[BUG] GPU training causes invalid OpenCL state

[xEcho1337]

Bug Description

Using the GPU for medium sized training tasks causes invalid OpenCL state (CL_OUT_OF_RESOURCES).

Steps to Reproduce

1. Create a model
2. Use the MNIST dataset
3. Start the training using the GPU

Expected Behavior

Loss should decrease over time and the GPU should speed up training.

Actual Behavior

Brain4J crashes after few computed batches.
What actually happened (including any error messages, stack traces, or console output).

Environment

• brain4j version: 3.0.0
• Java version: 21
• Operating system: Windows

Additional Context

MLP with 784 -> 128 (ReLU) -> 64 (ReLU) -> 10 (Softmax), MNIST dataset, 128 batch size.

[Adversing]

NOTE: This issue is currently related to NVIDIA GPU ecosystem only.

Tested devices: RTX 4070S, RTX 5070.

[xEcho1337]

NOTE: This issue is currently related to NVIDIA GPU ecosystem only.

Tested devices: RTX 4070S, RTX 5070.

I can confirm, it works fine (almost, there's still a memory leak) on my MacBook Air M3, this makes the issue even stranger.

I'm starting to suspect it's a bug in Windows's OpenCL driver.