Exploring Action-Centric Representations Through the Lens of Rate-Distortion Theory

Organisms have to keep track of the information in the environment that is relevant for adaptive behaviour. Transmitting information in an economical and efficient way becomes crucial for limited-resourced agents living in high-dimensional environments. The efficient coding hypothesis claims that organisms seek to maximize the information about the sensory input in an efficient manner. Under Bayesian inference, this means that the role of the brain is to efficiently allocate resources in order to make predictions about the hidden states that cause sensory data. However, neither of those frameworks accounts for how that information is exploited downstream, leaving aside the action-oriented role of the perceptual system. Rate-distortion theory, which defines optimal lossy compression under constraints, has gained attention as a formal framework to explore goal-oriented efficient coding. In this work, we explore action-centric representations in the context of rate-distortion theory. We also provide a mathematical definition of abstractions and we argue that, as a summary of the relevant details, they can be used to fix the content of action-centric representations. We model action-centric representations using VAEs and we find that such representations i) are efficient lossy compressions of the data; ii) capture the task-dependent invariances necessary to achieve successful behaviour; and iii) are not in service of reconstructing the data. Thus, we conclude that full reconstruction of the data is rarely needed to achieve optimal behaviour, consistent with a teleological approach to perception.

Tree Attention: Topology-aware Decoding for Long-Context Attention on GPU clusters

Self-attention is the core mathematical operation of modern transformer architectures and is also a significant computational bottleneck due to its quadratic complexity in the sequence length. In this work, we derive the scalar energy function whose gradient computes the self-attention block, thus elucidating the theoretical underpinnings of self-attention, providing a Bayesian interpretation of the operation and linking it closely with energy-based models such as Hopfield Networks. Our formulation reveals that the reduction across the sequence axis can be efficiently computed in parallel through a tree reduction. Our algorithm, for parallelizing attention computation across multiple GPUs enables cross-device decoding to be performed asymptotically faster (up to 8x faster in our experiments) than alternative approaches such as Ring Attention, while also requiring significantly less communication volume and incurring 2x less peak memory. Our code is publicly available here: \url{https://github.com/Zyphra/tree_attention}.

Zyda: A 1.3T Dataset for Open Language Modeling

The size of large language models (LLMs) has scaled dramatically in recent years and their computational and data requirements have surged correspondingly. State-of-the-art language models, even at relatively smaller sizes, typically require training on at least a trillion tokens. This rapid advancement has eclipsed the growth of open-source datasets available for large-scale LLM pretraining. In this paper, we introduce Zyda (Zyphra Dataset), a dataset under a permissive license comprising 1.3 trillion tokens, assembled by integrating several major respected open-source datasets into a single, high-quality corpus. We apply rigorous filtering and deduplication processes, both within and across datasets, to maintain and enhance the quality derived from the original datasets. Our evaluations show that Zyda not only competes favorably with other open datasets like Dolma, FineWeb, and RefinedWeb, but also substantially improves the performance of comparable models from the Pythia suite. Our rigorous data processing methods significantly enhance Zyda's effectiveness, outperforming even the best of its constituent datasets when used independently.

Toward Conversational Agents with Context and Time Sensitive Long-term Memory

There has recently been growing interest in conversational agents with long-term memory which has led to the rapid development of language models that use retrieval-augmented generation (RAG). Until recently, most work on RAG has focused on information retrieval from large databases of texts, like Wikipedia, rather than information from long-form conversations. In this paper, we argue that effective retrieval from long-form conversational data faces two unique problems compared to static database retrieval: 1) time/event-based queries, which requires the model to retrieve information about previous conversations based on time or the order of a conversational event (e.g., the third conversation on Tuesday), and 2) ambiguous queries that require surrounding conversational context to understand. To better develop RAG-based agents that can deal with these challenges, we generate a new dataset of ambiguous and time-based questions that build upon a recent dataset of long-form, simulated conversations, and demonstrate that standard RAG based approaches handle such questions poorly. We then develop a novel retrieval model which combines chained-of-table search methods, standard vector-database retrieval, and a prompting method to disambiguate queries, and demonstrate that this approach substantially improves over current methods at solving these tasks. We believe that this new dataset and more advanced RAG agent can act as a key benchmark and stepping stone towards effective memory augmented conversational agents that can be used in a wide variety of AI applications.

Zamba: A Compact 7B SSM Hybrid Model

In this technical report, we present Zamba, a novel 7B SSM-transformer hybrid model which achieves competitive performance against leading open-weight models at a comparable scale. Zamba is trained on 1T tokens from openly available datasets and is the best non-transformer model at this scale. Zamba pioneers a unique architecture combining a Mamba backbone with a single shared attention module, thus obtaining the benefits of attention at minimal parameter cost. Due to its architecture, Zamba is significantly faster at inference than comparable transformer models and requires substantially less memory for generation of long sequences. Zamba is pretrained in two phases: the first phase is based on existing web datasets, while the second one consists of annealing the model over high-quality instruct and synthetic datasets, and is characterized by a rapid learning rate decay. We open-source the weights and all checkpoints for Zamba, through both phase 1 and annealing phases.

Associative Memories in the Feature Space

An autoassociative memory model is a function that, given a set of data points, takes as input an arbitrary vector and outputs the most similar data point from the memorized set. However, popular memory models fail to retrieve images even when the corruption is mild and easy to detect for a human evaluator. This is because similarities are evaluated in the raw pixel space, which does not contain any semantic information about the images. This problem can be easily solved by computing \emph{similarities} in an embedding space instead of the pixel space. We show that an effective way of computing such embeddings is via a network pretrained with a contrastive loss. As the dimension of embedding spaces is often significantly smaller than the pixel space, we also have a faster computation of similarity scores. We test this method on complex datasets such as CIFAR10 and STL10. An additional drawback of current models is the need of storing the whole dataset in the pixel space, which is often extremely large. We relax this condition and propose a class of memory models that only stores low-dimensional semantic embeddings, and uses them to retrieve similar, but not identical, memories. We demonstrate a proof of concept of this method on a simple task on the MNIST dataset.

BlackMamba: Mixture of Experts for State-Space Models

State-space models (SSMs) have recently demonstrated competitive performance to transformers at large-scale language modeling benchmarks while achieving linear time and memory complexity as a function of sequence length. Mamba, a recently released SSM model, shows impressive performance in both language modeling and long sequence processing tasks. Simultaneously, mixture-of-expert (MoE) models have shown remarkable performance while significantly reducing the compute and latency costs of inference at the expense of a larger memory footprint. In this paper, we present BlackMamba, a novel architecture that combines the Mamba SSM with MoE to obtain the benefits of both. We demonstrate that BlackMamba performs competitively against both Mamba and transformer baselines, and outperforms in inference and training FLOPs. We fully train and open-source 340M/1.5B and 630M/2.8B BlackMamba models on 300B tokens of a custom dataset. We show that BlackMamba inherits and combines both of the benefits of SSM and MoE architectures, combining linear-complexity generation from SSM with cheap and fast inference from MoE. We release all weights, checkpoints, and inference code open-source. Inference code at: https://github.com/Zyphra/BlackMamba

Causal Inference via Predictive Coding

Bayesian and causal inference are fundamental processes for intelligence. Bayesian inference models observations: what can be inferred about y if we observe a related variable x? Causal inference models interventions: if we directly change x, how will y change? Predictive coding is a neuroscience-inspired method for performing Bayesian inference on continuous state variables using local information only. In this work, we go beyond Bayesian inference, and show how a simple change in the inference process of predictive coding enables interventional and counterfactual inference in scenarios where the causal graph is known. We then extend our results, and show how predictive coding can be generalized to cases where this graph is unknown, and has to be inferred from data, hence performing causal discovery. What results is a novel and straightforward technique that allows us to perform end-to-end causal inference on predictive-coding-based structural causal models, and demonstrate its utility for potential applications in machine learning.

Designing Ecosystems of Intelligence from First Principles

This white paper lays out a vision of research and development in the field of artificial intelligence for the next decade (and beyond). Its denouement is a cyber-physical ecosystem of natural and synthetic sense-making, in which humans are integral participants$\unicode{x2014}$what we call ''shared intelligence''. This vision is premised on active inference, a formulation of adaptive behavior that can be read as a physics of intelligence, and which inherits from the physics of self-organization. In this context, we understand intelligence as the capacity to accumulate evidence for a generative model of one's sensed world$\unicode{x2014}$also known as self-evidencing. Formally, this corresponds to maximizing (Bayesian) model evidence, via belief updating over several scales: i.e., inference, learning, and model selection. Operationally, this self-evidencing can be realized via (variational) message passing or belief propagation on a factor graph. Crucially, active inference foregrounds an existential imperative of intelligent systems; namely, curiosity or the resolution of uncertainty. This same imperative underwrites belief sharing in ensembles of agents, in which certain aspects (i.e., factors) of each agent's generative world model provide a common ground or frame of reference. Active inference plays a foundational role in this ecology of belief sharing$\unicode{x2014}$leading to a formal account of collective intelligence that rests on shared narratives and goals. We also consider the kinds of communication protocols that must be developed to enable such an ecosystem of intelligences and motivate the development of a shared hyper-spatial modeling language and transaction protocol, as a first$\unicode{x2014}$and key$\unicode{x2014}$step towards such an ecology.

Interpreting Neural Networks through the Polytope Lens

Mechanistic interpretability aims to explain what a neural network has learned at a nuts-and-bolts level. What are the fundamental primitives of neural network representations? Previous mechanistic descriptions have used individual neurons or their linear combinations to understand the representations a network has learned. But there are clues that neurons and their linear combinations are not the correct fundamental units of description: directions cannot describe how neural networks use nonlinearities to structure their representations. Moreover, many instances of individual neurons and their combinations are polysemantic (i.e. they have multiple unrelated meanings). Polysemanticity makes interpreting the network in terms of neurons or directions challenging since we can no longer assign a specific feature to a neural unit. In order to find a basic unit of description that does not suffer from these problems, we zoom in beyond just directions to study the way that piecewise linear activation functions (such as ReLU) partition the activation space into numerous discrete polytopes. We call this perspective the polytope lens. The polytope lens makes concrete predictions about the behavior of neural networks, which we evaluate through experiments on both convolutional image classifiers and language models. Specifically, we show that polytopes can be used to identify monosemantic regions of activation space (while directions are not in general monosemantic) and that the density of polytope boundaries reflect semantic boundaries. We also outline a vision for what mechanistic interpretability might look like through the polytope lens.