Spectral Quantum Memory for Implicit Neural Representations
Organizations: Department of Computer Science Ben Gurion University · Department of Electrical Engineering Ben Gurion University
Abstract
Implicit neural representations (INRs) model signals as continuous functions from coordinates to values, but their performance depends strongly on how spectral capacity is allocated. Yet most Quantum INR (QINR) architectures treat spectral components as independent outputs of separate quantum blocks. We introduce the Spectral Quantum Memory Model (SQMM), a memory coupled quantum INR in which sequential hybrid data reuploading layers (bands) communicate through a shared coherent quantum memory register. Each band reads the memory, computes on its own coordinate upload, and writes part of its state back, so later bands can build on the computation of earlier ones. When memory read and write layers are input independent, we prove that the memory state forms an operator valued Fourier series whose support is contained in the Minkowski sum of the initial memory spectrum and the local band spectra. Downstream bands then condition their Fourier coefficients on the spectral content stored by earlier bands, and the memory coupling creates no new primitive frequencies. This theory motivates a spectral memory alignment loss that encourages memory to encode the spectrum expressed by the deployed prefix of the model. Across audio and image reconstruction benchmarks, SQMM improves over compact QINR and INR baselines.
Figures & tables
| QIREN | QVF | IGA | FINER | SIREN | MSA-INR | CAFE (enc.) | |||
| Params | 642 | 642 | 649 | 673 | 691 | 673 | 673 | 703 | |
| Cello | 0.48 0.08 | 0.76 0.06 | 6.79 0.35 | 4.87 0.01 | 3.73 0.44 | 5.02 0.08 | 7.04 1.23 | 3.60 0.60 | |
| Params | 654 | 654 | 657 | 671 | 687 | 697 | 687 | 697 | 709 |
| Astronaut | 1.12 0.10 | 2.86 0.53 | 5.90 0.49 | 12.61 1.65 | 2.01 0.53 | 2.80 0.31 | 2.15 0.22 | 2.03 0.28 | 10.46 5.92 |
| Camera | 0.38 0.08 | 1.27 0.16 | 2.19 0.31 | 4.43 1.38 | 1.07 0.46 | 1.96 0.62 | 0.76 0.47 | 1.14 0.14 | 3.36 1.41 |
| Coffee | 0.78 0.05 | 1.79 0.16 | 2.52 0.67 | 14.38 6.80 | 1.63 0.42 | 1.85 0.31 | 2.26 0.40 | 1.37 0.19 | 5.17 1.06 |
| Setting | Params | MSE ( ) |
|---|---|---|
| No memory, local qubits | 650 | 0.53 0.06 |
| Memory | 642 | 0.52 0.08 |
| 642 | 0.54 0.06 | |
| (Ours) | 642 | 0.48 0.08 |
| Cello | Astronaut | Camera | Coffee | |||||
|---|---|---|---|---|---|---|---|---|
| Metric | No loss | No loss | No loss | No loss | ||||
| MSE ( ) | 0.52 0.08 | 0.48 0.08 | 1.23 0.21 | 1.12 0.10 | 0.42 0.06 | 0.38 0.08 | 0.88 0.11 | 0.78 0.05 |
| PSNR (dB) | 32.85 0.67 | 33.27 0.72 | 35.34 0.71 | 35.66 0.39 | 39.83 0.66 | 40.28 0.94 | 36.60 0.54 | 37.12 0.28 |
| LSD (1D) | 4.70 0.25 | 4.45 0.14 | ||||||
| SSIM ( ) (2D) | 99.23 0.10 | 99.32 0.07 | 98.21 0.26 | 98.45 0.32 | 98.93 0.10 | 99.06 0.04 | ||
| LPIPS ( ) (2D) | 3.54 0.84 | 2.64 0.66 | 3.14 1.26 | 2.82 0.90 | 3.55 0.99 | 3.40 0.92 | ||
| SQMM ⋆ | QIREN | FINER | SIREN | NeRF | MSA-INR | IGA | CAFE (enc.) | |
|---|---|---|---|---|---|---|---|---|
| Params | 642 | 649 | 691 | 673 | 698 | 673 | 673 | 703 |
| GTZAN | 4.71 0.17 | 44.19 1.88 | 5.14 1.06 | 8.65 0.61 | 8.89 0.44 | 42.11 3.39 | 8.45 0.56 | 6.42 1.34 |
| US8K | 6.11 0.27 | 51.83 0.86 | 8.56 2.35 | 15.55 1.49 | 13.96 0.72 | 52.21 1.55 | 15.47 1.22 | 10.83 3.75 |
| VCTK | 1.00 0.03 | 10.70 0.10 | 1.66 0.21 | 2.98 0.15 | 2.59 0.15 | 8.51 0.90 | 2.69 0.13 | 2.25 0.34 |
| SQMM ⋆ | QIREN+ | QVF | FINER | SIREN | IGA | MSA-INR | |
|---|---|---|---|---|---|---|---|
| Params | 654 | 801 | 671 | 697 | 687 | 687 | 697 |
| SSIM ( ) | 99.21 0.03 | 97.69 0.19 | 92.90 0.23 | 97.43 0.39 | 98.37 0.32 | 98.56 0.27 | 98.59 0.14 |
| LPIPS ( ) | 3.44 0.25 | 30.41 2.98 | 104.84 8.38 | 14.17 3.69 | 8.22 1.89 | 7.64 1.70 | 10.56 1.93 |
| MSE ( ) | 0.46 0.02 | 1.67 0.13 | 5.12 0.17 | 1.45 0.21 | 0.98 0.21 | 0.98 0.18 | 0.90 0.11 |
| Symbol | Meaning |
|---|---|
| , , | coordinate, target value, prediction |
| , , | positional encoding, upload variable of band , one time memory upload |
| , , | number of bands, band index, number of uploads in band |
| , , , | local data register of band and its qubit count, memory register and its qubit count |
| , | new data state of band , memory state after band ( is the initial memory) |
| , , | band channel, input dependent upload layer, input independent (trainable, read, write) block |
| Perturbation | Cello | Astronaut | Camera | Coffee |
|---|---|---|---|---|
| 0.48 0.08 | 1.12 0.10 | 0.38 0.08 | 0.78 0.05 | |
| 0.46 0.11 | 1.46 0.11 | 0.44 0.05 | 0.97 0.32 | |
| 0.56 0.06 | 1.62 0.14 | 0.68 0.05 | 1.10 0.16 | |
| 0.73 0.14 | 2.48 0.56 | 1.03 0.15 | 1.80 0.36 | |
| 1.13 0.19 | 4.89 0.58 | 2.38 0.76 | 3.90 1.00 |
| Setting | Params 1D | Params 2D | Cello | Astronaut | Camera | Coffee |
|---|---|---|---|---|---|---|
| No memory qubits | 534 | 544 | 0.82 0.06 | 2.94 0.54 | 0.89 0.07 | 1.59 0.42 |
| No memory qubits, wider decoder | 636 | 646 | 0.65 0.08 | 1.94 0.18 | 0.73 0.20 | 1.33 0.21 |
| No memory, local qubits | 650 | 662 | 0.53 0.06 | 1.41 0.18 | 0.47 0.09 | 0.80 0.06 |
| Read only memory | 626 | 638 | 0.61 0.13 | 1.52 0.09 | 0.48 0.09 | 1.00 0.21 |
| Read and write memory | 642 | 654 | 0.52 0.08 | 1.23 0.21 | 0.42 0.06 | 0.88 0.11 |
| Memory with | 642 | 654 | 0.54 0.06 | 1.20 0.15 | 0.42 0.07 | 0.88 0.08 |
| Setting | Cello | Astronaut | Camera | Coffee |
|---|---|---|---|---|
| SQMM ⋆ , memory on | 0.48 0.08 | 1.12 0.10 | 0.38 0.08 | 0.78 0.05 |
| Memory switched off at evaluation | 17.05 4.65 | 442.40 194.67 | 110.84 37.98 | 113.69 50.65 |
| Trained without memory qubits | 0.82 0.06 | 2.94 0.54 | 0.89 0.07 | 1.59 0.42 |
| Boundary | Cosine | JS | Coordinate spread | Write innovation |
|---|---|---|---|---|
| 1 | 0.75 | 0.12 | 1.64 | 0.21 |
| 2 | 0.82 | 0.11 | 1.26 | 0.15 |
| 3 | 0.82 | 0.14 | 1.02 | 0.08 |
| 4 | 0.82 | 0.17 | 0.80 | 0.09 |
| Shots | Cello | Astronaut | Camera | Coffee |
|---|---|---|---|---|
| 16 | 3.65 0.23 | 18.39 1.76 | 7.64 0.78 | 10.04 0.45 |
| 32 | 3.05 0.12 | 12.24 0.82 | 5.89 0.84 | 8.15 1.01 |
| 64 | 2.30 0.17 | 9.50 0.70 | 4.07 0.29 | 5.82 0.43 |
| 128 | 1.81 0.14 | 7.86 0.43 | 3.38 0.26 | 4.61 0.57 |
| 256 | 1.39 0.08 | 6.17 0.57 | 2.63 0.44 | 3.76 1.14 |
| 512 | 1.12 0.07 | 4.70 0.60 | 2.21 0.38 | 3.00 0.35 |
| Component | Setting |
|---|---|
| Bands | 5 bands of 2 data qubits with 2, 3, 4, 4 and 5 uploads |
| Encoding | the raw coordinate and 16 dyadic sinusoidal frequencies uploaded through in 1D, the raw coordinates and 8 consecutive frequencies per axis uploaded through and with a shared bias in 2D |
| Memory | 2 qubits coupled to every data qubit, no entangler of its own, initialized by and uploads of the encoding |
| Read | controlled before the first upload of every band, an extra controlled in band 5 |
| Write | couplings in bands 1 to 4, none in band 5 |
| Readout | Pauli and expectations of every band, concatenated |
| 1D (audio) | 2D (images) | |||
|---|---|---|---|---|
| Model | Shape | lr | Shape | lr |
| FINER | ||||
| SIREN | ||||
| NeRF | ||||
| KAN | ||||
| WIRE | ||||
| Model | Params 1D | Params 2D | Cello | Astronaut | Camera | Coffee |
|---|---|---|---|---|---|---|
| SQMM ⋆ | 642 | 654 | 0.38 | 1.01 | 0.28 | 0.73 |
| SQMM | 642 | 654 | 0.71 | 2.36 | 1.09 | 1.57 |
| QIREN | 649 | 657 | 6.25 | 5.34 | 1.80 | 1.73 |
| QVF | 671 | 11.00 | 3.22 | 7.76 | ||
| QIREN+ | 793 | 801 | 5.66 | 3.17 | 1.39 | 1.31 |
| FINER | 691 | 697 | 3.32 | 2.49 | 1.25 | 1.54 |
| Model | Params | GTZAN | US8K | VCTK |
|---|---|---|---|---|
| SQMM ⋆ | 642 | 4.71 0.17 | 6.11 0.27 | 1.00 0.03 |
| QIREN+ | 793 | 40.92 1.15 | 50.18 0.34 | 9.73 0.76 |
| QIREN | 649 | 44.19 1.88 | 51.83 0.86 | 10.70 0.10 |
| FINER | 691 | 5.14 1.06 | 8.56 2.35 | 1.66 0.21 |
| SIREN | 673 | 8.65 0.61 | 15.55 1.49 | 2.98 0.15 |
| NeRF | 698 | 8.89 0.44 | 13.96 0.72 | 2.59 0.15 |
| Corpus | MSE ( ) | PSNR (dB) | LSD |
|---|---|---|---|
| GTZAN | 4.71 0.17 | 24.55 0.17 | 7.10 0.10 |
| US8K | 6.11 0.27 | 22.92 0.22 | 5.38 0.08 |
| VCTK | 1.00 0.03 | 33.94 0.17 | 5.57 0.06 |
| Model | Params | MSE ( ) |
|---|---|---|
| SQMM ⋆ | 666 | 32.08 0.15 |
| FINER | 747 | 39.13 0.59 |
| SIREN | 717 | 34.44 0.61 |
| NeRF | 1023 | 34.79 0.40 |
| KAN | 1274 | 54.10 1.35 |
| WIRE | 1383 | 47.58 3.21 |
| Model | Params | native | ||||
|---|---|---|---|---|---|---|
| SQMM ⋆ | 666 | 1.58 0.10 | 5.43 0.03 | 11.54 0.05 | 18.52 0.04 | 32.08 0.15 |
| FINER | 747 | 5.82 0.36 | 10.18 0.26 | 17.32 0.24 | 25.03 0.34 | 39.13 0.59 |
| SIREN | 717 | 2.70 0.33 | 6.89 0.26 | 13.18 0.47 | 20.57 0.55 | 34.44 0.61 |
| NeRF | 1023 | 3.80 0.42 | 8.76 0.68 | 14.94 0.26 | 21.57 0.23 | 34.79 0.40 |
| KAN | 1274 | 19.49 1.72 | 25.53 1.26 | 34.69 1.71 | 42.10 2.41 | 54.10 1.35 |
| WIRE | 1383 | 19.27 1.68 | 52.10 3.73 | 53.46 9.73 | 37.37 3.40 | 47.58 3.21 |
| Model | Params | full | ||||
|---|---|---|---|---|---|---|
| SQMM ⋆ | 666 | 3.13 0.08 | 11.13 0.06 | 23.16 0.34 | 35.15 0.43 | 54.13 0.44 |
| FINER | 747 | 8.59 0.56 | 19.16 0.35 | 32.75 0.69 | 45.71 0.65 | 65.18 0.65 |
| SIREN | 717 | 5.22 0.74 | 14.37 0.75 | 27.07 1.02 | 39.34 0.74 | 58.41 1.14 |
| NeRF | 1023 | 6.91 0.20 | 18.57 0.66 | 30.87 0.38 | 42.68 1.26 | 61.08 0.53 |
| KAN | 1274 | 28.79 1.74 | 43.67 2.16 | 60.42 2.71 | 74.62 0.54 | 93.99 2.11 |
| WIRE | 1383 | 27.27 2.02 | 81.35 4.27 | 90.97 8.55 | 66.31 3.86 | 80.39 2.91 |
| Model | Params | SSIM ( ) | LPIPS ( ) | MSE ( ) |
|---|---|---|---|---|
| SQMM ⋆ | 654 | 99.21 0.03 | 3.44 0.25 | 0.46 0.02 |
| QIREN+ | 801 | 97.69 0.19 | 30.41 2.98 | 1.67 0.13 |
| QVF | 671 | 92.90 0.23 | 104.84 8.38 | 5.12 0.17 |
| FINER | 697 | 97.43 0.39 | 14.17 3.69 | 1.45 0.21 |
| SIREN | 687 | 98.37 0.32 | 8.22 1.89 | 0.98 0.21 |
| WIRE | 1311 | 83.21 2.51 | 306.79 62.91 | 18.23 2.68 |
Implicit Neural Representations
Implicit neural representations encode signals and fields as continuous functions that can be queried at arbitrary coordinates and resolutions.
See all 203 papers