SLeeLa Neural Networks
SLeeLa-facing package for the native CPU neural-network implementation in
artificial-intelligence/neural-nets/. The native C ABI remains authoritative
for allocation, inference, and SGD updates. The SLeeLa classes validate and
retain language-level configuration; they do not pretend to execute native
kernels when no host adapter is installed.
Public units
SLPackage— package identity and native ABI contract.SLNeuralNetworkConfig— layer widths, activation choices, and seed.SLNeuralNetwork— configuration state and inference/training contract.
Native bridge contract
The adapter must verify SLNN_ABI_VERSION == 1, widen the SLeeLa 32-bit seed
to the native uint64_t seed field, convert layer widths to size_t, and
map activation values without reordering them. It owns the opaque
slnn_network * handle and must destroy it exactly once. Map native status
codes 0–5 directly; status 6 is SLeeLa's STATUS_NATIVE_ADAPTER_REQUIRED
and is not a native ABI status.
The adapter creates handles with slnn_network_create, predicts with slnn_network_predict, trains with slnn_network_train_sample, and releases them with slnn_network_destroy.
Compile/link neural_nets.c (C11) and, when using C++, neural_nets.cpp
(C++17). The network is fully connected and CPU-backed. It supports linear,
ReLU, tanh, and sigmoid activations, deterministic seeded initialization,
single-sample inference, and online single-sample MSE/SGD training. A single
network instance is not thread-safe. GPU execution, batching, persistence, and
automatic differentiation are not part of this ABI.
Example
#sleela 3.0
material;
SLNeuralNetworkConfig config = new SLNeuralNetworkConfig();
config.setLayerSizes(new int[]{2, 8, 1});
config.setHiddenActivation(SLNeuralNetworkConfig.ACTIVATION_RELU);
config.setOutputActivation(SLNeuralNetworkConfig.ACTIVATION_LINEAR);
config.setSeed(42);
SLNeuralNetwork network = new SLNeuralNetwork();
int status = network.configure(config);
The example validates and configures the SLeeLa-side contract. Until the
embedding runtime registers the native adapter, predict and trainSample
return STATUS_NATIVE_ADAPTER_REQUIRED.