Skip to main content
. 2021 May 14;10:e67855. doi: 10.7554/eLife.67855

Appendix 1—table 3. Comparison of feature sets on the downstream task of predicting mouse identity given acoustic features of single syllables.

Classification accuracy, in percent, averaged over five disjoint, class-balanced splits of the data is reported. A class-weighted log-likelihood loss is targeted to help correct for class imbalance. Empirical standard deviation is shown in parentheses. Each MUPET acoustic feature is independently z-scored as a preprocessing step. Latent feature principal components are truncated when >99% of the feature variance is explained. The multilayer perceptron (MLP) classifiers are two-layer networks with a hidden layer size of 100, ReLU activations, and an L2 weight regularization parameter ‘alpha,’ trained with ADAM optimization with a learning rate of 10-3 for 200 epochs. Chance performance is 2.8% for top-1 accuracy and 13.9% for top-5 accuracy. D denotes the dimension of each feature set, with Gaussian random projections used to decrease the dimension of spectrograms.

Predicting mouse identity (Figure 4f)
Spectrogram MUPET Latent
D = 10 D = 30 D = 100 D = 9 D = 8
Top-1 accuracy
MLP (α = 0.01) 9.9 (0.2) 14.9 (0.2) 20.4 (0.4) 14.7 (0.2) 17.0 (0.3)
MLP (α = 0.001) 10.8 (0.1) 17.3 (0.4) 25.3 (0.3) 19.0 (0.3) 22.7 (0.5)
MLP (α = 0.0001) 10.7 (0.2) 17.3 (0.3) 25.1 (0.3) 20.6 (0.4) 24.0 (0.2)
Top-5 accuracy
MLP (α = 0.01) 36.6 (0.4) 45.1 (0.5) 55.0 (0.3) 46.5 (0.3) 49.9 (0.4)
MLP (α = 0.001) 38.6 (0.2) 50.7 (0.6) 62.9 (0.4) 54.0 (0.2) 59.2 (0.6)
MLP (α = 0.0001) 38.7 (0.5) 50.8 (0.3) 63.2 (0.4) 57.3 (0.4) 61.6 (0.4)