Deep learning models often have more parameters than observations, and still perform well. This is sometimes described as a paradox. In this work, we show experimentally that despite their huge number of parameters, deep neural networks can compress the data losslessly even when taking the cost of encoding the parameters into account. Such a compression viewpoint originally motivated the use of variational methods in neural networks. However, we show that these variational methods provide surprisingly poor compression bounds, despite being explicitly built to minimize such bounds. This might explain the relatively poor practical performance of variational methods in deep learning. Better encoding methods, imported from the Minimum Description Length (MDL) toolbox, yield much better compression values on deep networks.

However, the scalability, general applicability, and interpretabil-ity of such approaches remain a challenge.

Approximate inference, that is approximating posterior distributions and likelihood functions, is at the core of modern probabilistic machine learning.

Neural Information Processing Systems http://nips.cc/

Deep learning models often have more parameters than observations, and still perform well. This is sometimes described as a paradox. In this work, we show experimentally that despite their huge number of parameters, deep neural networks can compress the data losslessly even when taking the cost of encoding the parameters into account. Such a compression viewpoint originally motivated the use of variational methods in neural networks. However, we show that these variational methods provide surprisingly poor compression bounds, despite being explicitly built to minimize such bounds. This might explain the relatively poor practical performance of variational methods in deep learning. Better encoding methods, imported from the Minimum Description Length (MDL) toolbox, yield much better compression values on deep networks.

The paper introduces a probabilistic model for networks which assigns each node in the network to multiple, overlapping latent communities. Inference is done using a stochastic variational method and the experimental evaluations are performed on very large networks. The first thing I note is that you do not cite Morup et al. (2010) "Infinite multiple membership relational modelling for complex networks", which in truth was the first work to perform inference for a latent feature relational model on large datasets -- in effect, rendering your statement on 067-068 "... these innovations allow the first..." incorrect. This is a rather serious oversight, because their paper not only performs large scale inference, but their method is also an MCMC method, which is well-known to usually produce more accurate results than variational methods. I believe the strongest contribution from this paper is the application of a stochastic variational inference method to a relational data model.

Sparse Autoencoders (SAEs) have emerged as a promising approach for interpreting neural network representations by learning sparse, human-interpretable features from dense activations. We investigate whether incorporating variational methods into SAE architectures can improve feature organization and interpretability. We introduce the Variational Sparse Autoencoder (vSAE), which replaces deterministic ReLU gating with stochastic sampling from learned Gaussian posteriors and incorporates KL divergence regularization toward a standard normal prior. Our hypothesis is that this probabilistic sampling creates dispersive pressure, causing features to organize more coherently in the latent space while avoiding overlap. We evaluate a TopK vSAE against a standard TopK SAE on Pythia-70M transformer residual stream activations using comprehensive benchmarks including SAE Bench, individual feature interpretability analysis, and global latent space visualization through t-SNE. The vSAE underperforms standard SAE across core evaluation metrics, though excels at feature independence and ablation metrics. The KL divergence term creates excessive regularization pressure that substantially reduces the fraction of living features, leading to observed performance degradation. While vSAE features demonstrate improved robustness, they exhibit many more dead features than baseline. Our findings suggest that naive application of variational methods to SAEs does not improve feature organization or interpretability.

Gaussian process (GP) models form a core part of probabilistic machine learning. Considerable research effort has been made into attacking three issues with GP models: how to compute efficiently when the number of data is large; how to approximate the posterior when the likelihood is not Gaussian and how to estimate covariance function parameter posteriors. This paper simultaneously addresses these, using a variational approximation to the posterior which is sparse in support of the function but otherwise free-form. The result is a Hybrid Monte-Carlo sampling scheme which allows for a non-Gaussian approximation over the function values and covariance parameters simultaneously, with efficient computations based on inducing-point sparse GPs. Code to replicate each experiment in this paper is available at github.com/sparseMCMC .

One-dimensional signal decomposition is a well-established and widely used technique across various scientific fields. It serves as a highly valuable pre-processing step for data analysis. While traditional decomposition techniques often rely on mathematical models, recent research suggests that applying the latest deep learning models to this problem presents an exciting, unexplored area with promising potential. This work presents a novel method for the additive decomposition of one-dimensional signals. We leverage the Transformer architecture to decompose signals into their constituent components: piece-wise constant, smooth (low-frequency oscillatory), textured (high-frequency oscillatory), and a noise component. Our model, trained on synthetic data, achieves excellent accuracy in modeling and decomposing input signals from the same distribution, as demonstrated by the experimental results.