diff --git a/dlib/dnn/core.h b/dlib/dnn/core.h
index 03a699239..5d1ac529a 100644
--- a/dlib/dnn/core.h
+++ b/dlib/dnn/core.h
@@ -112,23 +112,6 @@ namespace dlib
         }
     }
 
-// ----------------------------------------------------------------------------------------
-
-    inline double log1pexp(double x)
-    {
-        using std::exp;
-        using namespace std; // Do this instead of using std::log1p because some compilers
-                             // error out otherwise (E.g. gcc 4.9 in cygwin)
-        if (x <= -37)
-            return exp(x);
-        else if (-37 < x && x <= 18)
-            return log1p(exp(x));
-        else if (18 < x && x <= 33.3)
-            return x + exp(-x);
-        else
-            return x;
-    }
-    
 // ----------------------------------------------------------------------------------------
 
     // Tell us if T is one of the special layer types (i.e. add_layer, repeat, add_tag_layer, or
@@ -442,24 +425,6 @@ namespace dlib
     {
         return item;
     }
-// ----------------------------------------------------------------------------------------
-
-    inline void randomize_parameters (
-        tensor& params,
-        unsigned long num_inputs_and_outputs,
-        dlib::rand& rnd
-    )
-    {
-        for (auto& val : params)
-        {
-            // Draw a random number to initialize the layer according to formula (16)
-            // from Understanding the difficulty of training deep feedforward neural
-            // networks by Xavier Glorot and Yoshua Bengio.
-            val = 2*rnd.get_random_float()-1;
-            val *= std::sqrt(6.0/(num_inputs_and_outputs));
-        }
-    }
-
 // ----------------------------------------------------------------------------------------
 
     template <typename T>
diff --git a/dlib/dnn/core_abstract.h b/dlib/dnn/core_abstract.h
index 86683d00e..11c7b8600 100644
--- a/dlib/dnn/core_abstract.h
+++ b/dlib/dnn/core_abstract.h
@@ -16,23 +16,6 @@ namespace dlib
 
 // ----------------------------------------------------------------------------------------
 
-    void randomize_parameters (
-        tensor& params,
-        unsigned long num_inputs_and_outputs,
-        dlib::rand& rnd
-    );
-    /*!
-        ensures
-            - This function assigns random values into params based on the given random
-              number generator.  In particular, it uses the parameter initialization method
-              of formula 16 from the paper "Understanding the difficulty of training deep
-              feedforward neural networks" by Xavier Glorot and Yoshua Bengio.
-            - It is assumed that the total number of inputs and outputs from the layer is
-              num_inputs_and_outputs.  That is, you should set num_inputs_and_outputs to
-              the sum of the dimensionalities of the vectors going into and out of the
-              layer that uses params as its parameters.
-    !*/
-
     template <
         typename... T 
         >
@@ -58,15 +41,6 @@ namespace dlib
               a non-std::tuple object is found.
     !*/
 
-    double log1pexp(
-        double x
-    );
-    /*!
-        ensures
-            - returns log(1+exp(x))
-              (except computes it using a numerically accurate method)
-    !*/
-
 // ----------------------------------------------------------------------------------------
 
     template <typename T>
diff --git a/dlib/dnn/layers.h b/dlib/dnn/layers.h
index 5af5ab5df..355c0c044 100644
--- a/dlib/dnn/layers.h
+++ b/dlib/dnn/layers.h
@@ -12,6 +12,7 @@
 #include "../string.h"
 #include "tensor_tools.h"
 #include "../vectorstream.h"
+#include "utilities.h"
 
 
 namespace dlib
diff --git a/dlib/dnn/utilities.h b/dlib/dnn/utilities.h
index 6bee976e0..116d5bbad 100644
--- a/dlib/dnn/utilities.h
+++ b/dlib/dnn/utilities.h
@@ -9,6 +9,41 @@
 namespace dlib
 {
 
+// ----------------------------------------------------------------------------------------
+
+    inline double log1pexp(double x)
+    {
+        using std::exp;
+        using namespace std; // Do this instead of using std::log1p because some compilers
+                             // error out otherwise (E.g. gcc 4.9 in cygwin)
+        if (x <= -37)
+            return exp(x);
+        else if (-37 < x && x <= 18)
+            return log1p(exp(x));
+        else if (18 < x && x <= 33.3)
+            return x + exp(-x);
+        else
+            return x;
+    }
+    
+// ----------------------------------------------------------------------------------------
+
+    inline void randomize_parameters (
+        tensor& params,
+        unsigned long num_inputs_and_outputs,
+        dlib::rand& rnd
+    )
+    {
+        for (auto& val : params)
+        {
+            // Draw a random number to initialize the layer according to formula (16)
+            // from Understanding the difficulty of training deep feedforward neural
+            // networks by Xavier Glorot and Yoshua Bengio.
+            val = 2*rnd.get_random_float()-1;
+            val *= std::sqrt(6.0/(num_inputs_and_outputs));
+        }
+    }
+
 // ----------------------------------------------------------------------------------------
 
     namespace impl
diff --git a/dlib/dnn/utilities_abstract.h b/dlib/dnn/utilities_abstract.h
index 8408d09dd..77e8d05df 100644
--- a/dlib/dnn/utilities_abstract.h
+++ b/dlib/dnn/utilities_abstract.h
@@ -8,6 +8,36 @@
 namespace dlib
 {
 
+// ----------------------------------------------------------------------------------------
+
+    double log1pexp(
+        double x
+    );
+    /*!
+        ensures
+            - returns log(1+exp(x))
+              (except computes it using a numerically accurate method)
+    !*/
+
+// ----------------------------------------------------------------------------------------
+
+    void randomize_parameters (
+        tensor& params,
+        unsigned long num_inputs_and_outputs,
+        dlib::rand& rnd
+    );
+    /*!
+        ensures
+            - This function assigns random values into params based on the given random
+              number generator.  In particular, it uses the parameter initialization method
+              of formula 16 from the paper "Understanding the difficulty of training deep
+              feedforward neural networks" by Xavier Glorot and Yoshua Bengio.
+            - It is assumed that the total number of inputs and outputs from the layer is
+              num_inputs_and_outputs.  That is, you should set num_inputs_and_outputs to
+              the sum of the dimensionalities of the vectors going into and out of the
+              layer that uses params as its parameters.
+    !*/
+
 // ----------------------------------------------------------------------------------------
 
     template <typename net_type>