diff --git a/examples/convolution.rs b/examples/convolution.rs
index a19e739..e59d38f 100644
--- a/examples/convolution.rs
+++ b/examples/convolution.rs
@@ -2,7 +2,7 @@
 #![feature(generic_const_exprs)]
 
 use neuramethyst::algebra::NeuraVector;
-use neuramethyst::derivable::reduce::Average;
+use neuramethyst::derivable::reduce::{Average, Max};
 use rust_mnist::Mnist;
 
 use neuramethyst::derivable::activation::{Linear, Relu};
@@ -55,17 +55,22 @@ fn main() {
 
     let mut network = neura_sequential![
         neura_layer!("unstable_reshape", 1, { 28 * 28 }),
-        neura_layer!("conv2d_pad", 1, {28 * 28}; 28, 3; neura_layer!("dense", {1 * 3 * 3}, 10; Relu)),
-        // neura_layer!("conv2d_pad", 28, 1; neura_layer!("dense", {30 * 1 * 1}, 10; Relu)),
+        neura_layer!("conv2d_pad", 1, {28 * 28}; 28, 3; neura_layer!("dense", {1 * 3 * 3}, 3; Relu)),
+        // neura_layer!("conv2d_block", 7, 7; 4; neura_layer!("dense", {3 * 4 * 4}, 8; Relu)),
+        // neura_layer!("conv2d_pad"; 28, 1; neura_layer!("dense", {30 * 1 * 1}, 10; Relu)),
+        neura_layer!("unstable_flatten"),
+        neura_layer!("dropout", 0.33),
+        neura_layer!("unstable_reshape", 3, { 28 * 28 }),
+        neura_layer!("conv2d_block", 14, 14; 2; neura_layer!("dense", {3 * 2 * 2}, 2; Relu)),
+        // neura_layer!("unstable_flatten"),
+        // neura_layer!("dropout", 0.33),
+        // neura_layer!("unstable_reshape", 2, { 14 * 14 }),
+        // neura_layer!("conv2d_pad"; 14, 5; neura_layer!("dense", {2 * 5 * 5}, 20; Relu)),
+        // neura_layer!("pool_global"; Max),
 
-        // neura_layer!("pool_global", 10, {28 * 28}; Average),
-        // neura_layer!("pool1d", 10, 28, 28; Average),
-        // neura_layer!("unstable_flatten", 10, 28),
-        neura_layer!("unstable_flatten", 10, { 28 * 28 }),
-        // neura_layer!("dense", 100; Relu),
-        // neura_layer!("dropout", 0.5),
-        // neura_layer!("dense", 30; Relu),
-        // neura_layer!("dropout", 0.5),
+        // neura_layer!("pool1d", {14 * 2}, 7; Max),
+        neura_layer!("unstable_flatten"),
+        neura_layer!("dropout", 0.2),
         neura_layer!("dense", 10; Linear),
         neura_layer!("softmax")
     ];
diff --git a/src/derivable/reduce.rs b/src/derivable/reduce.rs
index c692147..1070806 100644
--- a/src/derivable/reduce.rs
+++ b/src/derivable/reduce.rs
@@ -1,3 +1,5 @@
+use crate::utils::{argmax, one_hot};
+
 use super::*;
 
 #[derive(Clone, Copy, Debug, PartialEq)]
@@ -18,3 +20,25 @@ impl NeuraReducer<f64> for Average {
         NeuraVector::from_value(1.0 / inputs.len() as f64)
     }
 }
+
+#[derive(Clone, Copy, Debug, PartialEq)]
+pub struct Max;
+
+impl NeuraReducer<f64> for Max {
+    #[inline(always)]
+    fn eval<const LENGTH: usize>(&self, inputs: NeuraVector<LENGTH, f64>) -> f64 {
+        let mut max = 0.0;
+        for &i in inputs.iter() {
+            max = i.max(max);
+        }
+        max
+    }
+
+    #[inline(always)]
+    fn nabla<const LENGTH: usize>(
+        &self,
+        inputs: NeuraVector<LENGTH, f64>,
+    ) -> NeuraVector<LENGTH, f64> {
+        one_hot(argmax(inputs.as_ref()))
+    }
+}
diff --git a/src/layer/convolution.rs b/src/layer/convolution.rs
index 1e14905..c8fde6b 100644
--- a/src/layer/convolution.rs
+++ b/src/layer/convolution.rs
@@ -167,7 +167,7 @@ pub struct NeuraConv2DPadLayer<
     /// The width of the image, in grid units.
     ///
     /// **Class invariant:** `LAYER % width == 0`, `width > 0`
-    pub width: NonZeroUsize,
+    width: NonZeroUsize,
 }
 
 impl<
@@ -192,6 +192,10 @@ impl<
         }
     }
 
+    pub fn width(&self) -> NonZeroUsize {
+        self.width
+    }
+
     /// Iterates within the `(WINDOW, WINDOW)` window centered around `x, y`;
     /// Returns a 4-uple `(x' = x + δx, y' = y + δy, δy * WINDOW + δ, y' * width + x')`, with the last element
     /// being set to `None` if `x'` or `y'` are out of bound.
@@ -338,3 +342,148 @@ where
         self.inner_layer.cleanup();
     }
 }
+
+#[non_exhaustive]
+#[derive(Clone, Debug)]
+pub struct NeuraConv2DBlockLayer<
+    const WIDTH: usize,
+    const HEIGHT: usize,
+    const IN_FEATS: usize,
+    const BLOCK_SIZE: usize,
+    Layer: NeuraLayer<Input = NeuraVector<{ IN_FEATS * BLOCK_SIZE * BLOCK_SIZE }, f64>>,
+> {
+    pub inner_layer: Layer,
+}
+
+impl<
+        const WIDTH: usize,
+        const HEIGHT: usize,
+        const IN_FEATS: usize,
+        const BLOCK_SIZE: usize,
+        Layer: NeuraLayer<Input = NeuraVector<{ IN_FEATS * BLOCK_SIZE * BLOCK_SIZE }, f64>>,
+    > NeuraConv2DBlockLayer<WIDTH, HEIGHT, IN_FEATS, BLOCK_SIZE, Layer>
+{
+    pub fn new(layer: Layer) -> Self {
+        Self { inner_layer: layer }
+    }
+
+    fn iterate_blocks<'a>(
+        &'a self,
+        input: &'a NeuraMatrix<IN_FEATS, { WIDTH * HEIGHT * BLOCK_SIZE * BLOCK_SIZE }, f64>,
+    ) -> impl Iterator<
+        Item = (
+            usize,
+            usize,
+            usize,
+            NeuraVector<{ IN_FEATS * BLOCK_SIZE * BLOCK_SIZE }, f64>,
+        ),
+    > + 'a {
+        (0..HEIGHT).flat_map(move |y| {
+            (0..WIDTH).map(move |x| {
+                let output_index = y * WIDTH + x;
+                let mut virtual_input = NeuraVector::default();
+
+                for dy in 0..BLOCK_SIZE {
+                    let y = y * BLOCK_SIZE + dy;
+                    for dx in 0..BLOCK_SIZE {
+                        let x = x * BLOCK_SIZE + dx;
+                        let virtual_index = dy * BLOCK_SIZE + dx;
+
+                        for k in 0..IN_FEATS {
+                            virtual_input[virtual_index * IN_FEATS + k] = input[y * WIDTH + x][k];
+                        }
+                    }
+                }
+
+                (x, y, output_index, virtual_input)
+            })
+        })
+    }
+}
+
+impl<
+        const WIDTH: usize,
+        const HEIGHT: usize,
+        const IN_FEATS: usize,
+        const OUT_FEATS: usize,
+        const BLOCK_SIZE: usize,
+        Layer: NeuraLayer<
+            Input = NeuraVector<{ IN_FEATS * BLOCK_SIZE * BLOCK_SIZE }, f64>,
+            Output = NeuraVector<OUT_FEATS, f64>,
+        >,
+    > NeuraLayer for NeuraConv2DBlockLayer<WIDTH, HEIGHT, IN_FEATS, BLOCK_SIZE, Layer>
+where
+    [f64; WIDTH * HEIGHT * BLOCK_SIZE * BLOCK_SIZE]: Sized,
+{
+    type Input = NeuraMatrix<IN_FEATS, { WIDTH * HEIGHT * BLOCK_SIZE * BLOCK_SIZE }, f64>;
+
+    type Output = NeuraMatrix<OUT_FEATS, { WIDTH * HEIGHT }, f64>;
+
+    fn eval(&self, input: &Self::Input) -> Self::Output {
+        let mut res = Self::Output::default();
+
+        for (_, _, output_index, virtual_input) in self.iterate_blocks(input) {
+            res.set_row(output_index, self.inner_layer.eval(&virtual_input));
+        }
+
+        res
+    }
+}
+
+impl<
+        const WIDTH: usize,
+        const HEIGHT: usize,
+        const IN_FEATS: usize,
+        const OUT_FEATS: usize,
+        const BLOCK_SIZE: usize,
+        Layer: NeuraLayer<
+            Input = NeuraVector<{ IN_FEATS * BLOCK_SIZE * BLOCK_SIZE }, f64>,
+            Output = NeuraVector<OUT_FEATS, f64>,
+        >,
+    > NeuraTrainableLayer for NeuraConv2DBlockLayer<WIDTH, HEIGHT, IN_FEATS, BLOCK_SIZE, Layer>
+where
+    [f64; WIDTH * HEIGHT * BLOCK_SIZE * BLOCK_SIZE]: Sized,
+    Layer: NeuraTrainableLayer,
+{
+    type Delta = Layer::Delta;
+
+    fn backpropagate(
+        &self,
+        input: &Self::Input,
+        epsilon: Self::Output,
+    ) -> (Self::Input, Self::Delta) {
+        let mut gradient_sum = Layer::Delta::zero();
+        let mut next_epsilon = Self::Input::default();
+
+        for (x, y, output_index, virtual_input) in self.iterate_blocks(input) {
+            let (layer_next_epsilon, gradient) = self
+                .inner_layer
+                .backpropagate(&virtual_input, epsilon.get_row(output_index));
+
+            gradient_sum.add_assign(&gradient);
+
+            for dy in 0..BLOCK_SIZE {
+                let y = y * BLOCK_SIZE + dy;
+                for dx in 0..BLOCK_SIZE {
+                    let x = x * BLOCK_SIZE + dx;
+                    let input_index = y * WIDTH + x;
+
+                    for k in 0..IN_FEATS {
+                        next_epsilon[input_index][k] =
+                            layer_next_epsilon[(dy * BLOCK_SIZE + dx) * IN_FEATS + k];
+                    }
+                }
+            }
+        }
+
+        (next_epsilon, gradient_sum)
+    }
+
+    fn regularize(&self) -> Self::Delta {
+        self.inner_layer.regularize()
+    }
+
+    fn apply_gradient(&mut self, gradient: &Self::Delta) {
+        self.inner_layer.apply_gradient(gradient);
+    }
+}
diff --git a/src/layer/mod.rs b/src/layer/mod.rs
index 2b08a33..c89d835 100644
--- a/src/layer/mod.rs
+++ b/src/layer/mod.rs
@@ -2,7 +2,7 @@ mod dense;
 pub use dense::NeuraDenseLayer;
 
 mod convolution;
-pub use convolution::{NeuraConv1DPadLayer, NeuraConv2DPadLayer};
+pub use convolution::{NeuraConv1DPadLayer, NeuraConv2DBlockLayer, NeuraConv2DPadLayer};
 
 mod dropout;
 pub use dropout::NeuraDropoutLayer;
@@ -124,6 +124,14 @@ macro_rules! neura_layer {
         $crate::layer::NeuraConv2DPadLayer::new($layer, Default::default(), $width) as $crate::layer::NeuraConv2DPadLayer<_, _, $window, _>
     };
 
+    ( "conv2d_block", $feats:expr, $width:expr, $height:expr; $block_size:expr; $layer:expr ) => {
+        $crate::layer::NeuraConv2DBlockLayer::new($layer) as $crate::layer::NeuraConv2DBlockLayer<$width, $height, $feats, $block_size, _>
+    };
+
+    ( "conv2d_block", $width:expr, $height:expr; $block_size:expr; $layer:expr ) => {
+        $crate::layer::NeuraConv2DBlockLayer::new($layer) as $crate::layer::NeuraConv2DBlockLayer<$width, $height, _, $block_size, _>
+    };
+
     ( "pool_global"; $reduce:expr ) => {
         $crate::layer::NeuraGlobalPoolLayer::new($reduce) as $crate::layer::NeuraGlobalPoolLayer<_, _, _>
     };
diff --git a/src/layer/pool.rs b/src/layer/pool.rs
index 22ad4ae..9c1f056 100644
--- a/src/layer/pool.rs
+++ b/src/layer/pool.rs
@@ -150,7 +150,7 @@ where
                     block_input[k] = input[block * BLOCK_LENGTH + k];
                 }
 
-                let mut gradient = self.reducer.nabla(block_input);
+                let gradient = self.reducer.nabla(block_input);
 
                 for k in 0..BLOCK_LENGTH {
                     column_gradient[block * BLOCK_LENGTH + k] = gradient[k] * epsilon[block][j];