diff --git a/symbolic_nn_tests/__main__.py b/symbolic_nn_tests/__main__.py
index ac6d79b..3be6a0b 100644
--- a/symbolic_nn_tests/__main__.py
+++ b/symbolic_nn_tests/__main__.py
@@ -31,6 +31,7 @@ def main():
     run_test(semantic_loss.hasline_cross_entropy, "hasline_cross_entropy")
     run_test(semantic_loss.hasloop_cross_entropy, "hasloop_cross_entropy")
     run_test(semantic_loss.multisemantic_cross_entropy, "multisemantic_cross_entropy")
+    run_test(semantic_loss.garbage_cross_entropy, "garbage_cross_entropy")
 
 
 if __name__ == "__main__":
diff --git a/symbolic_nn_tests/semantic_loss.py b/symbolic_nn_tests/semantic_loss.py
index 7634e4b..ade67ed 100644
--- a/symbolic_nn_tests/semantic_loss.py
+++ b/symbolic_nn_tests/semantic_loss.py
@@ -73,3 +73,11 @@ MULTISEMANTIC_MATRIX = SIMILARITY_MATRIX * HASLINE_MATRIX * HASLOOP_MATRIX
 MULTISEMANTIC_MATRIX /= MULTISEMANTIC_MATRIX.sum()
 
 multisemantic_cross_entropy = create_semantic_cross_entropy(MULTISEMANTIC_MATRIX)
+
+# NOTE: As a final test, lets make something similar to tehse but where there's no knowledge,
+#   just random data. This will create a benchmark for the effects of this process wothout the
+#   "knowledge" component
+GARBAGE_MATRIX = torch.rand(10, 10).to("cuda")
+GARBAGE_MATRIX /= GARBAGE_MATRIX.sum()
+
+garbage_cross_entropy = create_semantic_cross_entropy(GARBAGE_MATRIX)