diff --git a/src/gradient_solver/backprop.rs b/src/gradient_solver/backprop.rs
index c88021b..c2c3d40 100644
--- a/src/gradient_solver/backprop.rs
+++ b/src/gradient_solver/backprop.rs
@@ -17,22 +17,22 @@ impl<Loss> NeuraBackprop<Loss> {
 impl<
         Input,
         Target,
-        Trainable: NeuraTrainableLayerBase + NeuraLayer<Input> + NeuraNetworkRec,
+        Trainable: NeuraLayer<Input> + NeuraNetworkRec,
         Loss: NeuraLoss<Trainable::Output, Target = Target> + Clone,
     > NeuraGradientSolver<Input, Target, Trainable> for NeuraBackprop<Loss>
 where
     <Loss as NeuraLoss<Trainable::Output>>::Output: ToPrimitive,
-    // Trainable: NeuraOldTrainableNetworkBase<Input, Gradient = <Trainable as NeuraTrainableLayerBase>::Gradient>,
+    // Trainable: NeuraOldTrainableNetworkBase<Input, Gradient = <Trainable as NeuraLayerBase>::Gradient>,
     // Trainable: for<'a> NeuraOldTrainableNetwork<Input, (&'a NeuraBackprop<Loss>, &'a Target)>,
     for<'a> (&'a NeuraBackprop<Loss>, &'a Target):
-        BackpropRecurse<Input, Trainable, <Trainable as NeuraTrainableLayerBase>::Gradient>,
+        BackpropRecurse<Input, Trainable, <Trainable as NeuraLayerBase>::Gradient>,
 {
     fn get_gradient(
         &self,
         trainable: &Trainable,
         input: &Input,
         target: &Target,
-    ) -> <Trainable as NeuraTrainableLayerBase>::Gradient {
+    ) -> <Trainable as NeuraLayerBase>::Gradient {
         let (_, gradient) = (self, target).recurse(trainable, input);
         // let (_, gradient) = trainable.traverse(input, &(self, target));
 
@@ -59,7 +59,7 @@ impl<Input, Loss: NeuraLoss<Input>> BackpropRecurse<Input, (), ()>
 
 impl<
         Input: Clone,
-        Network: NeuraNetworkRec + NeuraNetwork<Input> + NeuraTrainableLayerEval<Input>,
+        Network: NeuraNetworkRec + NeuraNetwork<Input> + NeuraLayer<Input>,
         Loss,
         Target,
     > BackpropRecurse<Input, Network, Network::Gradient> for (&NeuraBackprop<Loss>, &Target)
@@ -68,14 +68,13 @@ where
     for<'a> (&'a NeuraBackprop<Loss>, &'a Target): BackpropRecurse<
         Network::NodeOutput,
         Network::NextNode,
-        <Network::NextNode as NeuraTrainableLayerBase>::Gradient,
+        <Network::NextNode as NeuraLayerBase>::Gradient,
     >,
     // Verify that the current layer implements the right traits
-    Network::Layer: NeuraTrainableLayerSelf<Network::LayerInput>
-        + NeuraTrainableLayerBackprop<Network::LayerInput>,
+    Network::Layer: NeuraLayer<Network::LayerInput>,
     // Verify that the layer output can be cloned
     <Network::Layer as NeuraLayer<Network::LayerInput>>::Output: Clone,
-    Network::NextNode: NeuraTrainableLayerEval<Network::NodeOutput>,
+    Network::NextNode: NeuraLayer<Network::NodeOutput>,
 {
     fn recurse(&self, network: &Network, input: &Input) -> (Input, Network::Gradient) {
         let layer = network.get_layer();
diff --git a/src/gradient_solver/forward_forward.rs b/src/gradient_solver/forward_forward.rs
index dce173f..a4d5be2 100644
--- a/src/gradient_solver/forward_forward.rs
+++ b/src/gradient_solver/forward_forward.rs
@@ -3,7 +3,7 @@ use num::{traits::NumAssignOps, Float};
 
 use crate::{
     derivable::NeuraDerivable,
-    layer::NeuraTrainableLayerSelf,
+    layer::*,
     network::{NeuraNetwork, NeuraNetworkRec},
     prelude::NeuraLayer,
 };
@@ -69,18 +69,18 @@ impl<
         F: Float,
         Act: Clone + NeuraDerivable<f64>,
         Input: Clone,
-        Trainable: NeuraTrainableLayerBase + NeuraLayer<Input, Output = DVector<F>>,
+        Trainable: NeuraLayer<Input, Output = DVector<F>>,
     > NeuraGradientSolver<Input, bool, Trainable> for NeuraForwardForward<Act>
 where
     NeuraForwardPair<Act>:
-        ForwardForwardRecurse<Input, Trainable, <Trainable as NeuraTrainableLayerBase>::Gradient>,
+        ForwardForwardRecurse<Input, Trainable, <Trainable as NeuraLayerBase>::Gradient>,
 {
     fn get_gradient(
         &self,
         trainable: &Trainable,
         input: &Input,
         target: &bool,
-    ) -> <Trainable as NeuraTrainableLayerBase>::Gradient {
+    ) -> <Trainable as NeuraLayerBase>::Gradient {
         let target = *target;
         let pair = NeuraForwardPair {
             threshold: self.threshold,
@@ -137,13 +137,13 @@ impl<Act, Input> ForwardForwardRecurse<Input, (), ()> for NeuraForwardPair<Act>
 impl<Act, Input: Clone, Network: NeuraNetwork<Input> + NeuraNetworkRec>
     ForwardForwardRecurse<Input, Network, Network::Gradient> for NeuraForwardPair<Act>
 where
-    Network::Layer: NeuraTrainableLayerSelf<Network::LayerInput>,
+    Network::Layer: NeuraLayer<Network::LayerInput>,
     <Network::Layer as NeuraLayer<Network::LayerInput>>::Output: Clone,
     Self: ForwardForwardDerivate<<Network::Layer as NeuraLayer<Network::LayerInput>>::Output>,
     Self: ForwardForwardRecurse<
         Network::NodeOutput,
         Network::NextNode,
-        <Network::NextNode as NeuraTrainableLayerBase>::Gradient,
+        <Network::NextNode as NeuraLayerBase>::Gradient,
     >,
 {
     fn recurse(&self, network: &Network, input: &Input) -> Network::Gradient {
diff --git a/src/gradient_solver/mod.rs b/src/gradient_solver/mod.rs
index 740e0a2..68a7483 100644
--- a/src/gradient_solver/mod.rs
+++ b/src/gradient_solver/mod.rs
@@ -4,9 +4,9 @@ pub use backprop::NeuraBackprop;
 mod forward_forward;
 pub use forward_forward::NeuraForwardForward;
 
-use crate::layer::{NeuraTrainableLayerBase, NeuraTrainableLayerEval};
+use crate::layer::NeuraLayerBase;
 
-pub trait NeuraGradientSolver<Input, Target, Trainable: NeuraTrainableLayerBase> {
+pub trait NeuraGradientSolver<Input, Target, Trainable: NeuraLayerBase> {
     fn get_gradient(
         &self,
         trainable: &Trainable,
diff --git a/src/layer/dense.rs b/src/layer/dense.rs
index 8710ef9..9090260 100644
--- a/src/layer/dense.rs
+++ b/src/layer/dense.rs
@@ -1,7 +1,7 @@
 use std::marker::PhantomData;
 
-use nalgebra::{DMatrix, DVector};
-use num::Float;
+use nalgebra::{DMatrix, DVector, Scalar};
+use num::{traits::NumAssignOps, Float};
 use rand::Rng;
 
 use crate::{derivable::NeuraDerivable, err::NeuraDimensionsMismatch};
@@ -126,18 +126,10 @@ impl<F, Act, Reg, R: Rng> NeuraDenseLayerPartial<F, Act, Reg, R> {
     }
 }
 
-impl<F: Float, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>> NeuraShapedLayer
-    for NeuraDenseLayer<F, Act, Reg>
-{
-    fn output_shape(&self) -> NeuraShape {
-        NeuraShape::Vector(self.weights.shape().0)
-    }
-}
-
 impl<
-        F: Float + std::fmt::Debug + 'static,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
+        F: Float + Scalar + Send + NumAssignOps,
+        Act: NeuraDerivable<F> + Clone + std::fmt::Debug + 'static,
+        Reg: NeuraDerivable<F> + Clone + std::fmt::Debug + 'static,
         R: Rng,
     > NeuraPartialLayer for NeuraDenseLayerPartial<F, Act, Reg, R>
 where
@@ -158,8 +150,10 @@ where
     }
 }
 
-impl<F: Float, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>> NeuraPartialLayer
-    for NeuraDenseLayer<F, Act, Reg>
+impl<F: Float + Scalar + Send + NumAssignOps, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>>
+    NeuraPartialLayer for NeuraDenseLayer<F, Act, Reg>
+where
+    Self: Clone + std::fmt::Debug + 'static,
 {
     type Constructed = Self;
     type Err = NeuraDimensionsMismatch;
@@ -176,28 +170,10 @@ impl<F: Float, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>> NeuraPartialLayer
     }
 }
 
-impl<
-        F: Float + std::fmt::Debug + 'static + std::ops::AddAssign + std::ops::MulAssign,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
-    > NeuraLayer<DVector<F>> for NeuraDenseLayer<F, Act, Reg>
-{
-    type Output = DVector<F>;
-
-    fn eval(&self, input: &DVector<F>) -> Self::Output {
-        assert_eq!(input.shape().0, self.weights.shape().1);
-
-        let evaluated = &self.weights * input + &self.bias;
-
-        evaluated.map(|x| self.activation.eval(x))
-    }
-}
-
-impl<
-        F: Float + std::fmt::Debug + 'static + std::ops::AddAssign + std::ops::MulAssign,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
-    > NeuraTrainableLayerBase for NeuraDenseLayer<F, Act, Reg>
+impl<F: Float + NumAssignOps + Scalar + Send, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>>
+    NeuraLayerBase for NeuraDenseLayer<F, Act, Reg>
+where
+    Self: Clone + std::fmt::Debug + 'static,
 {
     type Gradient = (DMatrix<F>, DVector<F>);
 
@@ -212,14 +188,25 @@ impl<
         self.weights += &gradient.0;
         self.bias += &gradient.1;
     }
+
+    fn output_shape(&self) -> NeuraShape {
+        NeuraShape::Vector(self.weights.shape().0)
+    }
+
+    fn regularize_layer(&self) -> Self::Gradient {
+        (
+            self.weights.map(|x| self.regularization.derivate(x)),
+            DVector::zeros(self.bias.shape().0),
+        )
+    }
 }
 
-impl<
-        F: Float + std::fmt::Debug + 'static + std::ops::AddAssign + std::ops::MulAssign,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
-    > NeuraTrainableLayerEval<DVector<F>> for NeuraDenseLayer<F, Act, Reg>
+impl<F: Float + NumAssignOps + Scalar + Send, Act: NeuraDerivable<F>, Reg: NeuraDerivable<F>>
+    NeuraLayer<DVector<F>> for NeuraDenseLayer<F, Act, Reg>
+where
+    Self: Clone + std::fmt::Debug + 'static,
 {
+    type Output = DVector<F>;
     type IntermediaryRepr = DVector<F>; // pre-activation values
 
     fn eval_training(&self, input: &DVector<F>) -> (Self::Output, Self::IntermediaryRepr) {
@@ -228,20 +215,6 @@ impl<
 
         (output, evaluated)
     }
-}
-
-impl<
-        F: Float + std::fmt::Debug + 'static + std::ops::AddAssign + std::ops::MulAssign,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
-    > NeuraTrainableLayerSelf<DVector<F>> for NeuraDenseLayer<F, Act, Reg>
-{
-    fn regularize_layer(&self) -> Self::Gradient {
-        (
-            self.weights.map(|x| self.regularization.derivate(x)),
-            DVector::zeros(self.bias.shape().0),
-        )
-    }
 
     fn get_gradient(
         &self,
@@ -266,14 +239,7 @@ impl<
 
         (weights_gradient, bias_gradient)
     }
-}
 
-impl<
-        F: Float + std::fmt::Debug + 'static + std::ops::AddAssign + std::ops::MulAssign,
-        Act: NeuraDerivable<F>,
-        Reg: NeuraDerivable<F>,
-    > NeuraTrainableLayerBackprop<DVector<F>> for NeuraDenseLayer<F, Act, Reg>
-{
     fn backprop_layer(
         &self,
         _input: &DVector<F>,
diff --git a/src/layer/dropout.rs b/src/layer/dropout.rs
index be56222..8989cba 100644
--- a/src/layer/dropout.rs
+++ b/src/layer/dropout.rs
@@ -35,13 +35,7 @@ impl<R: Rng> NeuraDropoutLayer<R> {
     }
 }
 
-impl<R: Rng> NeuraShapedLayer for NeuraDropoutLayer<R> {
-    fn output_shape(&self) -> NeuraShape {
-        self.shape
-    }
-}
-
-impl<R: Rng> NeuraPartialLayer for NeuraDropoutLayer<R> {
+impl<R: Rng + Clone + std::fmt::Debug + 'static> NeuraPartialLayer for NeuraDropoutLayer<R> {
     type Constructed = NeuraDropoutLayer<R>;
 
     type Err = ();
@@ -53,25 +47,15 @@ impl<R: Rng> NeuraPartialLayer for NeuraDropoutLayer<R> {
     }
 }
 
-impl<R: Rng, F: Float> NeuraLayer<DVector<F>> for NeuraDropoutLayer<R> {
-    type Output = DVector<F>;
-
-    fn eval(&self, input: &DVector<F>) -> Self::Output {
-        let mut output = input.clone();
-        self.apply_dropout(&mut output);
-        output
-    }
-}
-
-impl<R: Rng> NeuraTrainableLayerBase for NeuraDropoutLayer<R> {
+impl<R: Rng + Clone + std::fmt::Debug + 'static> NeuraLayerBase for NeuraDropoutLayer<R> {
     type Gradient = ();
 
     fn default_gradient(&self) -> Self::Gradient {
         ()
     }
 
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
+    fn output_shape(&self) -> NeuraShape {
+        self.shape
     }
 
     fn prepare_layer(&mut self, is_training: bool) {
@@ -98,30 +82,19 @@ impl<R: Rng> NeuraTrainableLayerBase for NeuraDropoutLayer<R> {
     }
 }
 
-impl<R: Rng, F: Float> NeuraTrainableLayerEval<DVector<F>> for NeuraDropoutLayer<R> {
+impl<R: Rng + Clone + std::fmt::Debug + 'static, F: Float> NeuraLayer<DVector<F>>
+    for NeuraDropoutLayer<R>
+{
+    type Output = DVector<F>;
+
     type IntermediaryRepr = ();
 
     fn eval_training(&self, input: &DVector<F>) -> (Self::Output, Self::IntermediaryRepr) {
-        (self.eval(input), ())
-    }
-}
-
-impl<R: Rng, F: Float> NeuraTrainableLayerSelf<DVector<F>> for NeuraDropoutLayer<R> {
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
-
-    fn get_gradient(
-        &self,
-        _input: &DVector<F>,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
+        let mut output = input.clone();
+        self.apply_dropout(&mut output);
+        (output, ())
     }
-}
 
-impl<R: Rng, F: Float> NeuraTrainableLayerBackprop<DVector<F>> for NeuraDropoutLayer<R> {
     fn backprop_layer(
         &self,
         _input: &DVector<F>,
diff --git a/src/layer/isolate.rs b/src/layer/isolate.rs
index f45c4cf..ea553b4 100644
--- a/src/layer/isolate.rs
+++ b/src/layer/isolate.rs
@@ -24,12 +24,6 @@ impl NeuraIsolateLayer {
     }
 }
 
-impl NeuraShapedLayer for NeuraIsolateLayer {
-    fn output_shape(&self) -> NeuraShape {
-        self.end.sub(self.start).unwrap_or_else(|| unreachable!())
-    }
-}
-
 impl NeuraPartialLayer for NeuraIsolateLayer {
     type Constructed = NeuraIsolateLayer;
     type Err = NeuraIsolateLayerErr;
@@ -70,19 +64,7 @@ impl NeuraPartialLayer for NeuraIsolateLayer {
     }
 }
 
-impl<F: Clone + Scalar> NeuraLayer<DVector<F>> for NeuraIsolateLayer {
-    type Output = DVector<F>;
-
-    fn eval(&self, input: &DVector<F>) -> Self::Output {
-        let (NeuraShape::Vector(start), NeuraShape::Vector(end)) = (self.start, self.end) else {
-            panic!("NeuraIsolateLayer expected a value of dimension {}, got a vector", self.start.dims());
-        };
-
-        DVector::from_iterator(end - start, input.iter().cloned().skip(start).take(end))
-    }
-}
-
-impl NeuraTrainableLayerBase for NeuraIsolateLayer {
+impl NeuraLayerBase for NeuraIsolateLayer {
     type Gradient = ();
 
     #[inline(always)]
@@ -90,41 +72,26 @@ impl NeuraTrainableLayerBase for NeuraIsolateLayer {
         ()
     }
 
-    #[inline(always)]
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
+    fn output_shape(&self) -> NeuraShape {
+        self.end.sub(self.start).unwrap_or_else(|| unreachable!())
     }
 }
 
-impl<F: Clone + Scalar> NeuraTrainableLayerEval<DVector<F>> for NeuraIsolateLayer {
+impl<F: Clone + Scalar + Default> NeuraLayer<DVector<F>> for NeuraIsolateLayer {
+    type Output = DVector<F>;
+
     type IntermediaryRepr = ();
 
     fn eval_training(&self, input: &DVector<F>) -> (Self::Output, Self::IntermediaryRepr) {
-        (self.eval(input), ())
-    }
-}
+        let (NeuraShape::Vector(start), NeuraShape::Vector(end)) = (self.start, self.end) else {
+            panic!("NeuraIsolateLayer expected a value of dimension {}, got a vector", self.start.dims());
+        };
 
-impl<Input> NeuraTrainableLayerSelf<Input> for NeuraIsolateLayer
-where
-    Self: NeuraTrainableLayerEval<Input>,
-{
-    #[inline(always)]
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
+        let res = DVector::from_iterator(end - start, input.iter().cloned().skip(start).take(end));
 
-    #[inline(always)]
-    fn get_gradient(
-        &self,
-        _input: &Input,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
+        (res, ())
     }
-}
 
-impl<F: Clone + Scalar + Default> NeuraTrainableLayerBackprop<DVector<F>> for NeuraIsolateLayer {
     fn backprop_layer(
         &self,
         input: &DVector<F>,
diff --git a/src/layer/lock.rs b/src/layer/lock.rs
index d8d8dc0..64bebe2 100644
--- a/src/layer/lock.rs
+++ b/src/layer/lock.rs
@@ -24,56 +24,35 @@ impl<Layer> NeuraLockLayer<Layer> {
     }
 }
 
-impl<Input, Layer: NeuraLayer<Input>> NeuraLayer<Input> for NeuraLockLayer<Layer> {
-    type Output = Layer::Output;
+impl<Layer: NeuraLayerBase> NeuraLayerBase for NeuraLockLayer<Layer> {
+    type Gradient = ();
 
-    fn eval(&self, input: &Input) -> Self::Output {
-        self.layer.eval(input)
+    fn output_shape(&self) -> NeuraShape {
+        self.layer.output_shape()
     }
-}
-
-impl<Layer: NeuraTrainableLayerBase> NeuraTrainableLayerBase for NeuraLockLayer<Layer> {
-    type Gradient = ();
 
     fn default_gradient(&self) -> Self::Gradient {
         ()
     }
 
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
+    fn prepare_layer(&mut self, is_training: bool) {
+        self.layer.prepare_layer(is_training);
     }
 }
 
-impl<Input, Layer: NeuraTrainableLayerEval<Input>> NeuraTrainableLayerEval<Input>
-    for NeuraLockLayer<Layer>
-{
-    type IntermediaryRepr = Layer::IntermediaryRepr;
+impl<Input, Layer: NeuraLayer<Input>> NeuraLayer<Input> for NeuraLockLayer<Layer> {
+    type Output = Layer::Output;
 
-    fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr) {
-        self.layer.eval_training(input)
+    fn eval(&self, input: &Input) -> Self::Output {
+        self.layer.eval(input)
     }
-}
 
-impl<Input, Layer: NeuraTrainableLayerEval<Input>> NeuraTrainableLayerSelf<Input>
-    for NeuraLockLayer<Layer>
-{
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
+    type IntermediaryRepr = Layer::IntermediaryRepr;
 
-    fn get_gradient(
-        &self,
-        _input: &Input,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
+    fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr) {
+        self.layer.eval_training(input)
     }
-}
 
-impl<Input, Layer: NeuraTrainableLayerBackprop<Input>> NeuraTrainableLayerBackprop<Input>
-    for NeuraLockLayer<Layer>
-{
     fn backprop_layer(
         &self,
         input: &Input,
diff --git a/src/layer/mod.rs b/src/layer/mod.rs
index 9dd6547..3f46cda 100644
--- a/src/layer/mod.rs
+++ b/src/layer/mod.rs
@@ -75,11 +75,35 @@ impl From<(usize, usize, usize)> for NeuraShape {
     }
 }
 
-pub trait NeuraLayer<Input> {
-    /// What type the layer outputs
-    type Output;
+pub trait NeuraLayerBase: std::fmt::Debug + Clone + 'static {
+    /// What type the gradient of the layer is
+    type Gradient: NeuraVectorSpace + Send + 'static;
+
+    /// What the desired output shape of the layer is
+    fn output_shape(&self) -> NeuraShape;
+
+    /// Constructor for the gradient, should return the zero vector
+    fn default_gradient(&self) -> Self::Gradient;
+
+    /// Applies `δW_l` to the weights of the layer, the default implementation is a noop
+    #[allow(unused_variables)]
+    #[inline(always)]
+    fn apply_gradient(&mut self, gradient: &Self::Gradient) {
+        // Noop
+    }
+
+    /// Arbitrary computation that can be executed at the start of an epoch
+    #[allow(unused_variables)]
+    #[inline(always)]
+    fn prepare_layer(&mut self, is_training: bool) {
+        // Noop
+    }
 
-    fn eval(&self, input: &Input) -> Self::Output;
+    /// Computes the regularization terms of the layer's gradient, called once per batch
+    #[inline(always)]
+    fn regularize_layer(&self) -> Self::Gradient {
+        self.default_gradient()
+    }
 
     fn lock_layer(self) -> NeuraLockLayer<Self>
     where
@@ -89,87 +113,34 @@ pub trait NeuraLayer<Input> {
     }
 }
 
-impl<Input: Clone> NeuraLayer<Input> for () {
-    type Output = Input;
+pub trait NeuraLayer<Input>: NeuraLayerBase {
+    /// What type the layer outputs, may depend on `Input`.
+    type Output;
+
+    /// A type that can hold data between calls to `eval_training`, `backprop_layer` and `get_gradient`
+    type IntermediaryRepr: 'static;
 
-    #[inline(always)]
     fn eval(&self, input: &Input) -> Self::Output {
-        input.clone()
+        self.eval_training(input).0
     }
-}
-
-pub trait NeuraShapedLayer {
-    fn output_shape(&self) -> NeuraShape;
-}
-
-pub trait NeuraPartialLayer {
-    type Constructed: NeuraShapedLayer;
-    type Err;
-
-    fn construct(self, input_shape: NeuraShape) -> Result<Self::Constructed, Self::Err>;
-}
-
-pub trait NeuraTrainableLayerBase {
-    /// The representation of the layer gradient as a vector space
-    type Gradient: NeuraVectorSpace;
-
-    fn default_gradient(&self) -> Self::Gradient;
-
-    /// Applies `δW_l` to the weights of the layer
-    fn apply_gradient(&mut self, gradient: &Self::Gradient);
-
-    /// Arbitrary computation that can be executed at the start of an epoch
-    #[allow(unused_variables)]
-    #[inline(always)]
-    fn prepare_layer(&mut self, is_training: bool) {}
-}
 
-pub trait NeuraTrainableLayerEval<Input>: NeuraTrainableLayerBase + NeuraLayer<Input> {
-    /// An intermediary object type to be passed to the various training methods
-    type IntermediaryRepr;
-
-    // TODO: move this into another trait
     fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr);
-}
-
-/// Contains methods relative to a layer's ability to compute its own weights gradients,
-/// given the derivative of the output variables.
-pub trait NeuraTrainableLayerSelf<Input>: NeuraTrainableLayerEval<Input> {
-    /// Computes the regularization
-    fn regularize_layer(&self) -> Self::Gradient;
 
     /// Computes the layer's gradient,
     ///
     /// `intermediary` is guaranteed to have been generated by a previous call to `eval_training`,
     /// without mutation of `self` in-between, and with the same `input`.
+    #[allow(unused_variables)]
+    #[inline(always)]
     fn get_gradient(
         &self,
         input: &Input,
         intermediary: &Self::IntermediaryRepr,
         epsilon: &Self::Output,
-    ) -> Self::Gradient;
-}
+    ) -> Self::Gradient {
+        self.default_gradient()
+    }
 
-// impl<Input, Layer: NeuraTrainableLayerBase<Input, Gradient = ()>> NeuraTrainableLayerSelf<Input>
-//     for Layer
-// {
-//     #[inline(always)]
-//     fn regularize_layer(&self) -> Self::Gradient {
-//         ()
-//     }
-
-//     #[inline(always)]
-//     fn get_gradient(
-//         &self,
-//         input: &Input,
-//         intermediary: &Self::IntermediaryRepr,
-//         epsilon: Self::Output,
-//     ) -> Self::Gradient {
-//         ()
-//     }
-// }
-
-pub trait NeuraTrainableLayerBackprop<Input>: NeuraTrainableLayerEval<Input> {
     /// Computes the backpropagation term and the derivative of the internal weights,
     /// using the `input` vector outputted by the previous layer and the backpropagation term `epsilon` of the next layer.
     ///
@@ -189,7 +160,31 @@ pub trait NeuraTrainableLayerBackprop<Input>: NeuraTrainableLayerEval<Input> {
     ) -> Input;
 }
 
-impl NeuraTrainableLayerBase for () {
+#[deprecated]
+pub trait NeuraShapedLayer: NeuraLayerBase {}
+
+pub trait NeuraPartialLayer {
+    type Constructed: NeuraLayerBase + 'static;
+    type Err;
+
+    fn construct(self, input_shape: NeuraShape) -> Result<Self::Constructed, Self::Err>;
+}
+
+#[deprecated]
+pub trait NeuraTrainableLayerBase: NeuraLayerBase {}
+
+#[deprecated]
+pub trait NeuraTrainableLayerEval<Input>: NeuraLayer<Input> {}
+
+/// Contains methods relative to a layer's ability to compute its own weights gradients,
+/// given the derivative of the output variables.
+#[deprecated]
+pub trait NeuraTrainableLayerSelf<Input>: NeuraLayer<Input> {}
+
+#[deprecated]
+pub trait NeuraTrainableLayerBackprop<Input>: NeuraLayer<Input> {}
+
+impl NeuraLayerBase for () {
     type Gradient = ();
 
     #[inline(always)]
@@ -201,35 +196,22 @@ impl NeuraTrainableLayerBase for () {
     fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
         // Noop
     }
+
+    fn output_shape(&self) -> NeuraShape {
+        panic!("() has no shape!");
+    }
 }
 
-impl<Input: Clone> NeuraTrainableLayerEval<Input> for () {
+impl<Input: Clone> NeuraLayer<Input> for () {
+    type Output = Input;
     type IntermediaryRepr = ();
 
     #[inline(always)]
     fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr) {
-        (self.eval(input), ())
+        (input.clone(), ())
     }
-}
 
-impl<Input: Clone> NeuraTrainableLayerSelf<Input> for () {
     #[inline(always)]
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
-
-    #[inline(always)]
-    fn get_gradient(
-        &self,
-        _input: &Input,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
-    }
-}
-
-impl<Input: Clone> NeuraTrainableLayerBackprop<Input> for () {
     fn backprop_layer(
         &self,
         _input: &Input,
diff --git a/src/layer/normalize.rs b/src/layer/normalize.rs
index c12d95b..5e2d6af 100644
--- a/src/layer/normalize.rs
+++ b/src/layer/normalize.rs
@@ -23,12 +23,6 @@ impl NeuraNormalizeLayer {
     }
 }
 
-impl NeuraShapedLayer for NeuraNormalizeLayer {
-    fn output_shape(&self) -> NeuraShape {
-        self.shape
-    }
-}
-
 impl NeuraPartialLayer for NeuraNormalizeLayer {
     type Constructed = NeuraNormalizeLayer;
 
@@ -39,7 +33,22 @@ impl NeuraPartialLayer for NeuraNormalizeLayer {
     }
 }
 
-impl<F: Float + Scalar> NeuraLayer<DVector<F>> for NeuraNormalizeLayer {
+impl NeuraLayerBase for NeuraNormalizeLayer {
+    fn output_shape(&self) -> NeuraShape {
+        self.shape
+    }
+
+    type Gradient = ();
+
+    fn default_gradient(&self) -> Self::Gradient {
+        ()
+    }
+}
+
+impl<F: Float + Scalar + NumAssignOps> NeuraLayer<DVector<F>> for NeuraNormalizeLayer {
+    // TODO: store the kroenecker term in the jacobian matrix (might as well)
+    type IntermediaryRepr = (DMatrix<F>, F); // Partial jacobian matrix (without the kroenecker term) and stddev
+
     type Output = DVector<F>;
 
     fn eval(&self, input: &DVector<F>) -> Self::Output {
@@ -54,22 +63,6 @@ impl<F: Float + Scalar> NeuraLayer<DVector<F>> for NeuraNormalizeLayer {
 
         output
     }
-}
-
-impl NeuraTrainableLayerBase for NeuraNormalizeLayer {
-    type Gradient = ();
-
-    fn default_gradient(&self) -> Self::Gradient {
-        ()
-    }
-
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
-    }
-}
-
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerEval<DVector<F>> for NeuraNormalizeLayer {
-    type IntermediaryRepr = (DMatrix<F>, F); // Partial jacobian matrix (without the kroenecker term) and stddev
 
     fn eval_training(&self, input: &DVector<F>) -> (Self::Output, Self::IntermediaryRepr) {
         let (mean, variance, len) = mean_variance(input);
@@ -85,26 +78,7 @@ impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerEval<DVector<F>> for N
 
         (input_centered / stddev, (jacobian_partial, stddev))
     }
-}
-
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerSelf<DVector<F>> for NeuraNormalizeLayer {
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
-
-    fn get_gradient(
-        &self,
-        _input: &DVector<F>,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
-    }
-}
 
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerBackprop<DVector<F>>
-    for NeuraNormalizeLayer
-{
     fn backprop_layer(
         &self,
         _input: &DVector<F>,
diff --git a/src/layer/softmax.rs b/src/layer/softmax.rs
index 8172a88..ebae80f 100644
--- a/src/layer/softmax.rs
+++ b/src/layer/softmax.rs
@@ -16,8 +16,30 @@ impl NeuraSoftmaxLayer {
     }
 }
 
+impl NeuraPartialLayer for NeuraSoftmaxLayer {
+    type Constructed = Self;
+    type Err = ();
+
+    fn construct(self, input_shape: NeuraShape) -> Result<Self::Constructed, Self::Err> {
+        Ok(Self { shape: input_shape })
+    }
+}
+
+impl NeuraLayerBase for NeuraSoftmaxLayer {
+    type Gradient = ();
+
+    fn output_shape(&self) -> NeuraShape {
+        self.shape
+    }
+
+    fn default_gradient(&self) -> Self::Gradient {
+        ()
+    }
+}
+
 impl<F: Float + Scalar + NumAssignOps> NeuraLayer<DVector<F>> for NeuraSoftmaxLayer {
     type Output = DVector<F>;
+    type IntermediaryRepr = Self::Output; // Result of self.eval
 
     fn eval(&self, input: &DVector<F>) -> Self::Output {
         let mut res = input.clone();
@@ -39,64 +61,12 @@ impl<F: Float + Scalar + NumAssignOps> NeuraLayer<DVector<F>> for NeuraSoftmaxLa
 
         res
     }
-}
-
-impl NeuraShapedLayer for NeuraSoftmaxLayer {
-    fn output_shape(&self) -> NeuraShape {
-        self.shape
-    }
-}
-
-impl NeuraPartialLayer for NeuraSoftmaxLayer {
-    type Constructed = Self;
-    type Err = ();
-
-    fn construct(self, input_shape: NeuraShape) -> Result<Self::Constructed, Self::Err> {
-        Ok(Self { shape: input_shape })
-    }
-}
-
-impl NeuraTrainableLayerBase for NeuraSoftmaxLayer {
-    type Gradient = ();
-
-    fn default_gradient(&self) -> Self::Gradient {
-        ()
-    }
-
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
-    }
-}
-
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerEval<DVector<F>> for NeuraSoftmaxLayer {
-    type IntermediaryRepr = Self::Output; // Result of self.eval
 
     fn eval_training(&self, input: &DVector<F>) -> (Self::Output, Self::IntermediaryRepr) {
         let res = self.eval(input);
         (res.clone(), res)
     }
-}
-
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerSelf<DVector<F>> for NeuraSoftmaxLayer {
-    #[inline(always)]
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
-
-    #[inline(always)]
-    fn get_gradient(
-        &self,
-        _input: &DVector<F>,
-        _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
-    }
-}
 
-impl<F: Float + Scalar + NumAssignOps> NeuraTrainableLayerBackprop<DVector<F>>
-    for NeuraSoftmaxLayer
-{
     fn backprop_layer(
         &self,
         input: &DVector<F>,
diff --git a/src/lib.rs b/src/lib.rs
index 270bd68..8b5885d 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -23,10 +23,7 @@ pub mod prelude {
 
     // Structs and traits
     pub use crate::gradient_solver::NeuraBackprop;
-    pub use crate::layer::{
-        NeuraLayer, NeuraPartialLayer, NeuraShape, NeuraTrainableLayerBackprop,
-        NeuraTrainableLayerBase, NeuraTrainableLayerSelf,
-    };
+    pub use crate::layer::{NeuraLayer, NeuraLayerBase, NeuraPartialLayer, NeuraShape};
     pub use crate::network::sequential::{
         NeuraSequential, NeuraSequentialLock, NeuraSequentialTail,
     };
diff --git a/src/network/graph/from.rs b/src/network/graph/from.rs
index 3336bae..47f1364 100644
--- a/src/network/graph/from.rs
+++ b/src/network/graph/from.rs
@@ -25,7 +25,7 @@ impl<Data> FromSequential<(), Data> for NeuraGraph<Data> {
     }
 }
 
-impl<Data: Clone + 'static, Layer: NeuraTrainableLayerFull<Data, Output = Data>, ChildNetwork>
+impl<Data: Clone + 'static, Layer: NeuraLayer<Data, Output = Data>, ChildNetwork>
     FromSequential<NeuraSequential<Layer, ChildNetwork>, Data> for NeuraGraph<Data>
 where
     NeuraGraph<Data>: FromSequential<ChildNetwork, Data>,
@@ -57,7 +57,7 @@ impl<Data> NeuraGraph<Data> {
     ) -> Self
     where
         NeuraGraph<Data>: FromSequential<NeuraSequential<Layer, ChildNetwork>, Data>,
-        NeuraSequential<Layer, ChildNetwork>: NeuraShapedLayer,
+        NeuraSequential<Layer, ChildNetwork>: NeuraLayerBase,
     {
         Self::from_sequential_rec(&network, vec![], input_shape)
     }
diff --git a/src/network/graph/mod.rs b/src/network/graph/mod.rs
index 1156e2e..3713fd9 100644
--- a/src/network/graph/mod.rs
+++ b/src/network/graph/mod.rs
@@ -1,10 +1,7 @@
 use std::any::Any;
 
 use crate::{
-    algebra::NeuraDynVectorSpace,
-    derivable::NeuraLoss,
-    layer::{NeuraShapedLayer, NeuraTrainableLayerEval},
-    prelude::*,
+    algebra::NeuraDynVectorSpace, derivable::NeuraLoss, layer::NeuraLayerBase, prelude::*,
 };
 
 mod node;
@@ -16,31 +13,10 @@ pub use partial::NeuraGraphPartial;
 mod from;
 pub use from::FromSequential;
 
-pub trait NeuraTrainableLayerFull<Input>:
-    NeuraTrainableLayerEval<Input>
-    + NeuraTrainableLayerBackprop<Input>
-    + NeuraTrainableLayerSelf<Input>
-    + NeuraShapedLayer
-    + Clone
-    + std::fmt::Debug
-    + 'static
-where
-    Self::IntermediaryRepr: 'static,
-{
-}
+#[deprecated]
+pub trait NeuraTrainableLayerFull<Input>: NeuraLayer<Input> {}
 
-impl<Input, T> NeuraTrainableLayerFull<Input> for T
-where
-    T: NeuraTrainableLayerEval<Input>
-        + NeuraTrainableLayerBackprop<Input>
-        + NeuraTrainableLayerSelf<Input>
-        + NeuraShapedLayer
-        + Clone
-        + std::fmt::Debug
-        + 'static,
-    T::IntermediaryRepr: 'static,
-{
-}
+impl<Input, T> NeuraTrainableLayerFull<Input> for T where T: NeuraLayer<Input> {}
 
 #[derive(Debug)]
 pub struct NeuraGraphNodeConstructed<Data> {
@@ -49,7 +25,17 @@ pub struct NeuraGraphNodeConstructed<Data> {
     output: usize,
 }
 
-#[derive(Debug)]
+impl<Data> Clone for NeuraGraphNodeConstructed<Data> {
+    fn clone(&self) -> Self {
+        Self {
+            node: dyn_clone::clone_box(&*self.node),
+            inputs: self.inputs.clone(),
+            output: self.output.clone(),
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
 pub struct NeuraGraph<Data> {
     /// ## Class invariants
     ///
@@ -65,10 +51,16 @@ pub struct NeuraGraph<Data> {
     buffer_size: usize,
 }
 
-impl<Data> NeuraShapedLayer for NeuraGraph<Data> {
+impl<Data: Clone + std::fmt::Debug + 'static> NeuraLayerBase for NeuraGraph<Data> {
+    type Gradient = ();
+
     fn output_shape(&self) -> NeuraShape {
         self.output_shape
     }
+
+    fn default_gradient(&self) -> Self::Gradient {
+        unimplemented!("NeuraGraph cannot be used as a layer yet")
+    }
 }
 
 impl<Data> NeuraGraph<Data> {
@@ -178,8 +170,9 @@ impl<Data> NeuraGraph<Data> {
     }
 }
 
-impl<Data: Clone> NeuraLayer<Data> for NeuraGraph<Data> {
+impl<Data: Clone + std::fmt::Debug + 'static> NeuraLayer<Data> for NeuraGraph<Data> {
     type Output = Data;
+    type IntermediaryRepr = ();
 
     fn eval(&self, input: &Data) -> Self::Output {
         let mut buffer = self.create_buffer();
@@ -190,6 +183,21 @@ impl<Data: Clone> NeuraLayer<Data> for NeuraGraph<Data> {
             .take()
             .expect("Unreachable: output was not set")
     }
+
+    #[allow(unused)]
+    fn eval_training(&self, input: &Data) -> (Self::Output, Self::IntermediaryRepr) {
+        unimplemented!("NeuraGraph cannot be used as a trainable layer yet");
+    }
+
+    #[allow(unused)]
+    fn backprop_layer(
+        &self,
+        input: &Data,
+        intermediary: &Self::IntermediaryRepr,
+        epsilon: &Self::Output,
+    ) -> Data {
+        unimplemented!("NeuraGraph cannot be used as a trainable layer yet");
+    }
 }
 
 #[cfg(test)]
diff --git a/src/network/graph/node.rs b/src/network/graph/node.rs
index 9d368f2..da7458f 100644
--- a/src/network/graph/node.rs
+++ b/src/network/graph/node.rs
@@ -4,7 +4,6 @@ use std::{any::Any, fmt::Debug};
 use crate::{
     algebra::NeuraDynVectorSpace,
     err::NeuraAxisErr,
-    layer::{NeuraShapedLayer, NeuraTrainableLayerEval},
     network::residual::{NeuraAxisDefault, NeuraCombineInputs, NeuraSplitInputs},
     prelude::{NeuraPartialLayer, NeuraShape},
 };
@@ -81,9 +80,9 @@ impl<Axis, Layer> NeuraGraphNode<Axis, Layer> {
             + 'static,
         Layer: NeuraPartialLayer + Clone + Debug + 'static,
         Layer::Constructed:
-            NeuraTrainableLayerFull<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
+            NeuraLayer<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
         Layer::Err: Debug,
-        <Layer::Constructed as NeuraTrainableLayerEval<
+        <Layer::Constructed as NeuraLayer<
             <Axis as NeuraCombineInputs<Data>>::Combined,
         >>::IntermediaryRepr: 'static,
         <Axis as NeuraCombineInputs<Data>>::Combined: 'static,
@@ -97,7 +96,7 @@ impl<Axis, Layer> NeuraGraphNode<Axis, Layer> {
     ) -> &'a Intermediary<Axis::Combined, Layer>
     where
         Axis: NeuraCombineInputs<Data>,
-        Layer: NeuraTrainableLayerFull<Axis::Combined>,
+        Layer: NeuraLayer<Axis::Combined>,
         Axis::Combined: 'static,
     {
         intermediary
@@ -106,7 +105,7 @@ impl<Axis, Layer> NeuraGraphNode<Axis, Layer> {
     }
 }
 
-struct Intermediary<Combined, Layer: NeuraTrainableLayerFull<Combined>>
+struct Intermediary<Combined, Layer: NeuraLayer<Combined>>
 where
     Layer::IntermediaryRepr: 'static,
 {
@@ -117,7 +116,7 @@ where
 impl<
         Data: Clone,
         Axis: NeuraSplitInputs<Data> + Clone + Debug,
-        Layer: NeuraTrainableLayerFull<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
+        Layer: NeuraLayer<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
     > NeuraGraphNodeEval<Data> for NeuraGraphNode<Axis, Layer>
 where
     Layer::IntermediaryRepr: 'static,
@@ -188,9 +187,9 @@ impl<
         Layer: NeuraPartialLayer + Clone + Debug,
     > NeuraGraphNodePartial<Data> for NeuraGraphNode<Axis, Layer>
 where
-    Layer::Constructed: NeuraTrainableLayerFull<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
+    Layer::Constructed: NeuraLayer<<Axis as NeuraCombineInputs<Data>>::Combined, Output = Data>,
     Layer::Err: Debug,
-    <Layer::Constructed as NeuraTrainableLayerEval<<Axis as NeuraCombineInputs<Data>>::Combined>>::IntermediaryRepr: 'static,
+    <Layer::Constructed as NeuraLayer<<Axis as NeuraCombineInputs<Data>>::Combined>>::IntermediaryRepr: 'static,
     <Axis as NeuraCombineInputs<Data>>::Combined: 'static,
 {
     fn inputs<'a>(&'a self) -> &'a [String] {
diff --git a/src/network/graph/partial.rs b/src/network/graph/partial.rs
index 99ad1d8..cfaa118 100644
--- a/src/network/graph/partial.rs
+++ b/src/network/graph/partial.rs
@@ -122,7 +122,7 @@ impl<Data> NeuraGraphPartial<Data> {
     }
 }
 
-impl<Data> NeuraPartialLayer for NeuraGraphPartial<Data> {
+impl<Data: Clone + std::fmt::Debug + 'static> NeuraPartialLayer for NeuraGraphPartial<Data> {
     type Constructed = NeuraGraph<Data>;
 
     type Err = NeuraGraphErr;
diff --git a/src/network/residual/construct.rs b/src/network/residual/construct.rs
index b0aedfb..5c1159c 100644
--- a/src/network/residual/construct.rs
+++ b/src/network/residual/construct.rs
@@ -78,33 +78,10 @@ where
     }
 }
 
-impl<Layer, Axis> NeuraShapedLayer for NeuraResidualNode<Layer, (), Axis> {
-    #[inline(always)]
-    fn output_shape(&self) -> NeuraShape {
-        self.output_shape.unwrap()
-    }
-}
-
-impl<Layer, ChildNetwork: NeuraShapedLayer, Axis> NeuraShapedLayer
-    for NeuraResidualNode<Layer, ChildNetwork, Axis>
-{
-    #[inline(always)]
-    fn output_shape(&self) -> NeuraShape {
-        self.child_network.output_shape()
-    }
-}
-
-impl<Layers: NeuraShapedLayer> NeuraShapedLayer for NeuraResidual<Layers> {
-    #[inline(always)]
-    fn output_shape(&self) -> NeuraShape {
-        self.layers.output_shape()
-    }
-}
-
 impl<Layers: NeuraResidualConstruct> NeuraPartialLayer for NeuraResidual<Layers>
 where
     // Should always be satisfied:
-    Layers::Constructed: NeuraShapedLayer,
+    Layers::Constructed: NeuraLayerBase,
 {
     type Constructed = NeuraResidual<Layers::Constructed>;
     type Err = Layers::Err;
diff --git a/src/network/residual/last.rs b/src/network/residual/last.rs
index d6c0e1f..85619a3 100644
--- a/src/network/residual/last.rs
+++ b/src/network/residual/last.rs
@@ -63,7 +63,13 @@ impl NeuraResidualConstruct for NeuraResidualLast {
     }
 }
 
-impl NeuraShapedLayer for NeuraResidualLast {
+impl NeuraLayerBase for NeuraResidualLast {
+    type Gradient = ();
+
+    fn default_gradient(&self) -> Self::Gradient {
+        ()
+    }
+
     fn output_shape(&self) -> NeuraShape {
         self.output_shape
             .expect("Called NeuraResidualLast::output_shape before constructing it")
@@ -90,7 +96,7 @@ impl NeuraNetworkRec for NeuraResidualLast {
     #[inline(always)]
     fn merge_gradient(&self, _rec_gradient: (), _layer_gradient: ()) -> Self::Gradient
     where
-        Self::Layer: NeuraTrainableLayerBase,
+        Self::Layer: NeuraLayerBase,
     {
         ()
     }
@@ -135,62 +141,24 @@ impl<Data: Clone> NeuraNetwork<NeuraResidualInput<Data>> for NeuraResidualLast {
     }
 }
 
-impl NeuraTrainableLayerBase for NeuraResidualLast {
-    type Gradient = ();
-
-    #[inline(always)]
-    fn default_gradient(&self) -> Self::Gradient {
-        ()
-    }
-
-    #[inline(always)]
-    fn apply_gradient(&mut self, _gradient: &Self::Gradient) {
-        // Noop
-    }
-}
-
 impl<Data: Clone> NeuraLayer<NeuraResidualInput<Data>> for NeuraResidualLast {
     type Output = Data;
+    type IntermediaryRepr = ();
 
-    fn eval(&self, input: &NeuraResidualInput<Data>) -> Self::Output {
+    fn eval_training(&self, input: &NeuraResidualInput<Data>) -> (Self::Output, ()) {
         let result: Rc<Self::Output> = input.clone().get_first()
             .expect("Invalid NeuraResidual state: network returned no data, did you forget to link the last layer?")
             .into();
 
-        unwrap_or_clone(result)
+        (unwrap_or_clone(result), ())
     }
-}
 
-impl<Data: Clone> NeuraTrainableLayerEval<NeuraResidualInput<Data>> for NeuraResidualLast {
-    type IntermediaryRepr = ();
-
-    #[inline(always)]
-    fn eval_training(
+    fn backprop_layer(
         &self,
         input: &NeuraResidualInput<Data>,
-    ) -> (Self::Output, Self::IntermediaryRepr) {
-        (self.eval(input), ())
-    }
-}
-
-impl<Data: Clone> NeuraTrainableLayerSelf<NeuraResidualInput<Data>> for NeuraResidualLast {
-    #[inline(always)]
-    fn regularize_layer(&self) -> Self::Gradient {
-        ()
-    }
-
-    #[inline(always)]
-    fn get_gradient(
-        &self,
-        _input: &NeuraResidualInput<Data>,
         _intermediary: &Self::IntermediaryRepr,
-        _epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        ()
+        epsilon: &Self::Output,
+    ) -> NeuraResidualInput<Data> {
+        Cow::into_owned(self.map_gradient_out(input, epsilon, epsilon))
     }
 }
-
-// let epsilon = Rc::new(epsilon.clone());
-// let mut epsilon_residual = NeuraResidualInput::new();
-
-// epsilon_residual.push(0, epsilon);
diff --git a/src/network/residual/node.rs b/src/network/residual/node.rs
index 36ec981..6b04e9c 100644
--- a/src/network/residual/node.rs
+++ b/src/network/residual/node.rs
@@ -1,5 +1,3 @@
-use nalgebra::{DVector, Scalar};
-use num::Float;
 use std::borrow::Cow;
 
 use crate::network::*;
@@ -102,24 +100,6 @@ impl<Layer, ChildNetwork, Axis> NeuraResidualNode<Layer, ChildNetwork, Axis> {
     }
 }
 
-impl<F: Float + Scalar, Layer, ChildNetwork, Axis> NeuraLayer<NeuraResidualInput<DVector<F>>>
-    for NeuraResidualNode<Layer, ChildNetwork, Axis>
-where
-    Axis: NeuraCombineInputs<DVector<F>>,
-    Layer: NeuraLayer<Axis::Combined, Output = DVector<F>>,
-    ChildNetwork: NeuraLayer<NeuraResidualInput<DVector<F>>>,
-{
-    type Output = <ChildNetwork as NeuraLayer<NeuraResidualInput<DVector<F>>>>::Output;
-
-    fn eval(&self, input: &NeuraResidualInput<DVector<F>>) -> Self::Output {
-        let (layer_input, mut rest) = self.process_input(input);
-
-        self.combine_outputs(self.layer.eval(&layer_input), &mut rest);
-
-        self.child_network.eval(&rest)
-    }
-}
-
 #[allow(dead_code)]
 pub struct NeuraResidualIntermediary<LayerIntermediary, LayerOutput, ChildIntermediary> {
     layer_intermediary: LayerIntermediary,
@@ -127,9 +107,18 @@ pub struct NeuraResidualIntermediary<LayerIntermediary, LayerOutput, ChildInterm
     child_intermediary: Box<ChildIntermediary>,
 }
 
-impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase, Axis>
-    NeuraTrainableLayerBase for NeuraResidualNode<Layer, ChildNetwork, Axis>
+impl<
+        Layer: NeuraLayerBase,
+        ChildNetwork: NeuraLayerBase,
+        Axis: Clone + std::fmt::Debug + 'static,
+    > NeuraLayerBase for NeuraResidualNode<Layer, ChildNetwork, Axis>
 {
+    #[inline(always)]
+    fn output_shape(&self) -> NeuraShape {
+        todo!("output_shape for NeuraResidualNode is not yet ready");
+        self.child_network.output_shape()
+    }
+
     type Gradient = (Layer::Gradient, Box<ChildNetwork::Gradient>);
 
     fn default_gradient(&self) -> Self::Gradient {
@@ -148,25 +137,37 @@ impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase, Axis
         self.layer.prepare_layer(is_training);
         self.child_network.prepare_layer(is_training);
     }
+
+    fn regularize_layer(&self) -> Self::Gradient {
+        (
+            self.layer.regularize_layer(),
+            Box::new(self.child_network.regularize_layer()),
+        )
+    }
 }
 
-impl<
-        Data,
-        Axis: NeuraCombineInputs<Data>,
-        Layer: NeuraTrainableLayerEval<Axis::Combined, Output = Data>,
-        ChildNetwork: NeuraTrainableLayerEval<NeuraResidualInput<Data>>,
-    > NeuraTrainableLayerEval<NeuraResidualInput<Data>>
-    for NeuraResidualNode<Layer, ChildNetwork, Axis>
+impl<Data: Clone + 'static, Layer, ChildNetwork, Axis: Clone + std::fmt::Debug + 'static>
+    NeuraLayer<NeuraResidualInput<Data>> for NeuraResidualNode<Layer, ChildNetwork, Axis>
 where
-    NeuraResidualNode<Layer, ChildNetwork, Axis>:
-        NeuraLayer<NeuraResidualInput<Data>, Output = ChildNetwork::Output>,
+    Axis: NeuraCombineInputs<Data>,
+    Layer: NeuraLayer<Axis::Combined, Output = Data>,
+    ChildNetwork: NeuraLayer<NeuraResidualInput<Data>>,
 {
+    type Output = <ChildNetwork as NeuraLayer<NeuraResidualInput<Data>>>::Output;
     type IntermediaryRepr = NeuraResidualIntermediary<
         Layer::IntermediaryRepr,
         Layer::Output,
         ChildNetwork::IntermediaryRepr,
     >;
 
+    fn eval(&self, input: &NeuraResidualInput<Data>) -> Self::Output {
+        let (layer_input, mut rest) = self.process_input(input);
+
+        self.combine_outputs(self.layer.eval(&layer_input), &mut rest);
+
+        self.child_network.eval(&rest)
+    }
+
     fn eval_training(
         &self,
         input: &NeuraResidualInput<Data>,
@@ -186,25 +187,6 @@ where
 
         (output, intermediary)
     }
-}
-
-impl<
-        Data,
-        Axis: NeuraCombineInputs<Data>,
-        Layer: NeuraTrainableLayerSelf<Axis::Combined, Output = Data>,
-        ChildNetwork: NeuraTrainableLayerSelf<NeuraResidualInput<Data>>,
-    > NeuraTrainableLayerSelf<NeuraResidualInput<Data>>
-    for NeuraResidualNode<Layer, ChildNetwork, Axis>
-where
-    NeuraResidualNode<Layer, ChildNetwork, Axis>:
-        NeuraLayer<NeuraResidualInput<Data>, Output = ChildNetwork::Output>,
-{
-    fn regularize_layer(&self) -> Self::Gradient {
-        (
-            self.layer.regularize_layer(),
-            Box::new(self.child_network.regularize_layer()),
-        )
-    }
 
     #[allow(unused)]
     fn get_gradient(
@@ -213,7 +195,17 @@ where
         intermediary: &Self::IntermediaryRepr,
         epsilon: &Self::Output,
     ) -> Self::Gradient {
-        unimplemented!("NeuraResidualNode::get_gradient is not yet implemented, sorry");
+        unimplemented!("NeuraResidualNode::get_gradient is not yet implemented");
+    }
+
+    #[allow(unused)]
+    fn backprop_layer(
+        &self,
+        input: &NeuraResidualInput<Data>,
+        intermediary: &Self::IntermediaryRepr,
+        epsilon: &Self::Output,
+    ) -> NeuraResidualInput<Data> {
+        unimplemented!("NeuraResidualNode::backprop_layer is not yet implemented");
     }
 }
 
@@ -225,8 +217,11 @@ impl<Axis, Layer, ChildNetwork> NeuraNetworkBase for NeuraResidualNode<Layer, Ch
     }
 }
 
-impl<Axis, Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase> NeuraNetworkRec
-    for NeuraResidualNode<Layer, ChildNetwork, Axis>
+impl<
+        Axis: Clone + std::fmt::Debug + 'static,
+        Layer: NeuraLayerBase,
+        ChildNetwork: NeuraLayerBase,
+    > NeuraNetworkRec for NeuraResidualNode<Layer, ChildNetwork, Axis>
 {
     type NextNode = ChildNetwork;
 
@@ -236,8 +231,8 @@ impl<Axis, Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase
 
     fn merge_gradient(
         &self,
-        rec_gradient: <Self::NextNode as NeuraTrainableLayerBase>::Gradient,
-        layer_gradient: <Self::Layer as NeuraTrainableLayerBase>::Gradient,
+        rec_gradient: <Self::NextNode as NeuraLayerBase>::Gradient,
+        layer_gradient: <Self::Layer as NeuraLayerBase>::Gradient,
     ) -> Self::Gradient {
         (layer_gradient, Box::new(rec_gradient))
     }
diff --git a/src/network/residual/wrapper.rs b/src/network/residual/wrapper.rs
index f3cdc7e..5d3b1f4 100644
--- a/src/network/residual/wrapper.rs
+++ b/src/network/residual/wrapper.rs
@@ -1,6 +1,6 @@
 use std::borrow::Cow;
 
-use crate::network::*;
+use crate::{network::*, utils::unwrap_or_clone};
 
 use super::*;
 
@@ -45,18 +45,48 @@ impl<Layers> NeuraResidual<Layers> {
     }
 }
 
-impl<Input: Clone, Layers> NeuraLayer<Input> for NeuraResidual<Layers>
+impl<Data: Clone, Layers> NeuraLayer<Data> for NeuraResidual<Layers>
 where
-    Layers: NeuraLayer<NeuraResidualInput<Input>>,
+    Layers: NeuraLayer<NeuraResidualInput<Data>>,
 {
     type Output = Layers::Output;
+    type IntermediaryRepr = Layers::IntermediaryRepr;
 
-    fn eval(&self, input: &Input) -> Self::Output {
+    fn eval(&self, input: &Data) -> Self::Output {
         self.layers.eval(&self.input_to_residual_input(input))
     }
+
+    fn eval_training(&self, input: &Data) -> (Self::Output, Self::IntermediaryRepr) {
+        self.layers
+            .eval_training(&self.input_to_residual_input(input))
+    }
+
+    fn get_gradient(
+        &self,
+        input: &Data,
+        intermediary: &Self::IntermediaryRepr,
+        epsilon: &Self::Output,
+    ) -> Self::Gradient {
+        self.layers
+            .get_gradient(&self.input_to_residual_input(input), intermediary, &epsilon)
+    }
+
+    fn backprop_layer(
+        &self,
+        input: &Data,
+        intermediary: &Self::IntermediaryRepr,
+        epsilon: &Self::Output,
+    ) -> Data {
+        unwrap_or_clone(
+            self.layers
+                .backprop_layer(&self.input_to_residual_input(input), intermediary, &epsilon)
+                .get_first()
+                .unwrap(),
+        )
+    }
 }
 
-impl<Layers: NeuraTrainableLayerBase> NeuraTrainableLayerBase for NeuraResidual<Layers> {
+impl<Layers: NeuraLayerBase> NeuraLayerBase for NeuraResidual<Layers> {
     type Gradient = Layers::Gradient;
 
     #[inline(always)]
@@ -68,34 +98,13 @@ impl<Layers: NeuraTrainableLayerBase> NeuraTrainableLayerBase for NeuraResidual<
     fn apply_gradient(&mut self, gradient: &Self::Gradient) {
         self.layers.apply_gradient(gradient);
     }
-}
-
-impl<Data: Clone, Layers: NeuraTrainableLayerEval<NeuraResidualInput<Data>>>
-    NeuraTrainableLayerEval<Data> for NeuraResidual<Layers>
-{
-    type IntermediaryRepr = Layers::IntermediaryRepr;
-
-    fn eval_training(&self, input: &Data) -> (Self::Output, Self::IntermediaryRepr) {
-        self.layers
-            .eval_training(&self.input_to_residual_input(input))
-    }
-}
 
-impl<Data: Clone, Layers: NeuraTrainableLayerSelf<NeuraResidualInput<Data>>>
-    NeuraTrainableLayerSelf<Data> for NeuraResidual<Layers>
-{
     fn regularize_layer(&self) -> Self::Gradient {
         self.layers.regularize_layer()
     }
 
-    fn get_gradient(
-        &self,
-        input: &Data,
-        intermediary: &Self::IntermediaryRepr,
-        epsilon: &Self::Output,
-    ) -> Self::Gradient {
-        self.layers
-            .get_gradient(&self.input_to_residual_input(input), intermediary, &epsilon)
+    fn output_shape(&self) -> NeuraShape {
+        self.layers.output_shape()
     }
 }
 
@@ -108,7 +117,7 @@ impl<Layers> NeuraNetworkBase for NeuraResidual<Layers> {
     }
 }
 
-impl<Layers: NeuraTrainableLayerBase> NeuraNetworkRec for NeuraResidual<Layers> {
+impl<Layers: NeuraLayerBase> NeuraNetworkRec for NeuraResidual<Layers> {
     type NextNode = Layers;
 
     #[inline(always)]
@@ -119,8 +128,8 @@ impl<Layers: NeuraTrainableLayerBase> NeuraNetworkRec for NeuraResidual<Layers>
     #[inline(always)]
     fn merge_gradient(
         &self,
-        rec_gradient: <Self::NextNode as NeuraTrainableLayerBase>::Gradient,
-        _layer_gradient: <Self::Layer as NeuraTrainableLayerBase>::Gradient,
+        rec_gradient: <Self::NextNode as NeuraLayerBase>::Gradient,
+        _layer_gradient: <Self::Layer as NeuraLayerBase>::Gradient,
     ) -> Self::Gradient {
         rec_gradient
     }
diff --git a/src/network/sequential/construct.rs b/src/network/sequential/construct.rs
index a8b47eb..f8a50f5 100644
--- a/src/network/sequential/construct.rs
+++ b/src/network/sequential/construct.rs
@@ -1,4 +1,4 @@
-use crate::{err::NeuraRecursiveErr, layer::NeuraShapedLayer};
+use crate::err::NeuraRecursiveErr;
 
 use super::*;
 
@@ -39,19 +39,3 @@ impl<Layer: NeuraPartialLayer, ChildNetwork: NeuraPartialLayer> NeuraPartialLaye
         })
     }
 }
-
-impl<Layer: NeuraShapedLayer> NeuraShapedLayer for NeuraSequential<Layer, ()> {
-    #[inline(always)]
-    fn output_shape(&self) -> NeuraShape {
-        self.layer.output_shape()
-    }
-}
-
-impl<Layer, ChildNetwork: NeuraShapedLayer> NeuraShapedLayer
-    for NeuraSequential<Layer, ChildNetwork>
-{
-    #[inline(always)]
-    fn output_shape(&self) -> NeuraShape {
-        self.child_network.output_shape()
-    }
-}
diff --git a/src/network/sequential/layer_impl.rs b/src/network/sequential/layer_impl.rs
index aed7ef9..1d5aa20 100644
--- a/src/network/sequential/layer_impl.rs
+++ b/src/network/sequential/layer_impl.rs
@@ -1,19 +1,22 @@
 use super::*;
-use crate::layer::{NeuraTrainableLayerBackprop, NeuraTrainableLayerEval};
+use crate::layer::{NeuraLayer, NeuraLayerBase};
 
-impl<Input, Layer: NeuraLayer<Input>, ChildNetwork: NeuraLayer<Layer::Output>> NeuraLayer<Input>
+// impl<Layer: NeuraLayerBase> NeuraLayerBase for NeuraSequential<Layer, ()> {
+//     #[inline(always)]
+//     fn output_shape(&self) -> NeuraShape {
+//         self.layer.output_shape()
+//     }
+// }
+
+impl<Layer: NeuraLayerBase, ChildNetwork: NeuraLayerBase> NeuraLayerBase
     for NeuraSequential<Layer, ChildNetwork>
 {
-    type Output = ChildNetwork::Output;
-
-    fn eval(&self, input: &Input) -> Self::Output {
-        self.child_network.eval(&self.layer.eval(input))
+    #[inline(always)]
+    fn output_shape(&self) -> NeuraShape {
+        todo!("Have output_shape return Option");
+        self.child_network.output_shape()
     }
-}
 
-impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase> NeuraTrainableLayerBase
-    for NeuraSequential<Layer, ChildNetwork>
-{
     type Gradient = (Layer::Gradient, Box<ChildNetwork::Gradient>);
 
     fn default_gradient(&self) -> Self::Gradient {
@@ -32,16 +35,25 @@ impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase> Neur
         self.layer.apply_gradient(&gradient.0);
         self.child_network.apply_gradient(&gradient.1);
     }
+
+    fn regularize_layer(&self) -> Self::Gradient {
+        (
+            self.layer.regularize_layer(),
+            Box::new(self.child_network.regularize_layer()),
+        )
+    }
 }
 
-impl<
-        Input,
-        Layer: NeuraTrainableLayerEval<Input>,
-        ChildNetwork: NeuraTrainableLayerEval<Layer::Output>,
-    > NeuraTrainableLayerEval<Input> for NeuraSequential<Layer, ChildNetwork>
+impl<Input, Layer: NeuraLayer<Input>, ChildNetwork: NeuraLayer<Layer::Output>> NeuraLayer<Input>
+    for NeuraSequential<Layer, ChildNetwork>
 {
+    type Output = ChildNetwork::Output;
     type IntermediaryRepr = (Layer::IntermediaryRepr, Box<ChildNetwork::IntermediaryRepr>);
 
+    fn eval(&self, input: &Input) -> Self::Output {
+        self.child_network.eval(&self.layer.eval(input))
+    }
+
     fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr) {
         let (layer_output, layer_intermediary) = self.layer.eval_training(input);
         let (child_output, child_intermediary) = self.child_network.eval_training(&layer_output);
@@ -51,20 +63,6 @@ impl<
             (layer_intermediary, Box::new(child_intermediary)),
         )
     }
-}
-
-impl<
-        Input,
-        Layer: NeuraTrainableLayerSelf<Input>,
-        ChildNetwork: NeuraTrainableLayerSelf<Layer::Output> + NeuraTrainableLayerBackprop<Layer::Output>,
-    > NeuraTrainableLayerSelf<Input> for NeuraSequential<Layer, ChildNetwork>
-{
-    fn regularize_layer(&self) -> Self::Gradient {
-        (
-            self.layer.regularize_layer(),
-            Box::new(self.child_network.regularize_layer()),
-        )
-    }
 
     fn get_gradient(
         &self,
@@ -74,14 +72,7 @@ impl<
     ) -> Self::Gradient {
         unimplemented!("NeuraSequential::get_gradient is not yet implemented, sorry");
     }
-}
 
-impl<
-        Input,
-        Layer: NeuraTrainableLayerBackprop<Input>,
-        ChildNetwork: NeuraTrainableLayerBackprop<Layer::Output>,
-    > NeuraTrainableLayerBackprop<Input> for NeuraSequential<Layer, ChildNetwork>
-{
     fn backprop_layer(
         &self,
         input: &Input,
diff --git a/src/network/sequential/mod.rs b/src/network/sequential/mod.rs
index d6481a9..a4dca4e 100644
--- a/src/network/sequential/mod.rs
+++ b/src/network/sequential/mod.rs
@@ -1,10 +1,7 @@
 use std::borrow::Cow;
 
 use super::*;
-use crate::layer::{
-    NeuraLayer, NeuraPartialLayer, NeuraShape, NeuraTrainableLayerBase, NeuraTrainableLayerEval,
-    NeuraTrainableLayerSelf,
-};
+use crate::layer::{NeuraLayer, NeuraLayerBase, NeuraPartialLayer, NeuraShape};
 
 mod construct;
 mod layer_impl;
@@ -29,7 +26,7 @@ pub use tail::*;
 /// ## Notes on implemented traits
 ///
 /// The different implementations for `NeuraTrainableNetwork`,
-/// `NeuraLayer` and `NeuraTrainableLayerBase` each require that `ChildNetwork` implements those respective traits,
+/// `NeuraLayer` and `NeuraLayerBase` each require that `ChildNetwork` implements those respective traits,
 /// and that the output type of `Layer` matches the input type of `ChildNetwork`.
 ///
 /// If a method, like `eval`, is reported as missing,
@@ -96,7 +93,7 @@ impl<Layer, ChildNetwork> NeuraNetworkBase for NeuraSequential<Layer, ChildNetwo
     }
 }
 
-impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase> NeuraNetworkRec
+impl<Layer: NeuraLayerBase, ChildNetwork: NeuraLayerBase> NeuraNetworkRec
     for NeuraSequential<Layer, ChildNetwork>
 {
     type NextNode = ChildNetwork;
@@ -107,14 +104,14 @@ impl<Layer: NeuraTrainableLayerBase, ChildNetwork: NeuraTrainableLayerBase> Neur
 
     fn merge_gradient(
         &self,
-        rec_gradient: <Self::NextNode as NeuraTrainableLayerBase>::Gradient,
-        layer_gradient: <Self::Layer as NeuraTrainableLayerBase>::Gradient,
+        rec_gradient: <Self::NextNode as NeuraLayerBase>::Gradient,
+        layer_gradient: <Self::Layer as NeuraLayerBase>::Gradient,
     ) -> Self::Gradient {
         (layer_gradient, Box::new(rec_gradient))
     }
 }
 
-impl<Input: Clone, Layer: NeuraTrainableLayerEval<Input>, ChildNetwork> NeuraNetwork<Input>
+impl<Input: Clone, Layer: NeuraLayer<Input>, ChildNetwork> NeuraNetwork<Input>
     for NeuraSequential<Layer, ChildNetwork>
 where
     Layer::Output: Clone,
diff --git a/src/network/traits.rs b/src/network/traits.rs
index c2f25f9..d9d0f79 100644
--- a/src/network/traits.rs
+++ b/src/network/traits.rs
@@ -46,18 +46,18 @@ where
     ) -> Cow<'a, NodeInput>;
 }
 
-pub trait NeuraNetworkRec: NeuraNetworkBase + NeuraTrainableLayerBase {
+pub trait NeuraNetworkRec: NeuraNetworkBase + NeuraLayerBase {
     /// The type of the children network, it does not need to implement `NeuraNetworkBase`,
     /// although many functions will expect it to be either `()` or an implementation of `NeuraNetworkRec`.
-    type NextNode: NeuraTrainableLayerBase;
+    type NextNode: NeuraLayerBase;
 
     fn get_next(&self) -> &Self::NextNode;
 
     fn merge_gradient(
         &self,
-        rec_gradient: <Self::NextNode as NeuraTrainableLayerBase>::Gradient,
-        layer_gradient: <Self::Layer as NeuraTrainableLayerBase>::Gradient,
+        rec_gradient: <Self::NextNode as NeuraLayerBase>::Gradient,
+        layer_gradient: <Self::Layer as NeuraLayerBase>::Gradient,
     ) -> Self::Gradient
     where
-        Self::Layer: NeuraTrainableLayerBase;
+        Self::Layer: NeuraLayerBase;
 }
diff --git a/src/train.rs b/src/train.rs
index cd03dfc..b64b4d2 100644
--- a/src/train.rs
+++ b/src/train.rs
@@ -70,7 +70,7 @@ impl NeuraBatchedTrainer {
     pub fn train<
         Input: Clone,
         Target: Clone,
-        Network: NeuraTrainableLayerBase + NeuraTrainableLayerSelf<Input>,
+        Network: NeuraLayer<Input>,
         GradientSolver: NeuraGradientSolver<Input, Target, Network>,
         Inputs: IntoIterator<Item = (Input, Target)>,
     >(