^T_(1,a) - z^T_(1,b)‖² + ‖z^T_(2,a) - z^T_(2,b)‖² (Section 3.3, line 257) only compares same-timepoint replicates from the same subject, any static subject-specific feature is already identical across replicates (1,a) and (1,b), meaning z^T can freely encode subject identity with zero penalty in L_T and still minimize the future prediction loss D(z^S₁, ẑ^T₂, 0) (line 305). Second, nothing prevents temporally evolving biological state from leaking into the nuisance subspace z^N: the gradient-reversal adversarial classifier on z^N (Section 3.5, lines 281–287) only penalizes predictability of subject identity, not timepoint t, so z^N can absorb temporal variation to help minimize the individual reconstruction loss D(z^S_(1,a), z^T_(1,a), z^N_(1,a)) (line 314). Third, the predicted future temporal latent ẑ^T₂ = F(z^T₁, t₁, t₂, Δt) (Section 3.4, line 272) is only supervised through the decoder output in spectral space, with no latent-space consistency loss (such as ‖ẑ^T₂ - z^T₂‖²) to guarantee that ẑ^T₂ even remains aligned with the encoder's temporal subspace Z^T. Because the evaluation in Tables 1–2 only reports the single most predictive scalar coordinate across the entire 96-dimensional concatenated vector [z^S; z^T; z^N] rather than probing each subspace separately, there is no theoretical or empirical guarantee that the three subspaces actually isolate their intended factors.