In the meta optimizer, try to fix broadcast pattern change.

theano/gof/opt.py

@@ -898,7 +898,14 @@ class LocalMetaOptimizer(LocalOptimizer):
             timings.sort()
             if self.verbose:
                 print "= %s" % timings[0][2]
-            return timings[0][1]
+            ret = timings[0][1]
+            for i in range(len(ret)):
+                if ret[i].type != node.outputs[i].type:
+                    # If the type is different, it is probably due to
+                    # the broadcast pattern. Try to fix that.
+                    ret[i] = theano.tensor.patternbroadcast(
+                        ret[i], node.outputs[i].broadcastable)
+            return ret
         return
 
     def provide_inputs(self, node, inputs):