Machine Learning from Perspective of Information Theory
New description of components based on machine learning objectives:
Component | Function |
---|---|
information source | produce messages based on distribution |
transmitter | operates on messages to produce signal |
channel | medium to transmit signal |
receiver | reconstruct messages from signal |
destination | consume messages to reconstruct distribution |
Specific change:
In general communication system:
destination consumes messages
To learn source distribution:
destination consumes messages to reconstruct distribution
Parallels
New description of components based on machine learning objectives:
Component | Function |
---|---|
information source | distribution of data to be learned |
transmitter | draws sample + supervision from distribution |
channel | operates on samples to produce embeddings |
receiver | recover supervision from embedding |
destination | compare tx/rx supervision to learn distribution |
Corresponding machine learning concepts:
Component | Machine Learning concept |
---|---|
information source | dataset |
transmitter | sampling strategy |
channel | model (conv layers) |
receiver | model (fc layers) |
destination | loss function |