ResoLUT: Distilling Residual Networks into Lookup-Table Networks
Organizations: Wroclaw University of Science and Technology · Transcendent Logic · Vilnius University · Zurich University of Applied Sciences (ZHAW)
Abstract
Logic networks built from Boolean gates or lookup tables (LUTs) promise efficient hardware inference, but training them at depth and learning their connectivity remain difficult. We do not train logic as logic. Instead, ResoLUT distills a trained residual network (ResNet) layer by layer into LUT trees, preserving its multi-bit residual stream and BatchNorm while extracting connectivity from the teacher rather than learning it. To our knowledge, ResoLUT is the first convolutional logic network in computer vision to retain a ResNet-style multi-bit residual stream. Hard LUT decisions are used throughout training and inference, avoiding a discretization gap in the logic. ResoLUT-20 reaches 95.66% on CIFAR-10 with twice the trees, and ResoLUT-56 reaches 78.87% on CIFAR-100 and 53.69% on ImageNet at 64px, the best results of logic and lookup-table networks on CIFAR-10 and CIFAR-100, showing that LUT networks can scale in depth. We explore and quantify the role of the residuals and the tree budget and shape. The most optimized model, a ResoLUT-8 with one eighth of the trees and further optimizations, fits on an FPGA and reaches 90.04% at 189k images per second, the highest throughput among the compared models above 85% on CIFAR-10.
Figures & tables
| Dataset | Model | Teacher widths | Teacher source | Training time | Teacher acc. | ResoLUT acc. | Gap | LUT4 |
|---|---|---|---|---|---|---|---|---|
| CIFAR-10 | ResoLUT-8 | 128/256/512 | airbench-style net | 0.7 h | 93.42 | 92.32 0.14 | 1.10 | 1.81M |
| ResoLUT-20 | 128/256/512 | 4 WRN-28-10 | 4.6 h | 96.81 | 95.59 0.06 | 1.22 | 6.39M | |
| ResoLUT-32 | 128/256/512 | 4 WRN-28-10 | 10.4 h (2 GPUs) | 96.98 | 95.01 0.12 | 1.97 | 10.38M | |
| ResoLUT-56 | 128/256/512 | 4 WRN-28-10 | 31.0 h (2 GPUs) | 97.10 | 95.11 | 1.99 | 17.68M | |
| CIFAR-100 | ResoLUT-8 | 128/256/512 | 4 WRN-28-10 | 0.8 h | 73.26 | 71.21 0.14 | 2.05 | 2.09M |
| ResoLUT-20 | 192/384/768 | 4 WRN-28-10 | 6.8 h | 81.20 | 78.16 0.12 | 3.04 | 10.00M |
| Method | Type | Acc. (%) | Size (as printed) | Gate operation |
|---|---|---|---|---|
| CIFAR-10 | ||||
| DLGN IWP [ 40 ] | LGN | 57.47 | 512K LUT2 | 1.2M |
| DWN [ 4 ] | LUT | 57.51 | 48K LUT2 | 0.4M |
| LILogic Net [ 13 ] | LGN | 60.98 | 256K LUT2 | 1.1M |
| DiffLogic Net [ 35 ] | LGN | 62.14 | 5.12M LUT2 | 11M |
| DBN-1 [ 25 ] | Bool. | 67.60 | 2.47M LUT2 | 13M |
| Method | Acc. (%) | FPGA | Latency | Images/s |
|---|---|---|---|---|
| ReBNet, 3 residuals [ 16 ] | 86.98 | Zynq-7020 | – | 2.0K |
| H-WNN-L [ 23 ] | 88.23 | Zynq-7020 | 6.84 ms | 0.8K |
| Zhao et al . [ 54 ] | 88.54 | Zynq-7020 | 5.94 ms | 0.17K |
| FBNA CNV [ 18 ] | 88.61 | Zynq-7020 | – | 0.52K |
| FracBNN [ 53 ] | 89.10 | ZU3EG | – | 2.8K |
| Tridgell et al . [ 42 ] | 90.9 | VU9P | 29 s | 122K |
Efficient Neural Network Inference
Efficient neural network inference reduces the time, memory, and energy needed to run trained models across hardware and deployment settings.
See all 507 papers