Clustering is an unsupervised technique that separates the data into groups of similar data points. This package implements seismic-signal clustering based on deep neural networks.

The clustering is performed by simultaneously learning a set of cluster centers in the latent feature space and the parameters of a deep neural network that maps data points into latent feature space. It consists two phase:

1. parameter initialization with a deep convolutional autoencoder
2. parameter optimization (i.e., clustering), where in an iterative manner, an auxiliary target distribution is computed and the Kullback-Leibler (KL) divergence to it is minimized.

As input, the model uses spectrograms computed using the short-time Fourier transform (STFT).

• Python (== 3.9, with development headers)
• NumPy (== 1.22.3)
• SciPy (== 1.8.0)
• PyTorch (== 1.11.0)
• scikit-learn (== 1.0.2)
• xarray (== 2022.3.0)
• Pyrocko (== 2021.09.14)
• Joblib (== 1.1.0)

We use a fully convolutional autoencoder (CAE) to encode the salient features of input spectrograms into a lower-dimensional latent features, called the embedded layer. The size of the embedding layer is smaller than the input data, so the autoencoder learns the most salient features of the input data. This architecture is called under-complete autoencoder and is used for dimensionality reduction.

The autoencoder is composed of convolutional and transposed convolutional layers and minimizes the mean squared error (MSE) between an input spectrogram and its reconstruction.

The effectiveness of the autoencoder's ability to reconstruct the input spectrogram is illustrated in example figure above. Although some loss of resolution in time and frequency is expected due to the convolutional and transposed convolutional layers, the structure of the spectrogram is largely preserved, with the salient information of the input encoded to the latent feature space.

After the pretraining step, where the salient features of the data are learned by the autoencoder weights, clustering algorithms can be applied to the data's latent feature space. In this step, the decoder layers are discarded and only the encoder layers are used as initial mapping between the data space and the feature space. To initialize the cluster centers, the data is passed through the initialized deep network to get embedded data points and then standard k-means clustering is performed in the feature space to initialize cluster centroids.

In this step:

1. first, soft assignment is performed, meaning that the degree of similarity (or membership probability), , is measured between each embedded point, , and cluster centroids, , are calculated using Student’s t-distribution.

2. then, the soft assignments are used to compute an auxiliary target distribution, .

3. finally, clusters are refined by learning from their high confidence assignments with the help of the auxiliary target distributionby. The objective is to minimize the Kullback-Leibler (KL) divergence between the soft assignments, , and the target distribution, .

• Xie, J., Girshick, R., & Farhadi, A. (2016). Unsupervised deep embedding for clustering analysis. In International conference on machine learning (pp. 478-487). PMLR.

• Nima Nooshiri - [email protected]

Any feedback/suggestions are welcome.