PhotoKAN

Photonic Kolmogorov-Arnold Networks — vendor-agnostic PyTorch framework for photonic AI hardware.

---

Why PhotoKAN?

Standard MLPs use fixed activations on nodes and linear weights on edges. KANs invert this: learnable nonlinear functions sit on the edges, summed at nodes.

Photonic hardware is physically structured around edge-level nonlinear transforms — light through a waveguide produces exactly the kind of parametric nonlinear function a KAN edge needs. This is not an analogy; it is a direct structural match.

Published benchmarks show:

• 43% fewer parameters vs equivalent MLPs
• 46% fewer operations vs equivalent MLPs
• 30x energy efficiency gain on photonic NPU vs CMOS

---

Supported Hardware

Vendor	Technology	SDK	Install
Q.ANT	Thin-Film Lithium Niobate (TFLN)	Q.PAL	pip install photokan[qant]
Lightmatter	Silicon Photonics	Lightmatter SDK	pip install photokan[lightmatter]
Salience Labs	III-V Photonics (InP)	Salience SDK	pip install photokan[salience]
CPU Sim	Physics-accurate noise model	Built-in	pip install photokan

---

Installation

pip install photokan                        # CPU simulation (no hardware required)
pip install photokan[qant]                  # + Q.ANT NPU support
pip install photokan[lightmatter]            # + Lightmatter support
pip install photokan[salience]              # + Salience Labs support
pip install photokan[llm]                   # + HuggingFace / PEFT integration
pip install photokan[onnx]                  # + ONNX export
pip install photokan[dev]                   # + development tools

---

Quick Start

import torch
import photokan as pk

# Works on CPU sim if no hardware present — no code changes needed
model = pk.PhotoKAN(
    layer_sizes=[4, 16, 16, 1],
    activation='sine',   # 'sine' | 'fourier' | 'spline' | 'relu'
    backend='auto',      # auto-detects photonic hardware, falls back to CPU
)

x = torch.randn(32, 4)
y = model(x)

# Standard PyTorch training
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
loss = torch.nn.MSELoss()(y, torch.randn(32, 1))
loss.backward()
optimizer.step()

# Check what hardware is available
print(pk.available_backends())
# -> {'cpu': True, 'cuda': False, 'qant': False, 'lightmatter': False, 'salience': False}

# List registered vendors
print(pk.all_vendor_names())
# -> ['qant', 'lightmatter', 'salience']

# Get vendor-specific noise profiles
print(pk.get_noise_config('qant', 'npu1'))
# -> {'snr_db': 14.0, 'bit_depth': 6, 'phase_noise_rad': 0.02, 'technology': 'tfln', 'enabled': True}

---

Vendor-Specific Dispatch

Target a specific vendor or let the framework auto-detect:

model = pk.PhotoKAN(
    layer_sizes=[4, 16, 1],
    activation='sine',
    backend='qant',        # explicitly use Q.ANT hardware
)

# Or simulate a specific vendor's noise characteristics on CPU
model = pk.PhotoKAN(
    layer_sizes=[4, 16, 1],
    activation='sine',
    backend='cpu',
    noise_config=pk.get_noise_config('lightmatter', 'envise1'),
)

---

Activation Variants

Name	Formula	Best for	Photonic native
sine	sum w*sin(f*x + p)	Periodic targets, photonic deployment	Yes
fourier	a0 + sum [a*cos + b*sin]	Multi-frequency signals	Yes
spline	B-spline basis	Non-periodic, high precision	Via LUT
relu	sum w*ReLU(a*x + b)	Edge inference, speed	Yes

---

Photonic Noise Simulation

Test accuracy against realistic hardware impairments before deploying:

sim = pk.PhotonicSimulator()

# Use vendor-specific noise profile
sim.set_hardware_profile('npu2')   # Q.ANT NPU2: SNR=16dB, 8-bit

results = sim.sweep_snr(model, x_test, y_test,
                         snr_range=[8, 10, 12, 14, 16, 20])
sim.plot_snr_accuracy(results)

Noise Profiles by Vendor

Q.ANT (TFLN)

Profile	SNR	Bit Depth	Phase Noise
npu1	14 dB	6-bit	0.02 rad
npu2	16 dB	8-bit	0.01 rad
ideal	60 dB	16-bit	0.0 rad

Lightmatter (Silicon Photonics)

Profile	SNR	Bit Depth	Thermal Drift
envise1	12 dB	5-bit	0.005
mars1	15 dB	7-bit	0.002
ideal	60 dB	16-bit	0.0

Salience Labs (III-V/InP)

Profile	SNR	Bit Depth	Ring Crosstalk
mr100	18 dB	8-bit	0.003
mr200	22 dB	10-bit	0.001
ideal	60 dB	16-bit	0.0

---

Energy Estimation

Estimate energy consumption with published photonic efficiency numbers:

from photokan.utils import estimate_model_energy

reports = estimate_model_energy(model, batch_size=64)
# Layer 0: 8.214 uJ (CMOS) -> 0.267 uJ (Photonic), 30.8x better
# Layer 1: ...

---

Convolutional KAN

Use PhotoConvKAN for image and spatial workloads:

model = pk.PhotoConvKAN(
    in_channels=1, out_channels=16,
    kernel_size=3, activation='sine',
)
x = torch.randn(8, 1, 28, 28)
y = model(x)  # [8, 16, 28, 28]

---

LLM Integration

Replace MLP layers in HuggingFace transformers with PhotoKAN using LoRA-style adapters:

from photokan.llm import add_photo_lora, PhotoKANAttention

# Wrap a transformer's MLP layers
model = add_photo_lora(model, target_modules=["mlp"], n_basis=4)

---

AOT Compilation

Compile models to photonic deployment bundles (.npu):

compiler = pk.PhotonicCompiler()
program = compiler.compile(model, './my_model.npu')

# CPU validation (LUT interpreter)
y = program.run(x, backend='cpu')

# Hardware inference
y = program.run(x, backend='qant')      # Q.ANT
y = program.run(x, backend='lightmatter')  # Lightmatter

# Benchmark latency
stats = program.benchmark(x)
# -> {'mean_ms': 0.42, 'throughput_samples_per_sec': 76190}

---

Architecture

User PyTorch model
  |- PhotoKAN / PhotoKANLayer / PhotoConvKAN  (nn.Module, drop-in)
       |- EdgeActivations (sine / fourier / spline / relu)
            |- SimBackend -> CPU physics simulation (vendor noise profiles)
            |- Vendor dispatch -> Q.ANT / Lightmatter / Salience via pluggable backends
  |- PhotonicCompiler
       |- LUTCompiler -> int8 quantised lookup tables
       |- GraphBuilder -> photonic execution graph
       |- PhotonicProgram -> run on hardware or CPU LUT interpreter
  |- PhotonicSimulator -> SNR sweeps, transfer functions
  |- LLM integration (LoRA adapters, attention)

---

Adding a New Vendor

PhotoKAN uses a pluggable backend architecture. To add support for a new photonic vendor:

1. Create photokan/backends/<vendor>/__init__.py and backend.py
2. Implement the PhotonicBackend ABC:

from photokan.backends.base import PhotonicBackend
from photokan.backends.registry import register_vendor

class MyBackend(PhotonicBackend):
    @staticmethod
    def name() -> str: ...
    @staticmethod
    def is_available() -> bool: ...
    @staticmethod
    def execute(x, activation, op_type) -> torch.Tensor: ...
    @staticmethod
    def compute_gradient(grad_output, x, activation, op_type) -> tuple: ...
    @staticmethod
    def noise_profiles() -> dict[str, dict]: ...

register_vendor(MyBackend)

3. Add the vendor to _discover_vendors() in registry.py

---

Project Status

Phase	Features	Status
Phase 1	Activations, SimBackend, Layers, Energy, Profiler	Done
Phase 2	LUT compiler, execution graph, ONNX export, ConvKAN	Done
Phase 3	Vendor-agnostic backends (Q.ANT, Lightmatter, Salience)	Done
Phase 4	Hardware dispatch, LLM fine-tuning, arXiv paper	Upcoming

---

Running Tests

pip install -e ".[dev]"
pytest                          # 136+ tests
pytest -k "not slow"            # skip slow tests
pytest --cov=photokan           # with coverage

---

References

• Liu et al. (2024) -- KAN: Kolmogorov-Arnold Networks (arXiv 2404.19756)
• Peng et al. (2024) -- Photonic KAN via RAMZI (98% MNIST, 65x energy-area reduction)
• Reinhardt et al. (2024) -- SineKAN
• Q.ANT NPU -- https://qant.com/photonic-computing/
• Lightmatter -- https://lightmatter.co
• Salience Labs -- https://salience.io

---

PhotoKAN -- Build the bridge. Light does the math.