🪴 jaden lorenc

Search

Search IconIcon to open search

PatchMixer

Last updated Feb 8, 2024 Edit Source

#work/patientsim #paper_notes 2024-02-05

A convolutional neural network!! Haven’t seen one in a minute, and certainly didn’t expect it in time series processing.

PatchMixer-G578PTIA

Seems like the main innovation is in its patch-mixing, used as an alternative to self-attention. Patch-mixing relies on convolutions. First, there’s convolutions over each patch, then convolutions over those representations of the patches.

# 1. Convolutions Over Each Patch:

# 2. Convolutions Over Representations of the Patches:

This is not for heterogeneous data. seems like it’s mostly for continuous time series. In fact, it’s channel-independent, completely separating different features and not using them for predicting each other. Not ideal for patient similarity representation. PatchMixer-G578PTIA

[2310.00655] PatchMixer: A Patch-Mixing Architecture for Long-Term Time Series Forecasting