diff --git a/src/tensor.ts b/src/tensor.ts
index c82d9b5..148761c 100644
--- a/src/tensor.ts
+++ b/src/tensor.ts
@@ -717,7 +717,7 @@ export class Pow {
    */
   forward(a: Tensor, n: number): Tensor {
     // Build cache to use in backward step:
-    this.cache = a;
+    this.cache = [a, n];
 
     // Call recursive function:
     const z = new Tensor(
@@ -737,12 +737,12 @@ export class Pow {
 
   backward(dz: Tensor, z: Tensor) {
     // Get data from cache:
-    const a = this.cache;
+    const [a, n] = this.cache;
 
     // Find gradients relative to "a", and pass them downstream:
     if (requiresGrad(a)) {
-      // d/da(e^a) = e^a, apply the chain rule to the derivative of e^a:
-      const da = new Tensor(_mul(2, _mul(a.data, dz.data)));
+      // d/da(a^n) = na^(n-1), apply the chain rule to the derivative of a^n:
+      const da = new Tensor(_mul(n, _mul(_pow(a.data, n-1), dz.data)));
       a.backward(da, z);
     }
   }
@@ -1725,11 +1725,13 @@ function _matmul(a: Array<any>, b: Array<any>, kernel: any): Array<any> {
 
 function _pow(a: Array<any> | number, n: number): Array<any> | number {
   // If a is a number, exponentiate it. If not, exponentiate all of its elements:
-  let z = a;
-  for (let i = 0; i < n - 1; i++) {
-    z = _mul(z, a);
+  if (typeof a === "number") {
+    return a ** n;
+  } else if (a instanceof Array) {
+    return a.map((element: Array<any>) => _pow(element, n))
+  } else {
+    throw new TypeError("the input data is not a number.");
   }
-  return z;
 }
 
 function _sqrt(a: Array<any> | number): Array<any> | number {