diff --git a/GPy/likelihoods/EP.py b/GPy/likelihoods/EP.py
index 3e975436..c52fd8bf 100644
--- a/GPy/likelihoods/EP.py
+++ b/GPy/likelihoods/EP.py
@@ -18,12 +18,10 @@ class EP:
         self.likelihood_function = likelihood_function
         self.epsilon = epsilon
         self.eta, self.delta = power_ep
-        self.jitter = 1e-12 # TODO: is this needed?
+        self.is_heteroscedastic = True
 
-        """
-        Initial values - Likelihood approximation parameters:
-        p(y|f) = t(f|tau_tilde,v_tilde)
-        """
+        #Initial values - Likelihood approximation parameters:
+        #p(y|f) = t(f|tau_tilde,v_tilde)
         self.tau_tilde = np.zeros(self.N)
         self.v_tilde = np.zeros(self.N)
 
@@ -32,8 +30,11 @@ class EP:
         mu_tilde = self.v_tilde/self.tau_tilde #When calling EP, this variable is used instead of Y in the GP model
         sigma_sum = 1./self.tau_ + 1./self.tau_tilde
         mu_diff_2 = (self.v_/self.tau_ - mu_tilde)**2
-        Z_ep = np.sum(np.log(self.Z_hat)) + 0.5*np.sum(np.log(sigma_sum)) + 0.5*np.sum(mu_diff_2/sigma_sum) #Normalization constant
-        self.Y, self.beta, self.Z = self.tau_tilde[:,None], mu_tilde[:,None], Z_ep
+        self.Z = np.sum(np.log(self.Z_hat)) + 0.5*np.sum(np.log(sigma_sum)) + 0.5*np.sum(mu_diff_2/sigma_sum) #Normalization constant, aka Z_ep
+
+        self.Y =  mu_tilde[:,None]
+        self.precsion = self.tau_tilde
+        self.covariance_matrix = np.diag(1./self.precision)
 
     def fit_full(self,K):
         """
diff --git a/GPy/likelihoods/Gaussian.py b/GPy/likelihoods/Gaussian.py
index fe954b78..5b461537 100644
--- a/GPy/likelihoods/Gaussian.py
+++ b/GPy/likelihoods/Gaussian.py
@@ -2,6 +2,7 @@ import numpy as np
 
 class Gaussian:
     def __init__(self,data,variance=1.,normalize=False):
+        self.is_heteroscedastic = False
         self.data = data
         self.N,D = data.shape
         self.Z = 0. # a correction factor which accounts for the approximation made
@@ -19,6 +20,7 @@ class Gaussian:
         self.YYT = np.dot(self.Y,self.Y.T)
         self._set_params(np.asarray(variance))
 
+
     def _get_params(self):
         return np.asarray(self._variance)
 
@@ -27,7 +29,8 @@ class Gaussian:
 
     def _set_params(self,x):
         self._variance = x
-        self.variance = np.eye(self.N)*self._variance
+        self.covariance_matrix = np.eye(self.N)*self._variance
+        self.precision = 1./self._variance
 
     def fit(self):
         """
diff --git a/GPy/likelihoods/likelihood_functions.py b/GPy/likelihoods/likelihood_functions.py
index b94929d3..756d9eb0 100644
--- a/GPy/likelihoods/likelihood_functions.py
+++ b/GPy/likelihoods/likelihood_functions.py
@@ -8,18 +8,18 @@ import scipy as sp
 import pylab as pb
 from ..util.plot import gpplot
 
-class likelihood:
+class likelihood_function:
     """
     Likelihood class for doing Expectation propagation
 
     :param Y: observed output (Nx1 numpy.darray)
-    ..Note:: Y values allowed depend on the likelihood used
+    ..Note:: Y values allowed depend on the likelihood_function used
     """
     def __init__(self,location=0,scale=1):
         self.location = location
         self.scale = scale
 
-class probit(likelihood):
+class probit(likelihood_function):
     """
     Probit likelihood
     Y is expected to take values in {-1,1}
@@ -29,7 +29,7 @@ class probit(likelihood):
     $$
     """
     def __init__(self,location=0,scale=1):
-        likelihood.__init__(self,Y,location,scale)
+        likelihood_function.__init__(self,Y,location,scale)
 
     def moments_match(self,data_i,tau_i,v_i):
         """
@@ -64,7 +64,7 @@ class probit(likelihood):
     def _log_likelihood_gradients():
         return np.zeros(0) # there are no parameters of whcih to compute the gradients
 
-class poisson(likelihood):
+class poisson(likelihood_function):
     """
     Poisson likelihood
     Y is expected to take values in {0,1,2,...}
@@ -75,7 +75,7 @@ class poisson(likelihood):
     """
     def __init__(self,Y,location=0,scale=1):
         assert len(Y[Y<0]) == 0, "Output cannot have negative values"
-        likelihood.__init__(self,Y,location,scale)
+        likelihood_function.__init__(self,Y,location,scale)
 
     def moments_match(self,i,tau_i,v_i):
         """
@@ -160,7 +160,7 @@ class poisson(likelihood):
         if Z is not None:
             pb.plot(Z,Z*0+pb.ylim()[0],'k|',mew=1.5,markersize=12)
 
-class gaussian(likelihood):
+class gaussian(likelihood_function):
     """
     Gaussian likelihood
     Y is expected to take values in (-inf,inf)
diff --git a/GPy/models/sparse_GP.py b/GPy/models/sparse_GP.py
index 7b043209..fe7bcc3b 100644
--- a/GPy/models/sparse_GP.py
+++ b/GPy/models/sparse_GP.py
@@ -6,37 +6,36 @@ import pylab as pb
 from ..util.linalg import mdot, jitchol, chol_inv, pdinv
 from ..util.plot import gpplot
 from .. import kern
+from ..inference.likelihoods import likelihood
 from GP import GP
 
-
 #Still TODO:
 # make use of slices properly (kernel can now do this)
 # enable heteroscedatic noise (kernel will need to compute psi2 as a (NxMxM) array)
 
 class sparse_GP(GP):
     """
-    Variational sparse GP model (Regression)
+    Variational sparse GP model
 
     :param X: inputs
     :type X: np.ndarray (N x Q)
-    :param Y: observed data
-    :type Y: np.ndarray of observations (N x D)
+    :param likelihood: a likelihood instance, containing the observed data
+    :type likelihood: GPy.likelihood.(Gaussian | EP)
     :param kernel : the kernel/covariance function. See link kernels
     :type kernel: a GPy kernel
-    :param Z: inducing inputs (optional, see note)
-    :type Z: np.ndarray (M x Q) | None
     :param X_uncertainty: The uncertainty in the measurements of X (Gaussian variance)
     :type X_uncertainty: np.ndarray (N x Q) | None
+    :param Z: inducing inputs (optional, see note)
+    :type Z: np.ndarray (M x Q) | None
     :param Zslices: slices for the inducing inputs (see slicing TODO: link)
     :param M : Number of inducing points (optional, default 10. Ignored if Z is not None)
     :type M: int
-    :param beta: noise precision. TODO> ignore beta if doing EP
-    :type beta: float
     :param normalize_(X|Y) : whether to normalize the data before computing (predictions will be in original scales)
     :type normalize_(X|Y): bool
     """
 
-    def __init__(self,X,Y=None,kernel=None,X_uncertainty=None,beta=100.,Z=None,Zslices=None,M=10,normalize_X=False,normalize_Y=False,likelihood=None,method_ep='DTC',epsilon_ep=1e-3,power_ep=[1.,1.]):
+    def __init__(self,X,likelihood,kernel, X_uncertainty=None, Z=None,Zslices=None,M=10,normalize_X=False):
+        self.scale_factor = 1000.0# a scaling factor to help keep the algorithm stable
 
         if Z is None:
             self.Z = np.random.permutation(X.copy())[:M]
@@ -52,140 +51,91 @@ class sparse_GP(GP):
             self.has_uncertain_inputs=True
             self.X_uncertainty = X_uncertainty
 
-        GP.__init__(self, X=X, Y=Y, kernel=kernel, normalize_X=normalize_X, normalize_Y=normalize_Y,likelihood=likelihood,epsilon_ep=epsilon_ep,power_ep=power_ep)
+        GP.__init__(self, X, Y, kernel=kernel, normalize_X=normalize_X, Xslices=Xslices)
 
         #normalise X uncertainty also
         if self.has_uncertain_inputs:
             self.X_uncertainty /= np.square(self._Xstd)
 
-        if not self.EP:
-            self.trYYT = np.sum(np.square(self.Y))
-        else:
-            self.method_ep = method_ep
-
-        #normalise X uncertainty also
-        if self.has_uncertain_inputs:
-            self.X_uncertainty /= np.square(self._Xstd)
-
-    def _set_params(self, p):
-        self.Z = p[:self.M*self.Q].reshape(self.M, self.Q)
-        if not self.EP:
-            self.beta = p[self.M*self.Q]
-            self.kern._set_params(p[self.Z.size + 1:])
-        else:
-            self.kern._set_params(p[self.Z.size:])
-            if self.Y is None:
-                self.Y = np.ones([self.N,1])
-        self._compute_kernel_matrices()
-        self._computations()
-
-    def _get_params(self):
-        if not self.EP:
-            return np.hstack([self.Z.flatten(),self.beta,self.kern._get_params_transformed()])
-        else:
-            return np.hstack([self.Z.flatten(),self.kern._get_params_transformed()])
-
-    def _get_param_names(self):
-        if not self.EP:
-            return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + ['noise_precision']+self.kern._get_param_names_transformed()
-        else:
-            return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + self.kern._get_param_names_transformed()
-
-
-    def _compute_kernel_matrices(self):
-        # kernel computations, using BGPLVM notation
-        #TODO: slices for psi statistics (easy enough)
-
-        self.Kmm = self.kern.K(self.Z)
-        if self.has_uncertain_inputs:
-            if not self.EP:
-                self.psi0 = self.kern.psi0(self.Z,self.X, self.X_uncertainty)#.sum() NOTE psi0 is now a vector
-                self.psi1 = self.kern.psi1(self.Z,self.X, self.X_uncertainty).T
-                self.psi2 = self.kern.psi2(self.Z,self.X, self.X_uncertainty)
-                #self.psi2_beta_scaled = ?
-            else:
-                raise NotImplementedError, "uncertain_inputs not yet supported for EP"
-        else:
-            self.psi0 = self.kern.Kdiag(self.X,slices=self.Xslices)#.sum()
-            self.psi1 = self.kern.K(self.Z,self.X)
-            self.psi2 = np.dot(self.psi1,self.psi1.T)
-            self.psi2_beta_scaled = np.dot(self.psi1,self.beta*self.psi1.T)
-
     def _computations(self):
         # TODO find routine to multiply triangular matrices
-        self.V = self.beta*self.Y
+        #TODO: slices for psi statistics (easy enough)
+
+        sf = self.scale_factor
+        sf2 = sf**2
+
+        # kernel computations, using BGPLVM notation
+        self.Kmm = self.kern.K(self.Z)
+        if self.has_uncertain_inputs:
+            self.psi0 = self.kern.psi0(self.Z,self.X, self.X_uncertainty).sum()
+            self.psi1 = self.kern.psi1(self.Z,self.X, self.X_uncertainty).T
+            self.psi2 = self.kern.psi2(self.Z,self.X, self.X_uncertainty)
+            self.psi2_beta_scaled = (self.psi2*(self.beta/sf2)).sum(0)
+        else:
+            self.psi0 = self.kern.Kdiag(self.X,slices=self.Xslices).sum()
+            self.psi1 = self.kern.K(self.Z,self.X)
+            tmp = self.psi1*(np.sqrt(self.likelihood.beta)/sf)
+            self.psi2_beta_scaled = np.dot(tmp,tmp.T)
+
+        self.Kmmi, self.Lm, self.Lmi, self.Kmm_logdet = pdinv(self.Kmm)#+np.eye(self.M)*1e-3)
+
+        self.V = (self.likelihood.beta/self.scale_factor)*self.Y
+        self.A = mdot(self.Lmi, self.psi2_beta_scaled, self.Lmi.T)
+        self.B = np.eye(self.M)/sf2 + self.A
+
+        self.Bi, self.LB, self.LBi, self.B_logdet = pdinv(self.B)
+
         self.psi1V = np.dot(self.psi1, self.V)
         self.psi1VVpsi1 = np.dot(self.psi1V, self.psi1V.T)
-        self.Kmmi, self.Lm, self.Lmi, self.Kmm_logdet = pdinv(self.Kmm)
-        self.A = mdot(self.Lmi, self.psi2_beta_scaled, self.Lmi.T)
-        self.B = np.eye(self.M) + self.A
-        self.Bi, self.LB, self.LBi, self.B_logdet = pdinv(self.B)
-        self.LLambdai = np.dot(self.LBi, self.Lmi)
-        self.LBL_inv = mdot(self.Lmi.T, self.Bi, self.Lmi)
-        self.C = mdot(self.LLambdai, self.psi1V)
-        self.G =  mdot(self.LBL_inv, self.psi1VVpsi1, self.LBL_inv.T)
-        self.trace_K_beta_scaled = (self.psi0*self.beta).sum() - np.trace(self.A)
-        if not self.EP:
-            self.trace_K = self.psi0.sum() - np.trace(self.A)/self.beta
+        self.C = mdot(self.Lmi.T, self.Bi, self.Lmi)
+        self.E = mdot(self.C, self.psi1VVpsi1/sf2, self.C.T)
 
-        # Compute dL_dpsi
-        self.dL_dpsi1 = mdot(self.LLambdai.T,self.C,self.V.T)
-        if not self.EP:
-            self.dL_dpsi0 = - 0.5 * self.D * self.beta * np.ones(self.N)
-            if self.has_uncertain_inputs:
-                self.dL_dpsi2 = - 0.5 * self.beta * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
-            else:
-                self.dL_dpsi2_ = - 0.5 * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
-        else:
-            self.dL_dpsi0 = - 0.5 * self.D * self.beta.flatten()
-            if not self.has_uncertain_inputs:
-                self.dL_dpsi2_ = - 0.5 * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
+        # Compute dL_dpsi # FIXME
+        self.dL_dpsi0 = - 0.5 * self.D * self.beta * np.ones(self.N)
+        self.dL_dpsi1 = mdot(self.V, self.psi1V.T,self.C).T
+        self.dL_dpsi2 = 0.5 * self.beta * self.D * self.Kmmi[None,:,:] # dB
+        self.dL_dpsi2 += - 0.5 * self.beta/sf2 * self.D * self.C[None,:,:] # dC
+        self.dL_dpsi2 += - 0.5 * self.beta * self.E[None,:,:] # dD
 
         # Compute dL_dKmm
-        self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi) # dB
-        self.dL_dKmm += -0.5 * self.D * (- self.LBL_inv - 2.*mdot(self.LBL_inv, self.psi2_beta_scaled, self.Kmmi) + self.Kmmi) # dC
-        self.dL_dKmm +=  np.dot(np.dot(self.G,self.psi2_beta_scaled) - np.dot(self.LBL_inv, self.psi1VVpsi1), self.Kmmi) + 0.5*self.G # dE
+        self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi)*sf2 # dB
+        self.dL_dKmm += -0.5 * self.D * (- self.C/sf2 - 2.*mdot(self.C, self.psi2_beta_scaled, self.Kmmi) + self.Kmmi) # dC
+        self.dL_dKmm +=  np.dot(np.dot(self.E*sf2, self.psi2_beta_scaled) - np.dot(self.C, self.psi1VVpsi1), self.Kmmi) + 0.5*self.E # dD
 
-    def approximate_likelihood(self):
-        assert not isinstance(self.likelihood, gaussian), "EP is only available for non-gaussian likelihoods"
-        if self.method_ep == 'DTC':
-            self.ep_approx = DTC(self.Kmm,self.likelihood,self.psi1,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
-        elif self.method_ep == 'FITC':
-            self.ep_approx = FITC(self.Kmm,self.likelihood,self.psi1,self.psi0,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
-        else:
-            self.ep_approx = Full(self.X,self.likelihood,self.kernel,inducing=None,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
-        self.beta, self.Y, self.Z_ep = self.ep_approx.fit_EP()
-        self.trbetaYYT = np.sum(np.square(self.Y)*self.beta)
+
+    def _set_params(self, p):
+        self.Z = p[:self.M*self.Q].reshape(self.M, self.Q)
+        self.beta = p[self.M*self.Q] # FIXME
+        self.kern._set_params(p[self.Z.size + 1:])
         self._computations()
 
+    def _get_params(self):
+        return np.hstack([self.Z.flatten(),GP._get_params(self))
+
+    def _get_param_names(self):
+        return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + GP._get_param_names(self)
+
     def log_likelihood(self):
-        """
-        Compute the (lower bound on the) log marginal likelihood
-        """
-        if not self.EP:
-            A = -0.5*self.N*self.D*(np.log(2.*np.pi) - np.log(self.beta))
-            D = -0.5*self.beta*self.trYYT
-        else:
-            A = -0.5*self.D*(self.N*np.log(2.*np.pi) - np.sum(np.log(self.beta)))
-            D = -0.5*self.trbetaYYT
-        B = -0.5*self.D*self.trace_K_beta_scaled
-        C = -0.5*self.D * self.B_logdet
-        E = +0.5*np.sum(self.psi1VVpsi1 * self.LBL_inv)
-        return A+B+C+D+E
+        """ Compute the (lower bound on the) log marginal likelihood """
+        sf2 = self.scale_factor**2
+        A = -0.5*self.N*self.D*(np.log(2.*np.pi) - np.log(self.beta)) -0.5*self.beta*self.trYYT # FIXME
+        B = -0.5*self.D*(self.beta*self.psi0-np.trace(self.A)*sf2)# FIXME
+        C = -0.5*self.D * (self.B_logdet + self.M*np.log(sf2))
+        D = +0.5*np.sum(self.psi1VVpsi1 * self.C)
+        return A+B+C+D
 
+    def _log_likelihood_gradients(self):
+        return np.hstack([self.dL_dZ().flatten(), GP._log_likelihood_gradients(self)])
+
+        # FIXME: move this into the lieklihood class
     def dL_dbeta(self):
-        """
-        Compute the gradient of the log likelihood wrt beta.
-        """
-        #TODO: suport heteroscedatic noise
-        dA_dbeta =   0.5 * self.N*self.D/self.beta
-        dB_dbeta = - 0.5 * self.D * self.trace_K
+        sf2 = self.scale_factor**2
+        dA_dbeta =   0.5 * self.N*self.D/self.beta - 0.5 * self.trYYT
+        dB_dbeta = - 0.5 * self.D * (self.psi0 - np.trace(self.A)/self.beta*sf2)
         dC_dbeta = - 0.5 * self.D * np.sum(self.Bi*self.A)/self.beta
-        dD_dbeta = - 0.5 * self.trYYT
-        tmp = mdot(self.LBi.T, self.LLambdai, self.psi1V)
-        dE_dbeta = (np.sum(np.square(self.C)) - 0.5 * np.sum(self.A * np.dot(tmp, tmp.T)))/self.beta
+        dD_dbeta = np.sum((self.C - 0.5 * mdot(self.C,self.psi2_beta_scaled,self.C) ) * self.psi1VVpsi1 )/self.beta
 
-        return np.squeeze(dA_dbeta + dB_dbeta + dC_dbeta + dD_dbeta + dE_dbeta)
+        return np.squeeze(dA_dbeta + dB_dbeta + dC_dbeta + dD_dbeta)
 
     def dL_dtheta(self):
         """
@@ -195,10 +145,10 @@ class sparse_GP(GP):
         if self.has_uncertain_inputs:
             dL_dtheta += self.kern.dpsi0_dtheta(self.dL_dpsi0, self.Z,self.X,self.X_uncertainty)
             dL_dtheta += self.kern.dpsi1_dtheta(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
-            dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # for multiple_beta, dL_dpsi2 will be a different shape
+            dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.dL_dpsi1.T, self.Z,self.X, self.X_uncertainty)
         else:
             #re-cast computations in psi2 back to psi1:
-            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2_,self.beta.T*self.psi1) #dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1)
+            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2.sum(0),self.psi1)
             dL_dtheta += self.kern.dK_dtheta(dL_dpsi1,self.Z,self.X)
             dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X)
 
@@ -208,48 +158,36 @@ class sparse_GP(GP):
         """
         The derivative of the bound wrt the inducing inputs Z
         """
-        dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm,self.Z,)#factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ
+        dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm,self.Z)#factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ
         if self.has_uncertain_inputs:
-            dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
-            dL_dZ += self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty)
+            dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1,self.Z,self.X, self.X_uncertainty)
+            dL_dZ += 2.*self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # 'stripes'
         else:
             #re-cast computations in psi2 back to psi1:
-            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2_,self.beta.T*self.psi1)#dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1)
+            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2.sum(0),self.psi1)
             dL_dZ += self.kern.dK_dX(dL_dpsi1,self.Z,self.X)
         return dL_dZ
 
-    def _log_likelihood_gradients(self):
-        if not self.EP:
-            return np.hstack([self.dL_dZ().flatten(), self.dL_dbeta(), self.dL_dtheta()])
-        else:
-            return np.hstack([self.dL_dZ().flatten(), self.dL_dtheta()])
-
     def _raw_predict(self, Xnew, slices, full_cov=False):
         """Internal helper function for making predictions, does not account for normalisation"""
+
         Kx = self.kern.K(self.Z, Xnew)
-        mu = mdot(Kx.T, self.LBL_inv, self.psi1V)
-        phi = None
+        mu = mdot(Kx.T, self.C/self.scale_factor, self.psi1V)
+
         if full_cov:
             Kxx = self.kern.K(Xnew)
-            var = Kxx - mdot(Kx.T, (self.Kmmi - self.LBL_inv), Kx)
-            if not self.EP:
-                var += np.eye(Xnew.shape[0])/self.beta
-            else:
-                raise NotImplementedError, "full_cov = True not implemented for EP"
+            var = Kxx - mdot(Kx.T, (self.Kmmi - self.C/self.scale_factor**2), Kx)
         else:
             Kxx = self.kern.Kdiag(Xnew)
-            var = Kxx - np.sum(Kx*np.dot(self.Kmmi - self.LBL_inv, Kx),0)
-            if not self.EP:
-                var += 1./self.beta
-            else:
-                phi = self.likelihood.predictive_mean(mu,var)
-        return mu,var,phi
+            var = Kxx - np.sum(Kx*np.dot(self.Kmmi - self.C/self.scale_factor**2, Kx),0)
+
+        return mu,var
 
     def plot(self, *args, **kwargs):
         """
         Plot the fitted model: just call the GP_regression plot function and then add inducing inputs
         """
-        GP.plot(self,*args,**kwargs)
+        GP_regression.plot(self,*args,**kwargs)
         if self.Q==1:
             pb.plot(self.Z,self.Z*0+pb.ylim()[0],'k|',mew=1.5,markersize=12)
             if self.has_uncertain_inputs:
diff --git a/GPy/models/sparse_GP_old.py b/GPy/models/sparse_GP_old.py
new file mode 100644
index 00000000..7b043209
--- /dev/null
+++ b/GPy/models/sparse_GP_old.py
@@ -0,0 +1,258 @@
+# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
+# Licensed under the BSD 3-clause license (see LICENSE.txt)
+
+import numpy as np
+import pylab as pb
+from ..util.linalg import mdot, jitchol, chol_inv, pdinv
+from ..util.plot import gpplot
+from .. import kern
+from GP import GP
+
+
+#Still TODO:
+# make use of slices properly (kernel can now do this)
+# enable heteroscedatic noise (kernel will need to compute psi2 as a (NxMxM) array)
+
+class sparse_GP(GP):
+    """
+    Variational sparse GP model (Regression)
+
+    :param X: inputs
+    :type X: np.ndarray (N x Q)
+    :param Y: observed data
+    :type Y: np.ndarray of observations (N x D)
+    :param kernel : the kernel/covariance function. See link kernels
+    :type kernel: a GPy kernel
+    :param Z: inducing inputs (optional, see note)
+    :type Z: np.ndarray (M x Q) | None
+    :param X_uncertainty: The uncertainty in the measurements of X (Gaussian variance)
+    :type X_uncertainty: np.ndarray (N x Q) | None
+    :param Zslices: slices for the inducing inputs (see slicing TODO: link)
+    :param M : Number of inducing points (optional, default 10. Ignored if Z is not None)
+    :type M: int
+    :param beta: noise precision. TODO> ignore beta if doing EP
+    :type beta: float
+    :param normalize_(X|Y) : whether to normalize the data before computing (predictions will be in original scales)
+    :type normalize_(X|Y): bool
+    """
+
+    def __init__(self,X,Y=None,kernel=None,X_uncertainty=None,beta=100.,Z=None,Zslices=None,M=10,normalize_X=False,normalize_Y=False,likelihood=None,method_ep='DTC',epsilon_ep=1e-3,power_ep=[1.,1.]):
+
+        if Z is None:
+            self.Z = np.random.permutation(X.copy())[:M]
+            self.M = M
+        else:
+            assert Z.shape[1]==X.shape[1]
+            self.Z = Z
+            self.M = Z.shape[0]
+        if X_uncertainty is None:
+            self.has_uncertain_inputs=False
+        else:
+            assert X_uncertainty.shape==X.shape
+            self.has_uncertain_inputs=True
+            self.X_uncertainty = X_uncertainty
+
+        GP.__init__(self, X=X, Y=Y, kernel=kernel, normalize_X=normalize_X, normalize_Y=normalize_Y,likelihood=likelihood,epsilon_ep=epsilon_ep,power_ep=power_ep)
+
+        #normalise X uncertainty also
+        if self.has_uncertain_inputs:
+            self.X_uncertainty /= np.square(self._Xstd)
+
+        if not self.EP:
+            self.trYYT = np.sum(np.square(self.Y))
+        else:
+            self.method_ep = method_ep
+
+        #normalise X uncertainty also
+        if self.has_uncertain_inputs:
+            self.X_uncertainty /= np.square(self._Xstd)
+
+    def _set_params(self, p):
+        self.Z = p[:self.M*self.Q].reshape(self.M, self.Q)
+        if not self.EP:
+            self.beta = p[self.M*self.Q]
+            self.kern._set_params(p[self.Z.size + 1:])
+        else:
+            self.kern._set_params(p[self.Z.size:])
+            if self.Y is None:
+                self.Y = np.ones([self.N,1])
+        self._compute_kernel_matrices()
+        self._computations()
+
+    def _get_params(self):
+        if not self.EP:
+            return np.hstack([self.Z.flatten(),self.beta,self.kern._get_params_transformed()])
+        else:
+            return np.hstack([self.Z.flatten(),self.kern._get_params_transformed()])
+
+    def _get_param_names(self):
+        if not self.EP:
+            return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + ['noise_precision']+self.kern._get_param_names_transformed()
+        else:
+            return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + self.kern._get_param_names_transformed()
+
+
+    def _compute_kernel_matrices(self):
+        # kernel computations, using BGPLVM notation
+        #TODO: slices for psi statistics (easy enough)
+
+        self.Kmm = self.kern.K(self.Z)
+        if self.has_uncertain_inputs:
+            if not self.EP:
+                self.psi0 = self.kern.psi0(self.Z,self.X, self.X_uncertainty)#.sum() NOTE psi0 is now a vector
+                self.psi1 = self.kern.psi1(self.Z,self.X, self.X_uncertainty).T
+                self.psi2 = self.kern.psi2(self.Z,self.X, self.X_uncertainty)
+                #self.psi2_beta_scaled = ?
+            else:
+                raise NotImplementedError, "uncertain_inputs not yet supported for EP"
+        else:
+            self.psi0 = self.kern.Kdiag(self.X,slices=self.Xslices)#.sum()
+            self.psi1 = self.kern.K(self.Z,self.X)
+            self.psi2 = np.dot(self.psi1,self.psi1.T)
+            self.psi2_beta_scaled = np.dot(self.psi1,self.beta*self.psi1.T)
+
+    def _computations(self):
+        # TODO find routine to multiply triangular matrices
+        self.V = self.beta*self.Y
+        self.psi1V = np.dot(self.psi1, self.V)
+        self.psi1VVpsi1 = np.dot(self.psi1V, self.psi1V.T)
+        self.Kmmi, self.Lm, self.Lmi, self.Kmm_logdet = pdinv(self.Kmm)
+        self.A = mdot(self.Lmi, self.psi2_beta_scaled, self.Lmi.T)
+        self.B = np.eye(self.M) + self.A
+        self.Bi, self.LB, self.LBi, self.B_logdet = pdinv(self.B)
+        self.LLambdai = np.dot(self.LBi, self.Lmi)
+        self.LBL_inv = mdot(self.Lmi.T, self.Bi, self.Lmi)
+        self.C = mdot(self.LLambdai, self.psi1V)
+        self.G =  mdot(self.LBL_inv, self.psi1VVpsi1, self.LBL_inv.T)
+        self.trace_K_beta_scaled = (self.psi0*self.beta).sum() - np.trace(self.A)
+        if not self.EP:
+            self.trace_K = self.psi0.sum() - np.trace(self.A)/self.beta
+
+        # Compute dL_dpsi
+        self.dL_dpsi1 = mdot(self.LLambdai.T,self.C,self.V.T)
+        if not self.EP:
+            self.dL_dpsi0 = - 0.5 * self.D * self.beta * np.ones(self.N)
+            if self.has_uncertain_inputs:
+                self.dL_dpsi2 = - 0.5 * self.beta * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
+            else:
+                self.dL_dpsi2_ = - 0.5 * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
+        else:
+            self.dL_dpsi0 = - 0.5 * self.D * self.beta.flatten()
+            if not self.has_uncertain_inputs:
+                self.dL_dpsi2_ = - 0.5 * (self.D*(self.LBL_inv - self.Kmmi) + self.G)
+
+        # Compute dL_dKmm
+        self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi) # dB
+        self.dL_dKmm += -0.5 * self.D * (- self.LBL_inv - 2.*mdot(self.LBL_inv, self.psi2_beta_scaled, self.Kmmi) + self.Kmmi) # dC
+        self.dL_dKmm +=  np.dot(np.dot(self.G,self.psi2_beta_scaled) - np.dot(self.LBL_inv, self.psi1VVpsi1), self.Kmmi) + 0.5*self.G # dE
+
+    def approximate_likelihood(self):
+        assert not isinstance(self.likelihood, gaussian), "EP is only available for non-gaussian likelihoods"
+        if self.method_ep == 'DTC':
+            self.ep_approx = DTC(self.Kmm,self.likelihood,self.psi1,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
+        elif self.method_ep == 'FITC':
+            self.ep_approx = FITC(self.Kmm,self.likelihood,self.psi1,self.psi0,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
+        else:
+            self.ep_approx = Full(self.X,self.likelihood,self.kernel,inducing=None,epsilon=self.epsilon_ep,power_ep=[self.eta,self.delta])
+        self.beta, self.Y, self.Z_ep = self.ep_approx.fit_EP()
+        self.trbetaYYT = np.sum(np.square(self.Y)*self.beta)
+        self._computations()
+
+    def log_likelihood(self):
+        """
+        Compute the (lower bound on the) log marginal likelihood
+        """
+        if not self.EP:
+            A = -0.5*self.N*self.D*(np.log(2.*np.pi) - np.log(self.beta))
+            D = -0.5*self.beta*self.trYYT
+        else:
+            A = -0.5*self.D*(self.N*np.log(2.*np.pi) - np.sum(np.log(self.beta)))
+            D = -0.5*self.trbetaYYT
+        B = -0.5*self.D*self.trace_K_beta_scaled
+        C = -0.5*self.D * self.B_logdet
+        E = +0.5*np.sum(self.psi1VVpsi1 * self.LBL_inv)
+        return A+B+C+D+E
+
+    def dL_dbeta(self):
+        """
+        Compute the gradient of the log likelihood wrt beta.
+        """
+        #TODO: suport heteroscedatic noise
+        dA_dbeta =   0.5 * self.N*self.D/self.beta
+        dB_dbeta = - 0.5 * self.D * self.trace_K
+        dC_dbeta = - 0.5 * self.D * np.sum(self.Bi*self.A)/self.beta
+        dD_dbeta = - 0.5 * self.trYYT
+        tmp = mdot(self.LBi.T, self.LLambdai, self.psi1V)
+        dE_dbeta = (np.sum(np.square(self.C)) - 0.5 * np.sum(self.A * np.dot(tmp, tmp.T)))/self.beta
+
+        return np.squeeze(dA_dbeta + dB_dbeta + dC_dbeta + dD_dbeta + dE_dbeta)
+
+    def dL_dtheta(self):
+        """
+        Compute and return the derivative of the log marginal likelihood wrt the parameters of the kernel
+        """
+        dL_dtheta = self.kern.dK_dtheta(self.dL_dKmm,self.Z)
+        if self.has_uncertain_inputs:
+            dL_dtheta += self.kern.dpsi0_dtheta(self.dL_dpsi0, self.Z,self.X,self.X_uncertainty)
+            dL_dtheta += self.kern.dpsi1_dtheta(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
+            dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # for multiple_beta, dL_dpsi2 will be a different shape
+        else:
+            #re-cast computations in psi2 back to psi1:
+            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2_,self.beta.T*self.psi1) #dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1)
+            dL_dtheta += self.kern.dK_dtheta(dL_dpsi1,self.Z,self.X)
+            dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X)
+
+        return dL_dtheta
+
+    def dL_dZ(self):
+        """
+        The derivative of the bound wrt the inducing inputs Z
+        """
+        dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm,self.Z,)#factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ
+        if self.has_uncertain_inputs:
+            dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
+            dL_dZ += self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty)
+        else:
+            #re-cast computations in psi2 back to psi1:
+            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2_,self.beta.T*self.psi1)#dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1)
+            dL_dZ += self.kern.dK_dX(dL_dpsi1,self.Z,self.X)
+        return dL_dZ
+
+    def _log_likelihood_gradients(self):
+        if not self.EP:
+            return np.hstack([self.dL_dZ().flatten(), self.dL_dbeta(), self.dL_dtheta()])
+        else:
+            return np.hstack([self.dL_dZ().flatten(), self.dL_dtheta()])
+
+    def _raw_predict(self, Xnew, slices, full_cov=False):
+        """Internal helper function for making predictions, does not account for normalisation"""
+        Kx = self.kern.K(self.Z, Xnew)
+        mu = mdot(Kx.T, self.LBL_inv, self.psi1V)
+        phi = None
+        if full_cov:
+            Kxx = self.kern.K(Xnew)
+            var = Kxx - mdot(Kx.T, (self.Kmmi - self.LBL_inv), Kx)
+            if not self.EP:
+                var += np.eye(Xnew.shape[0])/self.beta
+            else:
+                raise NotImplementedError, "full_cov = True not implemented for EP"
+        else:
+            Kxx = self.kern.Kdiag(Xnew)
+            var = Kxx - np.sum(Kx*np.dot(self.Kmmi - self.LBL_inv, Kx),0)
+            if not self.EP:
+                var += 1./self.beta
+            else:
+                phi = self.likelihood.predictive_mean(mu,var)
+        return mu,var,phi
+
+    def plot(self, *args, **kwargs):
+        """
+        Plot the fitted model: just call the GP_regression plot function and then add inducing inputs
+        """
+        GP.plot(self,*args,**kwargs)
+        if self.Q==1:
+            pb.plot(self.Z,self.Z*0+pb.ylim()[0],'k|',mew=1.5,markersize=12)
+            if self.has_uncertain_inputs:
+                pb.errorbar(self.X[:,0], pb.ylim()[0]+np.zeros(self.N), xerr=2*np.sqrt(self.X_uncertainty.flatten()))
+        if self.Q==2:
+            pb.plot(self.Z[:,0],self.Z[:,1],'wo')
diff --git a/GPy/models/sparse_GP_regression.py b/GPy/models/sparse_GP_regression.py
deleted file mode 100644
index 07ce4d97..00000000
--- a/GPy/models/sparse_GP_regression.py
+++ /dev/null
@@ -1,205 +0,0 @@
-# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
-# Licensed under the BSD 3-clause license (see LICENSE.txt)
-
-import numpy as np
-import pylab as pb
-from ..util.linalg import mdot, jitchol, chol_inv, pdinv
-from ..util.plot import gpplot
-from .. import kern
-from ..inference.likelihoods import likelihood
-from GP_regression import GP_regression
-
-#Still TODO:
-# make use of slices properly (kernel can now do this)
-# enable heteroscedatic noise (kernel will need to compute psi2 as a (NxMxM) array)
-
-class sparse_GP_regression(GP_regression):
-    """
-    Variational sparse GP model (Regression)
-
-    :param X: inputs
-    :type X: np.ndarray (N x Q)
-    :param Y: observed data
-    :type Y: np.ndarray of observations (N x D)
-    :param kernel : the kernel/covariance function. See link kernels
-    :type kernel: a GPy kernel
-    :param Z: inducing inputs (optional, see note)
-    :type Z: np.ndarray (M x Q) | None
-    :param X_uncertainty: The uncertainty in the measurements of X (Gaussian variance)
-    :type X_uncertainty: np.ndarray (N x Q) | None
-    :param Zslices: slices for the inducing inputs (see slicing TODO: link)
-    :param M : Number of inducing points (optional, default 10. Ignored if Z is not None)
-    :type M: int
-    :param beta: noise precision. TODO> ignore beta if doing EP
-    :type beta: float
-    :param normalize_(X|Y) : whether to normalize the data before computing (predictions will be in original scales)
-    :type normalize_(X|Y): bool
-    """
-
-    def __init__(self,X,Y,kernel=None, X_uncertainty=None, beta=100., Z=None,Zslices=None,M=10,normalize_X=False,normalize_Y=False):
-        self.scale_factor = 1000.0
-        self.beta = beta
-        if Z is None:
-            self.Z = np.random.permutation(X.copy())[:M]
-            self.M = M
-        else:
-            assert Z.shape[1]==X.shape[1]
-            self.Z = Z
-            self.M = Z.shape[0]
-        if X_uncertainty is None:
-            self.has_uncertain_inputs=False
-        else:
-            assert X_uncertainty.shape==X.shape
-            self.has_uncertain_inputs=True
-            self.X_uncertainty = X_uncertainty
-
-        GP_regression.__init__(self, X, Y, kernel=kernel, normalize_X=normalize_X, normalize_Y=normalize_Y)
-        self.trYYT = np.sum(np.square(self.Y))
-
-        #normalise X uncertainty also
-        if self.has_uncertain_inputs:
-            self.X_uncertainty /= np.square(self._Xstd)
-
-    def _computations(self):
-        # TODO find routine to multiply triangular matrices
-        #TODO: slices for psi statistics (easy enough)
-
-        # kernel computations, using BGPLVM notation
-        self.Kmm = self.kern.K(self.Z)
-        if self.has_uncertain_inputs:
-            self.psi0 = self.kern.psi0(self.Z,self.X, self.X_uncertainty).sum()
-            self.psi1 = self.kern.psi1(self.Z,self.X, self.X_uncertainty).T
-            self.psi2 = self.kern.psi2(self.Z,self.X, self.X_uncertainty)
-            self.psi2_beta_scaled = (self.psi2*(self.beta/self.scale_factor**2)).sum(0)
-        else:
-            self.psi0 = self.kern.Kdiag(self.X,slices=self.Xslices).sum()
-            self.psi1 = self.kern.K(self.Z,self.X)
-            #self.psi2 = np.dot(self.psi1,self.psi1.T)
-            #self.psi2 = self.psi1.T[:,:,None]*self.psi1.T[:,None,:]
-            tmp = self.psi1/(self.scale_factor/np.sqrt(self.beta))
-            self.psi2_beta_scaled = np.dot(tmp,tmp.T)
-
-        sf = self.scale_factor
-        sf2 = sf**2
-
-        self.Kmmi, self.Lm, self.Lmi, self.Kmm_logdet = pdinv(self.Kmm)#+np.eye(self.M)*1e-3)
-
-        self.V = (self.beta/self.scale_factor)*self.Y
-        self.A = mdot(self.Lmi, self.psi2_beta_scaled, self.Lmi.T)
-        self.B = np.eye(self.M)/sf2 + self.A
-
-        self.Bi, self.LB, self.LBi, self.B_logdet = pdinv(self.B)
-
-        self.psi1V = np.dot(self.psi1, self.V)
-        self.psi1VVpsi1 = np.dot(self.psi1V, self.psi1V.T)
-        self.C = mdot(self.Lmi.T, self.Bi, self.Lmi)
-        self.E = mdot(self.C, self.psi1VVpsi1/sf2, self.C.T)
-
-        # Compute dL_dpsi
-        self.dL_dpsi0 = - 0.5 * self.D * self.beta * np.ones(self.N)
-        self.dL_dpsi1 = mdot(self.V, self.psi1V.T,self.C).T
-        self.dL_dpsi2 = 0.5 * self.beta * self.D * self.Kmmi[None,:,:] # dB
-        self.dL_dpsi2 += - 0.5 * self.beta/sf2 * self.D * self.C[None,:,:] # dC
-        self.dL_dpsi2 += - 0.5 * self.beta * self.E[None,:,:] # dD
-
-        # Compute dL_dKmm
-        self.dL_dKmm = -0.5 * self.D * mdot(self.Lmi.T, self.A, self.Lmi)*sf2 # dB
-        self.dL_dKmm += -0.5 * self.D * (- self.C/sf2 - 2.*mdot(self.C, self.psi2_beta_scaled, self.Kmmi) + self.Kmmi) # dC
-        self.dL_dKmm +=  np.dot(np.dot(self.E*sf2, self.psi2_beta_scaled) - np.dot(self.C, self.psi1VVpsi1), self.Kmmi) + 0.5*self.E # dD
-
-
-    def _set_params(self, p):
-        self.Z = p[:self.M*self.Q].reshape(self.M, self.Q)
-        self.beta = p[self.M*self.Q]
-        self.kern._set_params(p[self.Z.size + 1:])
-        self._computations()
-
-    def _get_params(self):
-        return np.hstack([self.Z.flatten(),self.beta,self.kern._get_params_transformed()])
-
-    def _get_param_names(self):
-        return sum([['iip_%i_%i'%(i,j) for i in range(self.Z.shape[0])] for j in range(self.Z.shape[1])],[]) + ['noise_precision']+self.kern._get_param_names_transformed()
-
-        
-    def log_likelihood(self):
-        """ Compute the (lower bound on the) log marginal likelihood """
-        sf2 = self.scale_factor**2
-        A = -0.5*self.N*self.D*(np.log(2.*np.pi) - np.log(self.beta)) -0.5*self.beta*self.trYYT
-        B = -0.5*self.D*(self.beta*self.psi0-np.trace(self.A)*sf2)
-        C = -0.5*self.D * (self.B_logdet + self.M*np.log(sf2))
-        D = +0.5*np.sum(self.psi1VVpsi1 * self.C)
-        return A+B+C+D
-
-    def _log_likelihood_gradients(self):
-        return np.hstack([self.dL_dZ().flatten(), self.dL_dbeta(), self.dL_dtheta()])
-    
-    def dL_dbeta(self):
-        """
-        Compute the gradient of the log likelihood wrt beta.
-        """
-        #TODO: suport heteroscedatic noise
-        sf2 = self.scale_factor**2
-        dA_dbeta =   0.5 * self.N*self.D/self.beta - 0.5 * self.trYYT
-        dB_dbeta = - 0.5 * self.D * (self.psi0 - np.trace(self.A)/self.beta*sf2)
-        dC_dbeta = - 0.5 * self.D * np.sum(self.Bi*self.A)/self.beta
-        dD_dbeta = np.sum((self.C - 0.5 * mdot(self.C,self.psi2_beta_scaled,self.C) ) * self.psi1VVpsi1 )/self.beta
-
-        return np.squeeze(dA_dbeta + dB_dbeta + dC_dbeta + dD_dbeta)
-
-    def dL_dtheta(self):
-        """
-        Compute and return the derivative of the log marginal likelihood wrt the parameters of the kernel
-        """
-        dL_dtheta = self.kern.dK_dtheta(self.dL_dKmm,self.Z)
-        if self.has_uncertain_inputs:
-            dL_dtheta += self.kern.dpsi0_dtheta(self.dL_dpsi0, self.Z,self.X,self.X_uncertainty)
-            dL_dtheta += self.kern.dpsi1_dtheta(self.dL_dpsi1.T,self.Z,self.X, self.X_uncertainty)
-            dL_dtheta += self.kern.dpsi2_dtheta(self.dL_dpsi2,self.dL_dpsi1.T, self.Z,self.X, self.X_uncertainty) # for multiple_beta, dL_dpsi2 will be a different shape
-        else:
-            #re-cast computations in psi2 back to psi1:
-            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2.sum(0),self.psi1)
-            dL_dtheta += self.kern.dK_dtheta(dL_dpsi1,self.Z,self.X)
-            dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X)
-
-        return dL_dtheta
-
-    def dL_dZ(self):
-        """
-        The derivative of the bound wrt the inducing inputs Z
-        """
-        dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm,self.Z)#factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ
-        if self.has_uncertain_inputs:
-            dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1,self.Z,self.X, self.X_uncertainty)
-            dL_dZ += 2.*self.kern.dpsi2_dZ(self.dL_dpsi2,self.Z,self.X, self.X_uncertainty) # 'stripes'
-        else:
-            #re-cast computations in psi2 back to psi1:
-            dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2.sum(0),self.psi1)
-            dL_dZ += self.kern.dK_dX(dL_dpsi1,self.Z,self.X)
-        return dL_dZ
-
-    def _raw_predict(self, Xnew, slices, full_cov=False):
-        """Internal helper function for making predictions, does not account for normalisation"""
-
-        Kx = self.kern.K(self.Z, Xnew)
-        mu = mdot(Kx.T, self.C/self.scale_factor, self.psi1V)
-
-        if full_cov:
-            Kxx = self.kern.K(Xnew)
-            var = Kxx - mdot(Kx.T, (self.Kmmi - self.C/self.scale_factor**2), Kx) + np.eye(Xnew.shape[0])/self.beta # TODO: This beta doesn't belong here in the EP case.
-        else:
-            Kxx = self.kern.Kdiag(Xnew)
-            var = Kxx - np.sum(Kx*np.dot(self.Kmmi - self.C/self.scale_factor**2, Kx),0) + 1./self.beta # TODO: This beta doesn't belong here in the EP case.
-
-        return mu,var
-
-    def plot(self, *args, **kwargs):
-        """
-        Plot the fitted model: just call the GP_regression plot function and then add inducing inputs
-        """
-        GP_regression.plot(self,*args,**kwargs)
-        if self.Q==1:
-            pb.plot(self.Z,self.Z*0+pb.ylim()[0],'k|',mew=1.5,markersize=12)
-            if self.has_uncertain_inputs:
-                pb.errorbar(self.X[:,0], pb.ylim()[0]+np.zeros(self.N), xerr=2*np.sqrt(self.X_uncertainty.flatten()))
-        if self.Q==2:
-            pb.plot(self.Z[:,0],self.Z[:,1],'wo')