{% extends "base.html" %} {% block title %}Feature Topology ยท Run #{{ run.id }}{% endblock %} {% block extra_head %} {% endblock %} {% block content %}
Structural Analysis

TOPOLOGY GRAPH / {{ run.families.count }} CLUSTERS

Re-Build Topology

This page visualizes the semantic topology of the features. It groups features into families based on their co-occurrence and correlation, allowing you to explore the relationships between different concepts learned by the SAE.

{% if families %}
Scroll to zoom, Drag to move, Click to view details.
Initializing Physics Engine...
{% else %}

No Topology Built

Run the MST algorithm to detect semantic clusters in the SAE latent space.

Start Graph Algorithm
{% endif %}
{% if run.matrix_s_heatmap or run.matrix_c_heatmap or run.matrix_d_heatmap %}
Matrix Analysis
Similarity (S)
Decoder Weights Cosine Sim
$$S = W_{dec} W_{dec}^T$$
{% if run.matrix_s_heatmap %} S Matrix {% else %} Not Available {% endif %}
Co-occurrence (C)
Normalized Co-activations
$$C_{ij}^{norm} = \frac{C_{ij}}{f_i + \epsilon}$$
{% if run.matrix_c_heatmap %} C Matrix {% else %} Not Available {% endif %}
Dense Similarity (D)
Activation Value Similarity
$$D_{ij} = \frac{V_i \cdot V_j}{\|V_i\| \|V_j\|}$$
{% if run.matrix_d_heatmap %} D Matrix {% else %} Not Available {% endif %}
{% if run.matrix_s_heatmap %} {% endif %} {% if run.matrix_c_heatmap %} {% endif %} {% if run.matrix_d_heatmap %} {% endif %} {% endif %} {% endblock %}