Sparse autoencoders (SAEs) are a promising approach to interpreting the internal representations of transformer language models. However, SAEs are usually trained separately on each transformer layer, making it difficult to use them to study how information flows across layers. To solve this problem, we introduce the multi-layer SAE (MLSAE): a single SAE trained on the residual stream activation vectors from every transformer layer. Given that the residual stream is understood to preserve information across layers, we expected MLSAE latents to 'switch on' at a token position and remain active at later layers. Interestingly, we find that individual latents are often active at a single layer for a given token or prompt, but the layer at which an individual latent is active may differ for different tokens or prompts. We quantify these phenomena by defining a distribution over layers and considering its variance. We find that the variance of the distributions of latent activations over layers is about two orders of magnitude greater when aggregating over tokens compared with a single token. For larger underlying models, the degree to which latents are active at multiple layers increases, which is consistent with the fact that the residual stream activation vectors at adjacent layers become more similar. Finally, we relax the assumption that the residual stream basis is the same at every layer by applying pre-trained tuned-lens transformations, but our findings remain qualitatively similar. Our results represent a new approach to understanding how representations change as they flow through transformers. We release our code to train and analyze MLSAEs at https://github.com/tim-lawson/mlsae.

Sparse autoencoders (SAEs) have been successfully used to discover sparse and human-interpretable representations of the latent activations of language models (LLMs). However, we would ultimately like to understand the computations performed by LLMs and not just their representations. The extent to which SAEs can help us understand computations is unclear because they are not designed to “sparsify” computations in any sense, only latent activations. To solve this, we propose Jacobian sparse autoencoders (JSAEs), which yield not only sparsity in the input and output activations of a given model component but also sparsity in the computation (formally, the Jacobian) connecting them. With a naïve implementation, the Jacobians in LLMs would be computationally intractable due to their size. Our key technical contribution is thus finding an efficient way of computing Jacobians in this setup. We find that JSAEs extract a relatively large degree of computational sparsity while preserving downstream LLM performance approximately as well as traditional SAEs. We also show that JSAEs achieve a greater degree of computational sparsity on pre-trained LLMs than on the equivalent randomized LLM. This shows that the sparsity of the computational graph appears to be a property that LLMs learn through training, and suggests that JSAEs might be more suitable for understanding learned transformer computations than standard SAEs.

It is difficult to estimate the mutual information between spike trains because established methods require more data than are usually available. Kozachenko-Leonenko estimators promise to solve this problem but include a smoothing parameter that must be set. We propose here that the smoothing parameter can be selected by maximizing the estimated unbiased mutual information. This is tested on fictive data and shown to work very well.

We propose that the critical function of sleep is to prevent uncontrolled neuronal feedback while allowing rapid responses and prolonged retention of short-term memories. Through learning, the brain is tuned to react optimally to environmental challenges. Optimal behavior often requires rapid responses and the prolonged retention of short-term memories. At a neuronal level, these correspond to recurrent activity in local networks. Unfortunately, when a network exhibits recurrent activity, small changes in the parameters or conditions can lead to runaway oscillations. Thus, the very changes that improve the processing performance of the network can put it at risk of runaway oscillation. To prevent this, stimulus-dependent network changes should be permitted only when there is a margin of safety around the current network parameters. We propose that the essential role of sleep is to establish this margin by exposing the network to a variety of inputs, monitoring for erratic behavior, and adjusting the parameters. When sleep is not possible, an emergency mechanism must come into play, preventing runaway behavior at the expense of processing efficiency. This is tiredness.

The Victor-Purpura spike train metric has recently been extended to a family of multineuron metrics and used to analyze spike trains recorded simultaneously from pairs of proximate neurons. The metric is one of the two metrics commonly used for quantifying the distance between two spike trains; the other is the van Rossum metric. Here, we suggest an extension of the van Rossum metric to a multineuron metric. We believe this gives a metric that is both natural and easy to calculate. Both types of multineuron metric are applied to simulated data and are compared.