Improve evaluation function with noise for assignments

Switched from '-1' sentinel value for silent neurons to randomly assigned labels. Also added noise for random tie-breaking in neuron assignments.

Author: Machuka

bindsnet/evaluation/evaluation.py

@@ -35,31 +35,36 @@ def assign_labels(
 
     # Sum over time dimension (spike ordering doesn't matter).
     spikes = spikes.sum(1)
-    
+
     for i in range(n_labels):
-        # Create mask (faster and allows future steps to stay on GPU).
-        mask = (labels == i)
+        # Create mask.
+        mask = labels == i
         # Count the number of samples with this label.
         n_labeled = mask.sum().float()
 
         if n_labeled > 0:
-            # Get indices of samples with this label (masking is faster and stays on the GPU).
+            # Get indices of samples with this label.
             label_sum = spikes[mask].sum(0)
             # Update rates.
             rates[:, i] = alpha * rates[:, i] + (label_sum / n_labeled)
 
-    # Compute proportions (and use 'torch.where' to avoid NaN bug).
+    # Compute proportions of spike activity per class.
     total_activity = rates.sum(1, keepdim=True)
-    proportions = torch.where(total_activity > 0, rates / total_activity, torch.zeros_like(rates))
+    proportions = torch.where(
+        total_activity > 0, rates / total_activity, torch.zeros_like(rates)
+    )
+
+    # Noise for random tie breaking.
+    eps = 1e-6  # Small enough not to distort real decisions
+    noise = eps * torch.randn_like(proportions)
 
     # Neuron assignments are the labels they fire most for.
-    max_vals, assignments = torch.max(proportions, 1)
-    
-    # Set unassigned (silent) neurons to -1 instead of defaulting to 0.
-    assignments[max_vals == 0] = -1 
+    assignments = torch.argmax(proportions + noise, dim=1)
 
     # Neurons that never fired have no class preference; mark them with -1 so
-    # they don't contribute a (biased) vote for class 0 in downstream schemes.
+    # they are masked out of downstream voting (issue #736: silent neurons must
+    # not default to a vote for class 0). Downstream schemes only match
+    # ``assignments == i`` for i >= 0, so -1 neurons are excluded.
     assignments[rates.sum(1) == 0] = -1
 
     return assignments, proportions, rates