H-LIF Neuron Family — FPGA & CUDA Reference Implementation

Patent reference repository. This repo contains the patent-scoped subset of the Dikmen spiking neuron library, the corresponding custom CUDA kernel for TH-LIF, and the FPGA energy/resource measurements from Arty A7-35T (post-route SAIF).

Patent: TÜRKPATENT application 2026/007632, filed 14.05.2026 "Hazard Function-Based Spiking Neuron Family, Liquid Extension, and N-Island Heterogeneous Neuromorphic Processing System" — owner: İstinye University.

.
├── hlif_neurons/      ← 17 patent-covered neuron implementations (PyTorch)
├── cuda/              ← custom CUDA kernel for TH-LIF (forward + backward)
├── fpga/              ← FPGA validation database + RTL (Arty A7-35T)
├── docs/              ← claim mapping, validation summary
└── examples/          ← minimal usage scripts

pip install -e .

import torch
from hlif_neurons import HLIFRegistry

# Pick any of 17 patent-covered models
neuron = HLIFRegistry.create("TH-LIF", size=128)

# Standard forward pass
x = torch.randn(32, 100, 128)   # (batch, timesteps, features)
spikes, state = neuron(x)        # spikes: (32, 100, 128), state: dict

# List by patent claim
print(HLIFRegistry.list_by_claim("claim_5_physics"))
# ['DB-LIF', 'RT-LIF', 'Thermal-LIF', 'Spin-LIF', 'Piezo-LIF']

cuda/kernels/th_lif_v1.cu is a fused forward+backward kernel for TH-LIF. Compared to a pure-PyTorch implementation, the kernel exploits the analytic natural-gradient property of the hazard formulation (no surrogate function needed) and a single launch per direction.

from cuda.wrappers.th_lif_v1 import CUDATHLIF
layer = CUDATHLIF(n_in=64, n_out=128, lam0=0.1, kappa=2.0, eta=1.5, phi0=1.0)
x = torch.randn(B, T, 64, device="cuda")   # float32 mandatory
spk = layer(x)                              # (B, T, 128)

Constraints: float32 only (AMP/half-precision overflows expf), shape (B, T, N), CUDA device. The reset step zeros the gradient through time, so BPTT does not cross spike events. See cuda/wrappers/th_lif_v1.py for a sanity-check script.

Single-neuron TH-LIF on Digilent Arty A7-35T (Xilinx Artix-7 xc7a35ticsg324-1L):

Arithmetic: Q4.12 signed fixed-point with saturating multiply/add/subtract. Tested on a 4-phase periodic square-wave current stimulus, 128-step cycle, 4 hardware cycles captured via ILA. Zero numerical drift, zero cycle-to-cycle variance.

See fpga/measurements.json for the full citable record. The Verilog RTL (th_lif_top.v, tb_th_lif_top.v) and LUT initialization files (softplus_1024x16.mem, tunneling_1024x16.mem) are in fpga/rtl/ and fpga/mem_inits/ respectively.

If you use this software or reference the patent in academic work, please cite:

@software{dikmen2026hlif,
  author    = {Dikmen, İsmail Can},
  title     = {H-LIF Neuron Family — FPGA \& CUDA Reference Implementation},
  year      = {2026},
  institution = {İstinye University},
  url       = {https://github.com/DrCanD/dikmen-hlif-neuron-family-fpga-cuda},
  license   = {Apache-2.0},
  note      = {Reference implementation for TÜRKPATENT 2026/007632}
}

@misc{dikmen2026patent_hlif,
  author    = {Dikmen, İsmail Can},
  title     = {Hazard Function-Based Spiking Neuron Family, Liquid Extension,
               and N-Island Heterogeneous Neuromorphic Processing System},
  year      = {2026},
  note      = {TÜRKPATENT application 2026/007632, filed 14.05.2026},
  institution = {İstinye University}
}

Apache License 2.0. See LICENSE and NOTICE.

The Apache-2.0 license includes an explicit patent grant from contributors to users of the software. Independent of that grant, the TÜRKPATENT application 2026/007632 is the property of İstinye University; commercial use of the patented inventions in jurisdictions where the application proceeds to grant should be coordinated through the İstinye University Technology Transfer Office (İstinye TTO).

İsmail Can Dikmen — Assistant Professor, Electrical and Electronics Engineering, İstinye University, Istanbul, Türkiye.

• ORCID: 0000-0002-7747-7777
• Web: Google Scholar
• IEEE Senior Member