acceptodds
Withdrawn from ICLR 2027

Quantum Memory Units for Compact Learning-Memory Separation in Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Memory is a fundamental component of intelligent systems, yet most neural architectures implement memory through dense hidden states, attention caches, or state-space transitions that entangle feature learning and memory storage within the same computational substrate. Inspired by the functional separation between learning and memory in biological systems, we propose Quantum Memory Units (QMUs), a scalable quantum-inspired memory module for hybrid classical-quantum neural networks. In this framework, classical neural networks extract task-relevant features, while parameterized quantum circuits encode compact memory states through interactions between input qubits and memory qubits. We instantiate two forms of QMU: an episodic QMU for retaining temporal histories through entanglement-mediated sequence compression, and a semantic QMU for aggregating feature-level statistical regularities. To couple classical observations with quantum-inspired memory states, we further introduce a Kalman-filter-based fusion mechanism. We evaluate QMU-enhanced networks on five memory-related tasks, including function approximation, Mackey-Glass prediction, permuted sequential MNIST classification, character-trajectory classification, and copy-memory. The results show that QMUs achieve strong performance on selected episodic memory tasks and competitive early-stage learning behavior under compact parameterization. Additional ROC-based evaluations further indicate that QMU representations can provide meaningful discriminative ability on classification tasks.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum memory units, Hybrid classical-quantum neural networks, Sequence modeling, Memory representation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uspuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether a separate, compact memory module can architecturally separate sequence feature extraction from historical information storage. The authors propose Quantum Memory Units (QMUs), assigning a parameterized quantum circuit to each feature dimension. The episodic QMU compresses history, while the semantic QMU aggregates statistical information and combines it with classical features through Kalman-style fusion. Results are reported on function approximation, Mackey–Glass, PSMNIST, character trajectories, and copy-memory, primarily emphasizing early-stage learning performance with relatively few parameters. All quantum circuits are simulated on classical computers.

Critical strengths
  1. Modularizing feature extraction and compact recurrent memory is a research question of some value; assigning memory modules to individual features also facilitates controlled analysis.

  2. The paper attempts to evaluate continuous-history reconstruction, statistical aggregation, and exact symbolic recall separately, rather than evaluating the method on a single task.

  3. The authors explicitly clarify that Hilbert-space dimensionality does not equal directly readable classical storage, disclose the high simulation cost, and do not claim practical quantum speedup or energy advantages.

Critical weaknesses
  1. The main evaluation metrics are insufficiently defined, and their correspondence with the figures is unclear. Tab. 1 is described as reporting early-window averages, but the window ranges and aggregation methods are not specified for each task and model, and neither random seeds nor variance estimates are reported. In Fig. 4, the QMU curves extend only to approximately 230 steps, whereas the other models extend to approximately 1,600 steps, which is inconsistent with the claim of “complete convergence curves” in the contributions. Estimates from the plotted curves suggest that, on Trajectory, the tabulated QMU result is close to its average over approximately the first 230 steps, while some baseline results are close to their averages over the entire approximately 1,600-step range. Moreover, the values in the two episodic columns of Tab. 1 closely match the errors computed from the individual prediction curves in Fig. 3, which does not show training dynamics. Consequently, the claimed early-stage learning advantage cannot currently be evaluated reliably.

  2. The baselines perform anomalously, limiting the informativeness of the comparisons. In Fig. 3(a), the errors of Mamba, LMU, LSTM, and Transformer are all higher than that of an always-zero predictor. On PSMNIST, Mamba and LMU achieve near-chance accuracy within the displayed training range, while LSTM and RC also achieve low accuracy (Fig. 4(a)); however, the paper does not establish whether these baseline configurations and training budgets provide an effective comparison. On MG, the QMU error estimated from the curves is of the same order as first-order persistence, but the paper does not report such simple baselines. On the copy task, QMU and several baselines remain at similarly low accuracy levels, without a memory-free reference, so these results do not establish that QMU provides long-term retention.

  3. Key controls needed to attribute the results to QMU are missing. The paper does not compare three conditions under the same classical backbone: removing QMU, replacing it with a small classical recurrent module with a similar parameter count, and removing Kalman fusion. For the structure in Fig. 1(b), if a fresh input qubit independent of the previous memory is prepared at each step, discarded after the interaction, and only local expectation values of the memory qubits are read out, these local readouts can be computed exactly through a recurrence involving one three-dimensional Bloch vector per memory qubit—a very small classical recurrence. The authors need to explain what empirical benefit this structure provides over an inexpensive classical implementation. Furthermore, the feature extractor itself may be an LSTM, so the observed memory capability cannot be attributed directly to QMU.

  4. The method is insufficiently specified for reproduction. Eq. 4 does not explain what state is retained for the next time step, how the input qubit is reset, or when the memory is initialized. Discarding the input qubit at each step generally produces a mixed state, which also requires reconciliation with the pure-state notation in the equation. Likewise, the paper does not explain how the covariance and observation matrices in Kalman fusion are specified, learned, or propagated.

  5. The basis for the parameter-efficiency comparison is insufficient. The paper sets the classical models’ hidden size to 256 based on the Hilbert-space dimensionality of eight qubits, which does not ensure comparable parameter counts or effective state sizes. Some values in Tab. 2 are also difficult to reconcile with the stated configurations: for a standard, unprojected LSTM with hidden size 256 and fully trainable recurrent weights, the recurrent weights alone exceed the reported 208k parameters; if Mamba’s reported size of 256 refers to its model width, its 6k parameter count is inconsistent with the standard implementation. The paper needs to clarify the exact meaning of hidden size. Without a module-wise parameter breakdown or reasonably configured compact classical baselines, the claim of competitive performance with few parameters is difficult to assess.

  6. The novelty relative to prior work is insufficiently established. Takaki et al. (PRA 2021) already employed a quantum recurrent structure in which some qubits retain history while the others are measured and reset at each step to encode new inputs. The paper needs to explain QMU’s specific additions in state updates, readout, or architecture. Furthermore, separating learning and memory has been explored extensively in classical approaches, including external-memory and retrieval-augmented models, but the paper does not discuss this literature.

Questions for authors
  1. How exactly is each value in Tab. 1 computed, including the window, aggregation unit, and number of runs? Why do the QMU curves in Fig. 4 extend only to approximately 230 steps? Could you provide complete curves and final metrics under the same training budget, together with means and standard deviations over at least three random seeds?

  2. Could you provide pseudocode for the step-by-step state updates, specifying the retained state, how the input qubit is reset, the observables read out, and when the memory is initialized?

  3. Could you add the controls described in W3: removing QMU, replacing it with a classical recurrent module with a similar parameter count, implementing Fig. 1(b) through the exact Bloch recurrence, and removing Kalman fusion from the semantic model?

  4. What are the inputs and outputs for function approximation and MG? For example, does MG involve next-step or 15-step prediction, and are ground-truth historical inputs used? Could you add constant-prediction, persistence, and linear AR baselines, and explain why Mamba and LMU remain near chance on PSMNIST?

  5. Please provide the complete baseline configurations and module-wise parameter counts, explain values such as LSTM’s 208k and Mamba’s 6k parameters, and clarify the specific differences between QMU and Takaki et al. (2021).

Reviewer_p5wVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Quantum Memory Units (QMUs), simulated parameterized quantum circuits operating on features from classical neural networks. Episodic QMUs use multiple memory qubits to represent temporal histories, while semantic QMUs use one memory qubit per feature to aggregate feature-level statistics. Semantic architectures additionally fuse memory-derived priors and current classical features through a Kalman-filter-based update. The evaluation covers function approximation, Mackey–Glass prediction, permuted sequential MNIST, character trajectories, and copy-memory, with LSTM, reservoir computing, LMU, Mamba, and transformer baselines. The main table reports averages over early training windows. The strongest reported gains are on the two continuous-signal tasks, while classification and copy-memory results are mixed. All QMU results use circuit simulation, and the paper explicitly disclaims practical quantum speedup.

Critical strengths
  • Promising episodic results. Table 1 reports function-approximation MSE of 0.002 for QMU10 and Mackey–Glass MSE of 0.001 for QMU4/QMU8, compared with the best reported classical early-window values of 0.066 and 0.004, respectively. These results motivate further investigation of the circuit representation for continuous temporal signals, although they do not establish an advantage at convergence or under matched resources.

  • Explicitly bounded scope. The separation of episodic history compression from semantic aggregation gives the architecture an interpretable design motivation. The paper reports unfavorable classification and copy-memory comparisons as well as favorable results, distinguishes Hilbert-space dimension from readable classical capacity, and acknowledges the substantial simulation cost. These qualifications make its exploratory contribution easier to assess.

Critical weaknesses
  1. The recurrent quantum state is underspecified. Equation (4) gives an expectation-value output from a circuit acting on an input and a previous pure memory state, but does not define the state passed to the next time step. After entanglement, discarding the input generally leaves a mixed memory state. Retaining that reduced state, conditioning on a measurement, and measuring then reencoding define different recurrent models. Sections 3.2–3.4 and the appendices do not specify which rule is implemented, the measured observables, or the input reset procedure. They also leave the memory reset boundaries unclear, although semantic aggregation is described as operating across time or samples. This prevents a reproducible interpretation of the central mechanism, without implying that a valid implementation is impossible or that data leakage has occurred.

  2. The Kalman fusion mechanism lacks an operational specification. Equations (5)–(7) give the gain and posterior updates, but do not specify initialization and propagation of the prior covariance, the observation covariance, or the observation matrix. The text also does not fully explain how these quantities are obtained from the QMU outputs and classical feature observations. Different choices can produce substantially different fusion behavior. Although fusion is not claimed as an independent contribution, it is part of the semantic architecture being evaluated, so these details are needed to reproduce and interpret its results.

  3. The compact-memory benefit is not isolated by controlled comparisons. Comparing classical hidden size 256 with an 8-qubit Hilbert space does not match accessible memory or trainable parameter count, a limitation the paper acknowledges. Table 2 shows large task-dependent differences. For example, Mamba has 6k trainable parameters on both function approximation and Mackey–Glass, compared with 21k and 34k, respectively, for QMU8. No reported parameter-matched classical-memory comparison or shared-backbone ablation isolates QMU from the feature extractor, prediction head, and Kalman fusion. Consequently, the favorable episodic numbers cannot yet establish that the improvement specifically comes from circuit memory or learning-memory separation. This is a concern about attribution of the claimed compact representation benefit, rather than a requirement to demonstrate quantum speedup.

  4. The evaluation protocol is insufficiently precise to assess robustness. Appendix B describes early averages over the first few hundred or thousand batches without exact task-specific windows, the averaging procedure, or uncertainty across independent runs. Appendix A describes Mackey–Glass prediction as either next-step or 15-step, without identifying which setting produces Table 1. Copy-1 is described as one interval while five fixed intervals are listed in Appendix A, leaving the table's setting unclear. The temporal training/evaluation split is also insufficiently described. Figures 4–5 contain training dynamics, but do not replace final held-out results under a fully specified protocol. These omissions limit confidence in the claimed early-learning advantage and its stability.

Questions for authors
  1. Operational memory definition. What state is passed from one step to the next in Equation (4)? Please specify its initialization, the input reset or discard operation, measured observables, and the state update after readout. If the input is discarded, is the retained state the reduced density matrix, or is a different rule used? Please state the reset policy between independent sequences or examples and between training and evaluation. These are clarification requests about the implemented method.

  2. Kalman implementation. How are the prior covariance, observation covariance, and observation matrix in Equations (5)–(7) initialized, selected or learned, and propagated? Please specify how the classical features and QMU readouts enter the prior and observation models. These details would determine whether the reported semantic architecture can be reproduced and how much of its behavior comes from fusion.

  3. Controlled evidence for QMU. Are there parameter-matched classical-memory comparisons or shared-backbone ablations replacing QMU, omitting QMU, or omitting Kalman fusion? Please identify any existing controlled evidence that attributes the episodic gains to the proposed memory rather than to different backbones or model sizes. Such evidence could materially change my recommendation.

  4. Exact evaluation protocol. What are the task-specific early-window boundaries, averaging procedure, number of independent runs, and variability? Which prediction horizon produces the Mackey–Glass entry, which retention interval produces Copy-1, and how are temporal train/evaluation segments separated? Please provide final held-out results, if available, to distinguish rapid early learning from performance after convergence.

Reviewer_DuyqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Quantum Memory Units (QMUs), parameterized quantum circuits that serve as memory modules on top of features extracted by a classical network (CNN or LSTM). The design is motivated by the separation of learning and memory in the brain. Each feature dimension gets its own QMU. Its value is normalized, mapped to an angle θ = (π/2)·s̃ and encoded on an "input qubit" (Eq. 1–2), and expectation values measured on "memory qubits" form the memory output (Eq. 4). There are two variants:

Episodic QMU: one input qubit and k memory qubits. In the structured version (Fig. 1b) the input qubit fans out CNOTs to each memory qubit. The permuted version (Fig. 1c) adds CNOT ladders among the memory qubits. The output feeds an FC prediction head. Semantic QMU: one input qubit and one memory qubit per feature (Fig. 1d). Its output is used as a "prior" and fused with the classical feature through a Kalman update (Eq. 5–7) before an MLP classifier.

The models are simulated in DeepQuantum and evaluated on function approximation, Mackey–Glass, permuted sequential MNIST (10% of the data), UCI character trajectories and copy-memory. They are compared against LSTM, reservoir computing, LMU, Mamba and a Transformer, all with hidden size 256. The headline metric (Table 1) is an "early-window average" over the first part of training. QMU is best on the two episodic regression tasks (e.g., function MSE 0.002 vs 0.066 for RC; MG 0.001 vs 0.004). It is behind the Transformer on PSMNIST (0.55 vs 0.82), the worst model on trajectories (0.75), and at the baseline level on copy-memory. The paper also says QMU uses far fewer trainable parameters (Table 2) and reports auxiliary ROC-AUC curves (Fig. 6).

Critical strengths
  • The paper says explicitly and repeatedly that it claims no speedup, energy advantage or hardware results. It says the 2^k Hilbert-space dimension is not readable classical storage, and that the neuroscience analogy is only an analogy (§1, §2.2, App. C–D). It also reports where QMU loses (trajectories, copy-memory) instead of hiding those tasks.
  • Putting a separate memory module on top of a classical feature extractor is easy to understand, and the episodic/semantic split gives two concrete, distinct circuit designs.
  • The five tasks cover smooth regression, chaotic forecasting, long-sequence classification, trajectory classification and symbolic recall, and the paper shows full training curves next to the summary table.
  • Simulation cost is disclosed (tens of hours vs tens of minutes, App. C), which is useful context that many hybrid-QML papers omit.
Critical weaknesses
  • Eq. 4 writes m_i^(t) = M(U_ψ(|φ_i^(t)⟩ ⊗ |h_i^(t−1)⟩)), but M returns expectation values, not a quantum state, and the paper never says what |h_i^(t−1)⟩ is. Possible readings: 1) The memory register's quantum state is carried forward unmeasured, and a fresh input qubit is prepared each step. The old input qubit must then be traced out (a mixed-state channel) or kept (a register that grows with T). 2) The measured expectations are re-encoded as angles at the next step. This would make the QMU a small classical recurrence with a trigonometric nonlinearity. 3) There is no recurrence inside the circuit, and the k+1 RY angles in Fig. 1b encode a window of k+1 values. The RY angles printed in Fig. 1 (e.g., 9.17, 11.3, 11.9) lie far outside the [0, π/2] range of Eq. 1, so they are not single normalized feature encodings as described. Each reading implies a different model, a different cost (with T = 784 for PSMNIST, (a) needs a 784-step coherent circuit), and different conclusions about where the "memory" lives. The paper's claims about compact quantum memory cannot be evaluated until this is specified precisely, ideally with pseudocode for one forward pass over a sequence.

  • In Fig. 1b, the only two-qubit gates are CNOTs from the input qubit q0 to each memory qubit, followed by single-qubit U gates. If the readout is single-qubit Pauli expectations on the memory qubits (Table 2 lists hidden size 8 for QMU8, i.e., 8 outputs), then each reduced state ρ_j depends only on q0's computational-basis populations and qubit j's own initial state. Tracing out q0 removes its coherences, and the CNOTs onto other targets commute with this. So each output is a function of two angles, and the 8 outputs are 8 independent two-qubit functions. Under this readout the 2^8 = 256-dimensional Hilbert space plays no role, and the "scale-matched to 256 hidden units" framing (§4.1, Table 2) misrepresents the model. The permuted circuit (Fig. 1c) does entangle memory qubits, but the paper does not say which circuit gives which result in Table 1.

  • The semantic QMU is a two-qubit circuit, a 4-dimensional state per feature, which a 4×4 complex matrix simulates exactly. The authors should state the measured observables and analyze how much of the Hilbert space the readout can actually reach (e.g., the dimension of the span of measured observables, or the rank of the output Jacobian).

  • Matching on Hilbert-space dimension is the wrong axis. The QMU exposes 8–16 numbers to the downstream head, while the baselines expose 256. The natural compact comparison is classical models (LSTM/GRU, LMU, a unitary RNN à la Arjovsky et al. 2016, a small linear SSM) with hidden size 8–16 and matched parameter counts. Without it, "58k vs 6,950k parameters" (§4.3) says only that the authors chose large baselines, not that the QMU is efficient.

  • Several parameter counts in Table 2 look implausible. LMU has 13.3M–14.5M parameters at hidden size 256 on a scalar regression task. Mamba goes from 6k (function approximation) to 38M (trajectories) at the same hidden size. RC has 132k–1.27M trainable parameters, although in reservoir computing normally only the readout is trained. These need to be explained, because they drive the paper's compactness claim.

  • Some baselines seem not to have learned at all. LMU reaches 0.11 on PSMNIST (chance is 0.10), with AUC 0.52, and RC has AUC 0.49. LMU is known to reach ~97% on psMNIST. Copy-memory accuracies of 0.25–0.46 for every model, and LMU's AUC of 0.44 (below chance), suggest that the task or the training budget is not discriminating between models. No hyperparameters or tuning are reported for any baseline beyond "Adam, default settings". Default Adam on a 13M-parameter LMU, or with an ESN whose spectral radius is not tuned, is not a fair reference.

  • There is no ablation of the QMU itself. Each hybrid model includes a CNN/LSTM feature extractor, an FC head or MLP, and (for semantic tasks) a Kalman fusion step. The paper does not report the same pipeline with (i) the QMU removed, (ii) the QMU replaced by a classical module of the same output width, or (iii) the Kalman fusion replaced by concatenation or addition. Without these it is impossible to attribute any result to the QMU. If the LSTM extractor processes the whole sequence, the "memory" may be held classically, which would undercut the learning-memory-separation claim.

  • Eq. 5–7 are the textbook update, but the paper does not say how P_k^−, R_k and H_k are obtained (fixed, learned or estimated), whether the dimensions of m and c match, or whether P is propagated across steps or samples. With fixed covariances, the update is a fixed convex combination of m and c, i.e., a linear layer. The paper says it is "not an independent contribution", but it sits on the path for every semantic result, so it needs specification and an ablation.* §3.4 says the input qubit interacts with the memory qubit "across time or samples". If the memory state is accumulated across samples (e.g., over a batch or over the training set), then the semantic QMU is a learned dataset-level statistic, closer to a normalization layer than to sequence memory. Test predictions could also depend on the order of evaluation or on batch composition. Please clarify whether the semantic memory is reset per sequence.

  • There is substantial prior work on quantum recurrent and memory models that the paper neither discusses nor compares against: quantum RNNs (Bausch, 2020), QLSTM (Chen et al., 2022), Hybrid QRNNs (Xu et. al, 2024, Xu, 2025) and variational quantum RNNs for temporal data (Takaki et al., 2021), and quantum reservoir computing (Fujii & Nakajima, 2017, and follow-ups), which uses quantum dynamics as fixed memory feeding a trained readout. Li et al. (2023) is cited but not compared.

  • On the classical side, "learning-memory separation" is closely related to external-memory architectures (NTM, DNC, memory networks) and to the LMU, whose fixed linear memory is already separated from its nonlinear component. Unitary RNNs (Arjovsky et al., 2016, cited only for the copy task) are the closest classical analogue of unitary memory dynamics and would be the most informative baseline.

  • Once the quantum circuits are classically simulated, the QMU is a specific classical architecture (small unitary or CPTP recurrences with trigonometric encodings and Pauli readouts). The paper needs to explain what inductive bias this provides that a comparably sized classical module does not, and test that.

  • The introduction says DNN memory capacities are "too limited to accurately capture the entirety of information conveyed within a sequence" and that the brain "achieves a vast memory capacity" by projecting information at lower resolution. Neither claim is substantiated or tested. Fig. 1a labels the qubits "somatic" and "synaptic", while the text uses "input" and "memory", and the mapping from dendrites, axon and spikes to the circuit is never explained. The paper never connects the episodic/semantic distinction to any measurable property beyond the choice of circuit.

  • The text does not distinguish which gates in Fig. 1 are data encodings and which are trainable. The same RY gate is drawn on every qubit, including the memory qubits.

  • Eq. 2 uses cos θ / sin θ, whereas the standard R_y(θ) gives cos(θ/2) / sin(θ/2). Please state the gate convention.

  • The total qubit count per model (d × (k+1)?) is not stated, nor is whether the per-feature QMUs share parameters.

  • No code is provided, and App. A–B lack learning rates, the feature-extractor architectures, window lengths, and the train/validation/test protocol for every task except PSMNIST.

  • There are typos, e.g., "multi-calss ROC" in Fig. 6.

Questions for authors
  1. Please give pseudocode for one forward pass over a length-T sequence. What exactly is |h^(t−1)⟩? Is the memory register's quantum state carried forward unmeasured, or is it re-encoded from measured expectations? If a fresh input qubit is prepared each step, what happens to the previous one? What is the circuit depth as a function of T for PSMNIST (T = 784) and for the copy-memory sweep (T up to 5,060)?

  2. Which observables are measured on the memory qubits? For the structured circuit in Fig. 1b, do you agree that single-qubit expectations depend only on (θ_0, θ_j)? If so, what role does the 2^k-dimensional state play?

  3. Which circuit (structured or permuted) produced each number in Table 1? What do the RY angles in Fig. 1 (e.g., 9.17, 11.9) represent, given that Eq. 1 maps features into [0, π/2]?

  4. Could you report the same hybrid pipeline (a) without the QMU, (b) with the QMU replaced by a classical recurrent module of matched output width (8–16) and parameter count, and (c) with the Kalman fusion replaced by concatenation?

  5. How do classical baselines with hidden size 8–16 (LSTM, LMU, unitary RNN, ESN with a tuned spectral radius) compare with QMU8 on function approximation and Mackey–Glass?

  6. Please explain the parameter counts in Table 2, in particular 13–14M for LMU, 6k vs 38M for Mamba at the same hidden size, and the trainable-parameter count for RC. How many of QMU8's 21k–233k parameters are circuit parameters and how many are classical?

  7. Exactly which training steps define the "early window" for each task, and how were they chosen? Can you add a table of final (converged) test performance, with mean ± std over at least 3–5 seeds?

  8. For function approximation and Mackey–Glass, what are the inputs, targets, horizon and train/test split? Is the reported MSE on held-out data?

  9. Is the semantic QMU's memory reset per sequence, or accumulated across samples or batches? If the latter, do test predictions depend on the order of evaluation?

  10. How are P^−, R and H in the Kalman fusion set or learned, and is P propagated over time?

  11. At what checkpoint is the ROC-AUC in Fig. 6 computed? How can the TF reach AUC 1.00 on copy-memory with 0.39–0.46 accuracy?

  12. How would the QMU's outputs degrade under finite-shot estimation (e.g., 1k shots) and a realistic noise model? This matters even for a "quantum-inspired" claim, since the motivation is eventual quantum execution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.