Embedding Datums

Overview

Embedding is a process of transforming high-dimensional data into a lower-dimensional space, often to simplify data analysis and visualization while preserving relevant information. In the context of Acadia, embedding datums (data points) helps in efficiently summarizing and comparing complex data such as text and images.

Why Embed Datums?

Embedding datums facilitates various machine learning and data analysis tasks, including:

Dimensionality Reduction: Reduces computational complexity and resource usage.
Visualization: Simplifies the representation of data, making it easier to visualize and understand complex patterns.
Similarity Measurement: Enhances the ability to measure similarity between datums by using their embedded vectors, which is useful in clustering and recommendation systems.

How to Embed Datums?

The process of embedding datums involves transforming the original data into a vector of continuous features. This transformation is typically done using a model specifically trained to capture the essential characteristics of the data.

Step-by-step process to embed datums in Acadia:

Define a Dimension Reduction Model: Before embedding, you often need to define how the high-dimensional data will be compressed. In Acadia, you can utilize dimension reduction models like PCA, t-SNE, or custom models that fit your data characteristics.
```
dim_reduction_model = MockDimensionReductionModel(
    dims=2, normalize=True, hyperparameter_1=0.5, hyperparameter_2=0.5
)
```

Define a Datum Embedding Model: With a dimension reduction model in place, you then specify a datum embedding model. This model determines how each type of data (e.g., text, images) is processed and transformed into embeddings.

embedding_model = MockDatumEmbeddingModel(
    task_context="This is some context to embed the image with",
    columns_to_embed={
        "caption_0": "text",
        "caption_1": "text",
        "image_0": "image",
    },
    dim_reduction_model=dim_reduction_model,
)

Embed Datums: Once the embedding model is defined, you can apply it to datums in your dataset. This step transforms each datum into a vector that represents its content in the embedded space.
```
acadia.datum.embed_datums(dataset, embedding_model)
```
After embedding, you can retrieve and inspect the embeddings:
```
print("Datums: ")
for datum in dataset.datums[:5]:
    print(datum.id, datum.embedding.embedding_value)
print("...")
```

This process integrates the datum embeddings directly into your data pipeline, enabling further analysis or machine learning tasks to be performed more effectively.

Creating Custom Dimension Reduction Models Creating Custom Datum Embedding Models