(paper) Transfer Learning for Time Series Classification

Transfer Learning for Time Series Classification

Contents

1. Abstract
2. Method
   1. Architecture
   2. Network Adaptation
   3. Inter-datsaet simliarity

0. Abstract

how to transfer deep CNNs for the TSC task

• Retrain a model & fine tune with other datasets
• total 85 datasets

1. Method

Notation

• TS : \(X=\left[x_1, x_2, \ldots x_T\right]\)
• dataset : \(D=\left\{\left(X_1, Y_1\right), \ldots,\left(X_N, Y_N\right)\right\}\)

1. adopted NN
2. how we adapt the network for TL process
3. DTW based method to compute inter-dataset similarities

(1) Architecture

• 1d FCN (Fully Convolutional NN)
• model input & output
  • input : TS of variable length
  • output : probability distn over \(C\) classes

(2) Network Adaptation

Procedure

• step 1) train NN on \(D_S\)

• step 2) remove the last layer

• step 3) add new softmax layer

• step 4) retrain (fine-tune) using \(D_T\)

Advantages of using a GAP layer

• do not need to re-scale the input TS, when tansferring models between TS of different length

(3) Inter-datsaet simliarity

challenge : which to choose as a source dataset ??

( Total 85 datasets… 1 target domain & 84 possible source domain )

\(\rightarrow\) propose to use DTW distance to compute simliarities between datasets

Step 1) reduce the number of TS for each dataset to 1 TS per class ( = prototype )

• computed by averaging the set of TS in the certian class
• use DTW Barycenter Averaging (DBA) method

Step 2) calculate distance

• distance between 2 datasets

  = minimum distance between the prototypes of their corresponding classes