Each of these concepts have been individually explored in previous studies.

Identifying the features learned by neural networks is a core challenge in mechanistic interpretability. Sparse autoencoders (SAEs), which learn a sparse, overcomplete dictionary that reconstructs a network's internal activations, have been used to identify these features. However, SAEs may learn more about the structure of the dataset than the computational structure of the network. There is therefore only indirect reason to believe that the directions found in these dictionaries are functionally important to the network. We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing the KL divergence between the output distributions of the original model and the model with SAE activations inserted.

One of the goals of neuroscience is to identify neural networks that correlate with important behaviors, environments, or genotypes. This work proposes a strategy for identifying neural networks characterized by time-and frequency-dependent connectivity patterns, using convolutional dictionary learning that links spike-train data to local field potentials (LFPs) across multiple areas of the brain. Analytical contributions are: ( i) modeling dynamic relationships between LFPs and spikes; ( ii) describing the relationships between spikes and LFPs, by analyzing the ability to predict LFP data from one region based on spiking information from across the brain; and ( iii) development of a clustering methodology that allows inference of similarities in neurons from multiple regions. Results are based on data sets in which spike and LFP data are recorded simultaneously from up to 16 brain regions in a mouse.

Identifying the features learned by neural networks is a core challenge in mechanistic interpretability. Sparse autoencoders (SAEs), which learn a sparse, overcomplete dictionary that reconstructs a network's internal activations, have been used to identify these features. However, SAEs may learn more about the structure of the dataset than the computational structure of the network. There is therefore only indirect reason to believe that the directions found in these dictionaries are functionally important to the network. We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing the KL divergence between the output distributions of the original model and the model with SAE activations inserted.

Dictionary learning has recently emerged as a promising approach for mechanistic interpretability of large transformer models. Disentangling high-dimensional transformer embeddings, however, requires algorithms that scale to high-dimensional data with large sample sizes. Recent work has explored sparse autoencoders (SAEs) for this problem. However, SAEs use a simple linear encoder to solve the sparse encoding subproblem, which is known to be NP-hard. It is therefore interesting to understand whether this structure is sufficient to find good solutions to the dictionary learning problem or if a more sophisticated algorithm could find better solutions. In this work, we propose Double-Batch KSVD (DB-KSVD), a scalable dictionary learning algorithm that adapts the classic KSVD algorithm. DB-KSVD is informed by the rich theoretical foundations of KSVD but scales to datasets with millions of samples and thousands of dimensions. We demonstrate the efficacy of DB-KSVD by disentangling embeddings of the Gemma-2-2B model and evaluating on six metrics from the SAEBench benchmark, where we achieve competitive results when compared to established approaches based on SAEs. By matching SAE performance with an entirely different optimization approach, our results suggest that (i) SAEs do find strong solutions to the dictionary learning problem and (ii) that traditional optimization approaches can be scaled to the required problem sizes, offering a promising avenue for further research. We provide an implementation of DB-KSVD at https://github.com/RomeoV/KSVD.jl.

A new Lévy process prior is proposed for an uncountable collection of covariatedependent feature-learning measures; the model is called the kernel beta process (KBP). Available covariates are handled efficiently via the kernel construction, with covariates assumed observed with each data sample ("customer"), and latent covariates learned for each feature ("dish"). Each customer selects dishes from an infinite buffet, in a manner analogous to the beta process, with the added constraint that a customer first decides probabilistically whether to "consider" a dish, based on the distance in covariate space between the customer and dish. If a customer does consider a particular dish, that dish is then selected probabilistically as in the beta process. The beta process is recovered as a limiting case of the KBP. An efficient Gibbs sampler is developed for computations, and state-of-the-art results are presented for image processing and music analysis tasks.

One of the goals of neuroscience is to identify neural networks that correlate with important behaviors, environments, or genotypes. This work proposes a strategy for identifying neural networks characterized by time-and frequency-dependent connectivity patterns, using convolutional dictionary learning that links spike-train data to local field potentials (LFPs) across multiple areas of the brain. Analytical contributions are: (i) modeling dynamic relationships between LFPs and spikes; (ii) describing the relationships between spikes and LFPs, by analyzing the ability to predict LFP data from one region based on spiking information from across the brain; and (iii) development of a clustering methodology that allows inference of similarities in neurons from multiple regions. Results are based on data sets in which spike and LFP data are recorded simultaneously from up to 16 brain regions in a mouse.

Identifying the features learned by neural networks is a core challenge in mechanistic interpretability. Sparse autoencoders (SAEs), which learn a sparse, overcomplete dictionary that reconstructs a network's internal activations, have been used to identify these features. However, SAEs may learn more about the structure of the datatset than the computational structure of the network. There is therefore only indirect reason to believe that the directions found in these dictionaries are functionally important to the network. We propose end-to-end (e2e) sparse dictionary learning, a method for training SAEs that ensures the features learned are functionally important by minimizing the KL divergence between the output distributions of the original model and the model with SAE activations inserted. Compared to standard SAEs, e2e SAEs offer a Pareto improvement: They explain more network performance, require fewer total features, and require fewer simultaneously active features per datapoint, all with no cost to interpretability. We explore geometric and qualitative differences between e2e SAE features and standard SAE features. E2e dictionary learning brings us closer to methods that can explain network behavior concisely and accurately.

Recent work has proposed the linear representation hypothesis: that language models perform computation by manipulating one-dimensional representations of concepts ("features") in activation space. In contrast, we explore whether some language model representations may be inherently multi-dimensional. We begin by developing a rigorous definition of irreducible multi-dimensional features based on whether they can be decomposed into either independent or non-co-occurring lower-dimensional features. Motivated by these definitions, we design a scalable method that uses sparse autoencoders to automatically find multi-dimensional features in GPT-2 and Mistral 7B. These auto-discovered features include strikingly interpretable examples, e.g. circular features representing days of the week and months of the year. We identify tasks where these exact circles are used to solve computational problems involving modular arithmetic in days of the week and months of the year. Finally, we provide evidence that these circular features are indeed the fundamental unit of computation in these tasks with intervention experiments on Mistral 7B and Llama 3 8B, and we find further circular representations by breaking down the hidden states for these tasks into interpretable components.

Nonparametric Bayesian methods are developed for analysis of multi-channel spike-train data, with the feature learning and spike sorting performed jointly. The feature learning and sorting are performed simultaneously across all channels. Dictionary learning is implemented via the beta-Bernoulli process, with spike sorting performed via the dynamic hierarchical Dirichlet process (dHDP), with these two models coupled. The dHDP is augmented to eliminate refractoryperiod violations, it allows the "appearance" and "disappearance" of neurons over time, and it models smooth variation in the spike statistics.