nmn 0.1.1__tar.gz → 0.1.2__tar.gz

{nmn-0.1.1 → nmn-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nmn
-Version: 0.1.1
+Version: 0.1.2
 Summary: a neuron that matter
 Project-URL: Homepage, https://github.com/mlnomadpy/nmn
 Project-URL: Bug Tracker, https://github.com/mlnomadpy/my_package/issues
@@ -30,6 +30,40 @@ Not the neurons we want, but the neurons we need
 
 > Deep Learning 2.0: Artificial Neurons that Matter: Reject Correlation - Embrace Orthogonality
 
+## Math
+
+Yat-Product: 
+$$
+ⵟ(\mathbf{w},\mathbf{x}) := \frac{\langle \mathbf{w}, \mathbf{x} \rangle^2}{\|\mathbf{w} - \mathbf{x}\|^2 + \epsilon} = \frac{ \|\mathbf{x}\|^2  \|\mathbf{w}\|^2 \cos^2 \theta}{\|\mathbf{w}\|^2 - 2\mathbf{w}^\top\mathbf{x} + \|\mathbf{x}\|^2 + \epsilon} = \frac{ \|\mathbf{x}\|^2  \|\mathbf{w}\|^2 \cos^2 \theta}{((\mathbf{x}-\mathbf{w})\cdot(\mathbf{x}-\mathbf{w})) + \epsilon}.
+$$
+
+**Explanation:**
+- $\mathbf{w}$ is the weight vector, $\mathbf{x}$ is the input vector.
+- $\langle \mathbf{w}, \mathbf{x} \rangle$ is the dot product between $\mathbf{w}$ and $\mathbf{x}$.
+- $\|\mathbf{w} - \mathbf{x}\|^2$ is the squared Euclidean distance between $\mathbf{w}$ and $\mathbf{x}$.
+- $\epsilon$ is a small constant for numerical stability.
+- $\theta$ is the angle between $\mathbf{w}$ and $\mathbf{x}$.
+
+This operation:
+- **Numerator:** Squares the similarity (dot product) between $\mathbf{w}$ and $\mathbf{x}$, emphasizing strong alignments.
+- **Denominator:** Penalizes large distances, so the response is high only when $\mathbf{w}$ and $\mathbf{x}$ are both similar in direction and close in space.
+- **No activation needed:** The non-linearity is built into the operation itself, allowing the layer to learn complex, non-linear relationships without a separate activation function.
+- **Geometric view:** The output is maximized when $\mathbf{w}$ and $\mathbf{x}$ are both large in norm, closely aligned (small $\theta$), and close together in Euclidean space.
+
+Yat-Conv:
+$$
+ⵟ^*(\mathbf{W}, \mathbf{X}) := \frac{\langle \mathbf{w}, \mathbf{x} \rangle^2}{\|\mathbf{w} - \mathbf{x}\|^2 + \epsilon}
+= \frac{\left(\sum_{i,j} w_{ij} x_{ij}\right)^2}{\sum_{i,j} (w_{ij} - x_{ij})^2 + \epsilon}
+$$
+
+Where:
+- $\mathbf{W}$ and $\mathbf{X}$ are local patches (e.g., kernel and input patch in convolution)
+- $w_{ij}$ and $x_{ij}$ are elements of the kernel and input patch, respectively
+- $\epsilon$ is a small constant for numerical stability
+
+This generalizes the Yat-product to convolutional (patch-wise) operations.
+
+
 ## Supported Frameworks & Tasks
 
 ### Flax (JAX)

{nmn-0.1.1 → nmn-0.1.2}/README.md

@@ -16,6 +16,40 @@ Not the neurons we want, but the neurons we need
 
 > Deep Learning 2.0: Artificial Neurons that Matter: Reject Correlation - Embrace Orthogonality
 
+## Math
+
+Yat-Product: 
+$$
+ⵟ(\mathbf{w},\mathbf{x}) := \frac{\langle \mathbf{w}, \mathbf{x} \rangle^2}{\|\mathbf{w} - \mathbf{x}\|^2 + \epsilon} = \frac{ \|\mathbf{x}\|^2  \|\mathbf{w}\|^2 \cos^2 \theta}{\|\mathbf{w}\|^2 - 2\mathbf{w}^\top\mathbf{x} + \|\mathbf{x}\|^2 + \epsilon} = \frac{ \|\mathbf{x}\|^2  \|\mathbf{w}\|^2 \cos^2 \theta}{((\mathbf{x}-\mathbf{w})\cdot(\mathbf{x}-\mathbf{w})) + \epsilon}.
+$$
+
+**Explanation:**
+- $\mathbf{w}$ is the weight vector, $\mathbf{x}$ is the input vector.
+- $\langle \mathbf{w}, \mathbf{x} \rangle$ is the dot product between $\mathbf{w}$ and $\mathbf{x}$.
+- $\|\mathbf{w} - \mathbf{x}\|^2$ is the squared Euclidean distance between $\mathbf{w}$ and $\mathbf{x}$.
+- $\epsilon$ is a small constant for numerical stability.
+- $\theta$ is the angle between $\mathbf{w}$ and $\mathbf{x}$.
+
+This operation:
+- **Numerator:** Squares the similarity (dot product) between $\mathbf{w}$ and $\mathbf{x}$, emphasizing strong alignments.
+- **Denominator:** Penalizes large distances, so the response is high only when $\mathbf{w}$ and $\mathbf{x}$ are both similar in direction and close in space.
+- **No activation needed:** The non-linearity is built into the operation itself, allowing the layer to learn complex, non-linear relationships without a separate activation function.
+- **Geometric view:** The output is maximized when $\mathbf{w}$ and $\mathbf{x}$ are both large in norm, closely aligned (small $\theta$), and close together in Euclidean space.
+
+Yat-Conv:
+$$
+ⵟ^*(\mathbf{W}, \mathbf{X}) := \frac{\langle \mathbf{w}, \mathbf{x} \rangle^2}{\|\mathbf{w} - \mathbf{x}\|^2 + \epsilon}
+= \frac{\left(\sum_{i,j} w_{ij} x_{ij}\right)^2}{\sum_{i,j} (w_{ij} - x_{ij})^2 + \epsilon}
+$$
+
+Where:
+- $\mathbf{W}$ and $\mathbf{X}$ are local patches (e.g., kernel and input patch in convolution)
+- $w_{ij}$ and $x_{ij}$ are elements of the kernel and input patch, respectively
+- $\epsilon$ is a small constant for numerical stability
+
+This generalizes the Yat-product to convolutional (patch-wise) operations.
+
+
 ## Supported Frameworks & Tasks
 
 ### Flax (JAX)

{nmn-0.1.1 → nmn-0.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "nmn"
-version = "0.1.1"
+version = "0.1.2"
 authors = [
     { name="Taha Bouhsine", email="yat@mlnomads.com" },
 ]