Update app.py

app.py

@@ -298,22 +298,27 @@ def create_interface():
                 gr.Markdown("""
                 # SAErch: Sparse Autoencoder-enhanced Semantic Search
 
-                Welcome to SAErch, an innovative approach to semantic search using Sparse Autoencoders (SAEs) trained on dense text embeddings.
+                Welcome to SAErch, an innovative approach to semantic search using Sparse Autoencoders (SAEs) trained on dense text embeddings. This tool builds upon recent advancements in the application of SAEs to language models and embeddings.
 
                 ## Key Concepts:
 
-                1. **Sparse Autoencoders (SAEs)**: Neural networks that learn to reconstruct input data using a sparse set of features, helping to disentangle complex representations.
+                1. **Sparse Autoencoders (SAEs)**: Neural networks that learn to reconstruct input data using a sparse set of features, helping to disentangle complex representations. SAEs have shown promising results in uncovering interpretable features in language models.
 
-                2. **Feature Families**: Groups of related SAE features that represent concepts at varying levels of abstraction.
+                2. **Feature Families**: Groups of related SAE features that represent concepts at varying levels of abstraction, allowing for multi-scale semantic analysis and manipulation.
 
-                3. **Embedding Interventions**: Technique to modify search queries by manipulating specific semantic features identified by the SAE.
+                3. **Embedding Interventions**: Technique to modify search queries by manipulating specific semantic features identified by the SAE, enabling fine-grained control over query semantics.
 
                 ## How It Works:
 
-                1. SAEs are trained on embeddings from scientific paper abstracts.
-                2. The SAE learns interpretable features that capture various semantic concepts.
-                3. Users can interact with these features to fine-tune search queries.
-                4. The system performs semantic search using the modified embeddings.
+                1. SAEs are trained on embeddings from scientific paper abstracts, learning interpretable features that capture various semantic concepts.
+                2. Users can interact with these features to fine-tune search queries.
+                3. The system performs semantic search using the modified embeddings, allowing for more precise and controllable results.
+
+                ## Key References:
+
+                - [Towards Monosemanticity: Decomposing Language Models With Dictionary Learning](https://transformer-circuits.pub/2023/monosemantic-features) - Anthropic's pioneering work on applying SAEs to language models.
+                - [Prism: Mapping Interpretable Concepts and Features in a Latent Space of Language](https://thesephist.com/posts/prism/#caveats-and-limitations) - An early application of SAEs to embeddings, demonstrating their potential for interpretable concept mapping.
+                - [Scaling and Evaluating Sparse Autoencoders](https://arxiv.org/html/2406.04093v1) - OpenAI's research on scaling SAEs, showcasing the effectiveness of top-k SAEs.
 
                 Explore the "SAErch" tab to try out the semantic search capabilities, or dive into the "Feature Visualisation" tab to examine the learned features in more detail.