pulsatrix::XorNetwork Class Reference

A tiny MLP (Linear(2,4) -> ReLU -> Linear(4,1)) trained on XOR – the canonical not-linearly-separable case, exactly representable by a small MLP, giving an unambiguous convergence target for Phase 1's "prove the training loop works" goal. More...

#include <xor_training_example.hpp>

Public Member Functions
	XorNetwork (DeviceBackend *backend)
	Constructs the network with fixed, non-zero initial weights.
Tensor	forward (const Tensor &input)
	Runs the network forward.
float	train_step (const Tensor &input, const Tensor &target, AdamOptimizer &optimizer, MetricsSink &sink, int step)
	Runs one training step: forward, loss, backward through every layer, one Adam update per layer's parameters, and logs the loss.
const Tensor &	linear1_weight () const
	Test/inspection accessor.
LinearModule &	linear1 ()
	Mutable layer accessors – Population Based Training's own exploit step needs to read/write this network's live weights directly (to copy a top performer's weights onto a bottom performer mid-training, via each LinearModule's own parameters()), which no existing const-only accessor supports. Logged as a small necessary scope addition, found by campaign_exai_dl_library_evolutionary_deep_learning's Phase 4 Mission 0.
LinearModule &	linear2 ()

Detailed Description

A tiny MLP (Linear(2,4) -> ReLU -> Linear(4,1)) trained on XOR – the canonical not-linearly-separable case, exactly representable by a small MLP, giving an unambiguous convergence target for Phase 1's "prove the training loop works" goal.

Note

Modules are chained directly (no ComputationGraph/Autograd involvement) – see mission_training_loop.md's design rationale. Batch size 1 (operator decision) – trains on one example at a time, no batching added to any module.

Deterministic, hand-picked non-zero initial weights, NOT a general Xavier/He scheme. All-zero initialization was tried first during this mission's design and found to break symmetry completely: with W=0 everywhere, every layer's forward output is 0 regardless of input, and every gradient that depends on multiplying by a zero weight or a zero (pre-activation) ReLU input is itself zero – linear1's weights and linear2's weights never receive a gradient at all, only linear2's bias does. Non-zero initial weights are required to break this symmetry; a small, fixed, sign-varied initialization is sufficient for a network this tiny. A general initialization scheme (Xavier/He) remains deferred until a real dataset needs one.

Constructor & Destructor Documentation

XorNetwork()

pulsatrix::XorNetwork::XorNetwork ( DeviceBackend *  backend )

explicit

Constructs the network with fixed, non-zero initial weights.

Parameters

backend

Backend to compute through. Not owned; must outlive this network.

Member Function Documentation

forward()

Tensor pulsatrix::XorNetwork::forward

(

const Tensor &

input

)

Runs the network forward.

Parameters

input

Shape (2,).

Returns

Shape (1,).

linear1()

LinearModule & pulsatrix::XorNetwork::linear1 ( )

inline

Mutable layer accessors – Population Based Training's own exploit step needs to read/write this network's live weights directly (to copy a top performer's weights onto a bottom performer mid-training, via each LinearModule's own parameters()), which no existing const-only accessor supports. Logged as a small necessary scope addition, found by campaign_exai_dl_library_evolutionary_deep_learning's Phase 4 Mission 0.

linear1_weight()

const Tensor & pulsatrix::XorNetwork::linear1_weight ( ) const

inline

Test/inspection accessor.

linear2()

LinearModule & pulsatrix::XorNetwork::linear2 ( )

inline

train_step()

float pulsatrix::XorNetwork::train_step	(	const Tensor &	input,
		const Tensor &	target,
		AdamOptimizer &	optimizer,
		MetricsSink &	sink,
		int	step
	)

Runs one training step: forward, loss, backward through every layer, one Adam update per layer's parameters, and logs the loss.

Parameters

input	Shape (2,).
target	Shape (1,).
optimizer	Optimizer to update this network's parameters with.
sink	Where the loss value is logged (tag "loss").
step	Training step number, passed through to sink.

Returns

The loss value for this example, before the update.

---

The documentation for this class was generated from the following file:

• include/pulsatrix/xor_training_example.hpp