Lifelong Adaptive Machine Learning for Sensor-Based Human Activity Recognition Using Prototypical Networks

. 2022 Sep 12;22(18):6881. doi: 10.3390/s22186881

Algorithm 1: LAPNet-HAR
Input: Initial network parameters $θ_{0}$ , hyperparameter $α$ , replay buffer size/class b
Data: Training data $(X, Y) = D$
$(X_{0}, Y_{0}) = D_{0}$	`// get base data for pretraining`
$C_{b a s e} =$ set of base classes in $D_{0}$
$M_{p} = {}; M_{r} = {}$	`//initialize empty prototype memory and replay`
`buffer`

`/* offline pretraining process`	`*/`
Update $θ_{0}$ with $(X_{0}, Y_{0})$
Store prototypes $p_{k}$ for $k \in C_{b a s e}$ in $M_{p}$ using $D_{0}$
Sample data from $D_{0}$ and store in $M_{r}$ for replay

while continually learning do
$(X_{t}, Y_{t}) = D_{t}$
$M_{p} \leftarrow$ UpdatePrototypeMemory( $D_{t}$ , $M_{p}$ , $θ_{t - 1}$ )	`// Figure 2a &`
`Equation (3)`
$Q \leftarrow D_{t} \cup M_{r}$ ; $(X_{q}, Y_{q}) = Q$	`// form combined query set`
Incur loss $L (f_{θ_{t - 1}} (X_{q}), Y_{q})$	`// Equation (5)`
Update model $θ_{t}$ with $(X_{q}, Y_{q})$
$M_{p} \leftarrow$ PrototypeAdaptation( $M_{r}$ , $M_{p}$ , $θ_{t}$ , $α$ )	`// Equation (6)`
$M_{r} \leftarrow$ UpdateReplayBuffer(Q, $M_{r}$ , $θ_{t}$ , b)
end