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
Figure 1: Sound representation results adapted from QIREN [ 85 ] , showing normalized amplitude over normalized time.
Figure 2: SQMM architecture. Coordinates are mapped to band uploads γi(x) and to the memory upload γM(x) , both of the form of Eq. ( 1 ). Sequential quantum bands share the memory register M and emit Pauli expectation features, which a classical decoder combines. Dashed red arrows mark the training only spectral alignment loss. Numbers follow the Intuition paragraph.
SQMM⋆
SQMM2048⋆
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
Table 1: MSE ↓(×10−3) on sound and image representation tasks, mean ± SD. ⋆ marks our model and its subscript the number of shots. QVF has no audio configuration. Best runs of all models are in Table 15 . Best in bold .
Setting
Params
MSE ↓ ( ×10−3 )
No memory, local qubits
650
0.53 ± 0.06
Memory
642
0.52 ± 0.08
Lout
642
0.54 ± 0.06
Lalign (Ours)
642
0.48 ± 0.08
Table 2: Mechanism ablation on Cello, MSE (×10−3) , mean ± SD. No memory, local qubits replaces the shared memory with two local qubits in every band and shares nothing between bands. All samples are in Table 9 . Best in bold .
Cello
Astronaut
Camera
Coffee
Metric
No loss
Lalign
No loss
Lalign
No loss
Lalign
No loss
Lalign
MSE ↓ ( ×10−3 )
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 ↑ ( ×10−2 ) (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 ↓ ( ×10−4 ) (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
Table 3: SQMM without and with Lalign on the base benchmark, mean ± SD. Cello reports LSD, the images SSIM and LPIPS. Best in bold .
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
Table 4: Audio MSE ↓(×10−3) , mean ± SD. ⋆ marks our model. All models are in Table 16 . Best in bold .
Figure 3: Ground truth (top) and SQMM (bottom), CIFAR10 at 32×32 , Kodak and DIV2K at 128×128 .
SQMM ⋆
QIREN+
QVF
FINER
SIREN
IGA
MSA-INR
Params
654
801
671
697
687
687
697
SSIM ↑ ( ×10−2 )
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 ↓ ( ×10−4 )
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 ↓ ( ×10−3 )
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
Table 5: Grayscale CIFAR10, mean ± SD. ⋆ marks our model. Table 21 has all models. Best in bold .
Symbol
Meaning
x∈Rd , y , y^(x)
coordinate, target value, prediction
PE(x) , γi(x) , γM(x)
positional encoding, upload variable of band i , one time memory upload
n , i , Li
number of bands, band index, number of uploads in band i
Di , qi , M , m
local data register of band i and its qubit count, memory register and its qubit count
ηDi , σM(i)(x)
new data state of band i , memory state after band i ( σM(0) is the initial memory)
Figure 4: Mixed state scaling ablation on Cello. Training time against the number of sequential bands (left) and the number of memory qubits (right). Each point is one run with 600 training epochs and a fixed depth per band.
Perturbation
Cello
Astronaut
Camera
Coffee
σ=0
0.48 ± 0.08
1.12 ± 0.10
0.38 ± 0.08
0.78 ± 0.05
σ=0.01
0.46 ± 0.11
1.46 ± 0.11
0.44 ± 0.05
0.97 ± 0.32
σ=0.03
0.56 ± 0.06
1.62 ± 0.14
0.68 ± 0.05
1.10 ± 0.16
σ=0.1
0.73 ± 0.14
2.48 ± 0.56
1.03 ± 0.15
1.80 ± 0.36
σ=0.3
1.13 ± 0.19
4.89 ± 0.58
2.38 ± 0.76
3.90 ± 1.00
Table 8: Gaussian gate angle noise with SD σ in training and evaluation of the final model, MSE ↓(×10−3) , mean ± SD. σ=0 is the noiseless model.
Figure 5: Reconstructions under gate angle noise of SD σ in training and evaluation. Cello shows samples 400 to 600 with the target in grey and the prediction in red.
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 Lout
642
654
0.54 ± 0.06
1.20 ± 0.15
0.42 ± 0.07
0.88 ± 0.08
Table 9: Mechanism and control ablation on the classic benchmark, MSE ↓(×10−3) , mean ± SD. Params give the live parameters of each setting. No memory qubits removes the memory qubits and their couplings, and No memory, local qubits is the control of Table 2 . Both one qubit memory rows use Lalign , and the added reuploads match the parameter count of SQMM. Best in bold .
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
Table 10: Memory switched off at evaluation on the classic benchmark, MSE ↓(×10−3) , mean ± SD over the five runs of Table 1 . The first two rows evaluate the same trained models, with every read and write coupling set to zero in the second. The last row is the model without memory qubits of Table 9 . ⋆ marks our model.
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
Table 11: Memory anatomy on Astronaut at each band boundary with Lalign .
Shots S
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
Table 12: MSE ↓(×10−3) with expectations estimated from S shots in training and evaluation, mean ± SD.
Figure 6: Reconstructions with expectations estimated from S shots in training and evaluation. Cello shows samples 400 to 600 with the target in grey and the prediction in red.
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 RY in 1D, the raw coordinates and 8 consecutive frequencies per axis uploaded through RY and RX with a shared bias in 2D
Memory
2 qubits coupled to every data qubit, no entangler of its own, initialized by RY and RX uploads of the encoding
Read
controlled RY before the first upload of every band, an extra controlled RX in band 5
Write
XY couplings in bands 1 to 4, none in band 5
Readout
Pauli Z and ZZ expectations of every band, concatenated
Table 13: Settings of SQMM.
1D (audio)
2D (images)
Model
Shape
lr
Shape
lr
FINER
10×6
10−3
24×1
3⋅10−3
SIREN
14×3
10−3
14×3
3⋅10−3
NeRF
17×1
10−2
12×3
3⋅10−2
KAN
2×4
3⋅10−2
2×4
10−2
WIRE
25×2
3⋅10−3
25×2
3⋅10−3
Table 14: Configurations of the classical baselines. Shape is hidden width × hidden layers. TUNER has no 1D configuration since its initialization requires 2D coordinates.
Model
Params 1D
Params 2D
Cello
Astronaut
Camera
Coffee
SQMM ⋆
642
654
0.38
1.01
0.28
0.73
SQMM 2048⋆
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
Table 15: Classic INR benchmark in the best run convention of QIREN, MSE ↓(×10−3) of the best single run. The SQMM rows take the best of the five runs in Table 1 . QIREN+ is QIREN with one extra spectrum layer. ⋆ marks our model. Best in bold .
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
Table 16: Audio reconstruction MSE ↓(×10−3) for all models, mean ± SD. QIREN+ adds one spectrum layer to QIREN. ⋆ marks our model. Best in bold .
Corpus
MSE ↓ ( ×10−3 )
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
Table 17: SQMM audio metrics per corpus, mean ± SD.
Model
Params
MSE ↓ ( ×10−3 )
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
Table 18: Kodak at native resolution ( 768×512 , 24 images) for all models, MSE (×10−3) , mean ± SD. ⋆ marks our model. Best in bold .
Model
Params
322
642
1282
2562
native 768×512
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
Table 19: Kodak resolution ladder (24 images, RGB), MSE ↓(×10−3) , mean ± SD. ⋆ marks our model. Best in bold .
Model
Params
322
642
1282
2562
full 5122
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
Table 20: DIV2K resolution ladder (10 images, RGB), MSE ↓(×10−3) , mean ± SD. The largest size is the 5122 prepared image. ⋆ marks our model. Best in bold .
Figure 7: SQMM reconstructions of Kodak and DIV2K images at 128×128 . Columns show the ground truth, the SQMM reconstruction, and the absolute error.
Figure 8: Band 3 of SQMM, recorded from the trained Cello model. Controlled RY reads from the memory qubits m1,m2 precede the four RY uploads on the data qubits q1,q2 , each followed by an entangling layer, and XY couplings write to the memory at the end of the band.
Model
Params
SSIM ↑ ( ×10−2 )
LPIPS ↓ ( ×10−4 )
MSE ↓ ( ×10−3 )
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
Table 21: Grayscale CIFAR10 metrics for all models, mean ± SD. MSE is on the [−1,1] range, SSIM and LPIPS on unit range images. QIREN+ adds one spectrum layer to QIREN. ⋆ marks our model. Best in bold .
Figure 9: The 100 grayscale CIFAR10 images (left) and their SQMM reconstructions (right).
A central design principle in modern machine learning and artificial intelligence is to align a model's inductive bias with the structure of its input data. For matrix-valued inputs, relevant matrix-level relationships can be characterised through spectral values and spectral subspaces; however, common coordinate-wise rotation-gate data-encoding unitaries used in most quantum machine learning models do not explicitly construct such a matrix-level representation. We introduce Quantum Spectral Models (QSMs), in which we construct the generator of the data-encoding unitary directly from each input matrix. We study three QSM variants based on symmetric, global block, and non-overlapping patch-local block Hamiltonians. Their outputs admit truncated Fourier representations in which input-dependent spectral gaps supply candidate phase carriers, while spectral subspaces help determine their coefficients. We evaluate the QSMs and comparison quantum models on two matrix representations of Pendigits and two controlled synthetic tasks defined by spectral statistics. At the largest evaluated circuit depth, QSM variants lead the tested quantum models in mean test accuracy across all four benchmarks. The patch-local QSM leads on Pendigits, whereas the global block-Hamiltonian QSM leads on the controlled spectral tasks. Ablations show a task-dependent reversal: subspace-preserving controls perform better on Pendigits, whereas spectral-value-only controls lead among the tested ablations on the synthetic tasks. Together, these results shed new light on quantum machine-learning model design by showing how input-conditioned spectral representations can provide an analysable inductive bias, while offering a broader perspective on structure-aware model design in machine learning and artificial intelligence.
Peiyong Wang, Udaya Parampalli, Casey R. Myers
CSIRO Technology · Research Way, Clayton VIC 3168, Australia · School of Computing and Information Systems +6
We study sinusoidal recurrence as an iterative mechanism for harmonic spectral enrichment in implicit neural representations (INRs). Our analysis reveals that sinusoidal activations induce a harmonic line spectrum, providing a spectral account of how recurrent unrolling enriches the effective spectral support. We realize this principle with a shared sinusoidal block that iteratively refines the latent representation. We empirically validate the resulting spectral behavior against feed-forward INRs, non-sinusoidal recurrent variants, and equilibrium-style sinusoidal models. Complementing this analysis, we evaluate the proposed architecture across image and 3D representation tasks. On RGB image benchmarks, our method achieves higher fidelity than feed-forward baselines with fewer parameters and fewer optimization steps, and it further transfers favorably to super-resolution, NeRF, and SDF tasks.
Hyunmin Cho, Jaejun Yoo, Kyong Hwan Jin
Department of Electrical Engineering, Korea University, Seoul, South Korea · Graduate School of Artificial Intelligence, UNIST, Ulsan, South Korea
Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, and how their accuracy depends on their size is known only empirically. We introduce FUTON (Fourier Tensor Network), which models signals as generalized Fourier series whose coefficients are parameterized by a low-rank tensor decomposition, read out by a linear or MLP head. FUTON combines complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure. Beyond universal approximation, we show that with a linear head, the best FUTON's error splits exactly into the truncation of the signal's series and the best low-rank error of its coefficients, which we bracket from the samples before training, with a FUTON that attains the upper bound. On image fitting, occupancy volumes, and novel view synthesis, against twelve INRs of matched size, FUTON achieves the best average quality, reaches the final quality of each task's strongest baseline 4-34x sooner than that baseline, infers fastest, and trains in the least memory.