From: Spatial and temporal learning representation for end-to-end recording device identification
Optimization of end-to-end parameter update algorithm based on deep-and-shallow loss |
---|
dataset: 45 device sources, each device source is 514 sentences for training |
1: initialized: Wtem, Wspa, Watt, Wclass |
2: for k=1,...,K (K=epoch) |
3: for t=1,...,T (T=514/batch size) |
4: deep-and-shallow loss: L(t)=λ1LT(t)+λ2LS(t)+(1−λ1−λ2)LA(t) |
5: backpropagation error:\(\frac {\partial L(t)}{\partial x_{i}(t)}\) |
6: \(W_{\text {tem}}(t+1){\gets }W_{\text {tem}}(t)-\mu * \frac {\partial L_{T}(t)}{\partial W_{\text {tem}}(t)}-\mu * \frac {\partial L_{A}(t)}{\partial W_{\text {tem}}(t)}\) |
7: \(W_{\text {spa}}(t+1){\gets }W_{\text {spa}}(t)-\mu * \frac {\partial L_{s}(t)}{\partial W_{\text {spa}}(t)}-\mu * \frac {\partial L_{A}(t)}{\partial W_{\text {spa}}(t)}\) |
8: \(W_{\text {att}}(t+1){\gets }W_{\text {att}}(t)-\mu * \frac {\partial L_{A}(t)}{\partial W_{\text {att}}(t)}\) |
9: \(W_{\text {cla}}(t+1){\gets }W_{\text {cla}}(t)-\mu * \frac {\partial L_{A}(t)}{\partial W_{\text {cla}}(t)}\) |
10: end for |
11: end for |
12: return Wtem, Wspa, Watt, Wclass |