Viktoria Schuster

Viktoria Schuster

Sparse autoencoders (SAEs) have lately been used to uncover interpretable latent features in large language models. By projecting dense embeddings into a much higher-dimensional and sparse space, learned features become disentangled and easier to interpret. This work explores the potential of SAEs for decomposing embeddings in complex and high-dimensional biological data. Using simulated data, it outlines the efficacy, hyperparameter landscape, and limitations of SAEs when it comes to extracting ground truth generative variables from latent space. The application to embeddings from pretrained single-cell models shows that SAEs can find and steer key biological processes and even uncover subtle biological signals that might otherwise be missed. This work further introduces scFeatureLens, an automated interpretability approach for linking SAE features and biological concepts from gene sets to enable large-scale analysis and hypothesis generation in single-cell gene expression models.

Viktoria Schuster, Anders Krogh

Learning low-dimensional representations of single-cell transcriptomics has become instrumental to its downstream analysis. The state of the art is currently represented by neural network models such as variational autoencoders (VAEs) which use a variational approximation of the likelihood for inference. We here present the Deep Generative Decoder (DGD), a simple generative model that computes model parameters and representations directly via maximum a posteriori (MAP) estimation. The DGD handles complex parameterized latent distributions naturally unlike VAEs which typically use a fixed Gaussian distribution, because of the complexity of adding other types. We first show its general functionality on a commonly used benchmark set, Fashion-MNIST. Secondly, we apply the model to multiple single-cell data sets. Here the DGD learns low-dimensional, meaningful and well-structured latent representations with sub-clustering beyond the provided labels. The advantages of this approach are its simplicity and its capability to provide representations of much smaller dimensionality than a comparable VAE.

Viktoria Schuster, Anders Krogh

Autoencoders are commonly used in representation learning. They consist of an encoder and a decoder, which provide a straightforward way to map n-dimensional data in input space to a lower m-dimensional representation space and back. The decoder itself defines an m-dimensional manifold in input space. Inspired by manifold learning, we show that the decoder can be trained on its own by learning the representations of the training samples along with the decoder weights using gradient descent. A sum-of-squares loss then corresponds to optimizing the manifold to have the smallest Euclidean distance to the training samples, and similarly for other loss functions. We derive expressions for the number of samples needed to specify the encoder and decoder and show that the decoder generally requires much less training samples to be well-specified compared to the encoder. We discuss training of autoencoders in this perspective and relate to previous work in the field that use noisy training examples and other types of regularization. On the natural image data sets MNIST and CIFAR10, we demonstrate that the decoder is much better suited to learn a low-dimensional representation, especially when trained on small data sets. Using simulated gene regulatory data, we further show that the decoder alone leads to better generalization and meaningful representations. Our approach of training the decoder alone facilitates representation learning even on small data sets and can lead to improved training of autoencoders. We hope that the simple analyses presented will also contribute to an improved conceptual understanding of representation learning.