Fig. 2

The atyPeak model learns correlation groups. In each case, the tensor at top is the original representation and the bottom one is what is rebuilt by the model. The model was trained on artificial data. There were 2 predefined correlation groups covering different subsets of dimensions (G1 and G2) defined in Additional file 2: Fig. S2. The thin colored lines are only here as a visual aid. In a when the model rebuilds the CRM representation, it rebuilds the entire correlation group when peaks from the group are present. This results in adding the other members that were not originally present as “phantom” peaks. In this case, it is the G1 group. In b however, we used a model with a less aggressive compression (too high information budget) and the rebuilding is too precise, learning smaller, non-significant groups instead of the entire G1 or G2 groups. Model parameters in a were a deep dimension of 32, 16 filters and a Learning Rate (LR) of 1E−3. b Used 48 filters, 256 deep dimension, a LR of 1E−4. Note that for B, that increased precision is not achieved with higher deep dim but default LR—we needed a lower LR. 48 epochs for all or early stopping (for a). Groups were equiprobable