From d2f8f7d0e7dbbf1d9dd2e1601af766ef3b52803b Mon Sep 17 00:00:00 2001
From: tiagofabre <tiagofabre@gmail.com>
Date: Sun, 5 Apr 2026 18:22:31 -0300
Subject: [PATCH] Add binary step and Gaussian activation functions

- Introduced new activation functions: Binary Step and Gaussian.
- Updated the activation functions map to include the new functions.
---
 index.html   |  2 ++
 src/nn.ts    | 16 ++++++++++++++++
 src/state.ts |  4 +++-
 3 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/index.html b/index.html
index 3f6060d6..28cb45e0 100644
--- a/index.html
+++ b/index.html
@@ -98,6 +98,8 @@ <h1 class="l--page">Tinker With a <b>Neural Network</b> <span class="optional">R
             <option value="tanh">Tanh</option>
             <option value="sigmoid">Sigmoid</option>
             <option value="linear">Linear</option>
+            <option value="step">Binary Step</option>
+            <option value="gaussian_z">Gaussian</option>
           </select>
         </div>
       </div>
diff --git a/src/nn.ts b/src/nn.ts
index e92a13de..f597421f 100644
--- a/src/nn.ts
+++ b/src/nn.ts
@@ -134,6 +134,22 @@ export class Activations {
     output: x => x,
     der: x => 1
   };
+  public static GAUSSIAN_Z: ActivationFunction = {
+    output: x => Math.exp(-x * x),
+    der: x => {
+      let o = Math.exp(-x * x);
+      return -2 * x * o;
+    }
+  };
+  /**
+   * Binary (Heaviside) step: 1 if pre-activation z >= 0, else 0.
+   * Derivative is 0 (no gradient through z); only layers above step units
+   * receive useful updates from the output error signal.
+   */
+  public static STEP: ActivationFunction = {
+    output: x => (x >= 0 ? 1 : 0),
+    der: () => 0
+  };
 }
 
 /** Build-in regularization functions */
diff --git a/src/state.ts b/src/state.ts
index 42dc8154..56edf502 100644
--- a/src/state.ts
+++ b/src/state.ts
@@ -24,7 +24,9 @@ export let activations: {[key: string]: nn.ActivationFunction} = {
   "relu": nn.Activations.RELU,
   "tanh": nn.Activations.TANH,
   "sigmoid": nn.Activations.SIGMOID,
-  "linear": nn.Activations.LINEAR
+  "linear": nn.Activations.LINEAR,
+  "step": nn.Activations.STEP,
+  "gaussian_z": nn.Activations.GAUSSIAN_Z
 };
 
 /** A map between names and regularization functions. */