Marcus Hütter

Jordi Grau-Moya, Grégoire Delétang, Markus Kunesch et al. · stanford

Meta-training agents with memory has been shown to culminate in Bayes-optimal agents, which casts Bayes-optimality as the implicit solution to a numerical optimization problem rather than an explicit modeling assumption. Bayes-optimal agents are risk-neutral, since they solely attune to the expected return, and ambiguity-neutral, since they act in new situations as if the uncertainty were known. This is in contrast to risk-sensitive agents, which additionally exploit the higher-order moments of the return, and ambiguity-sensitive agents, which act differently when recognizing situations in which they lack knowledge. Humans are also known to be averse to ambiguity and sensitive to risk in ways that aren't Bayes-optimal, indicating that such sensitivity can confer advantages, especially in safety-critical situations. How can we extend the meta-learning protocol to generate risk- and ambiguity-sensitive agents? The goal of this work is to fill this gap in the literature by showing that risk- and ambiguity-sensitivity also emerge as the result of an optimization problem using modified meta-training algorithms, which manipulate the experience-generation process of the learner. We empirically test our proposed meta-training algorithms on agents exposed to foundational classes of decision-making experiments and demonstrate that they become sensitive to risk and ambiguity.

Grégoire Delétang, Anian Ruoss, Paul-Ambroise Duquenne et al.

It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are powerful general-purpose predictors and that the compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively. Finally, we show that the prediction-compression equivalence allows us to use any compressor (like gzip) to build a conditional generative model.

Grégoire Delétang, Anian Ruoss, Jordi Grau-Moya et al.

Reliable generalization lies at the heart of safe ML and AI. However, understanding when and how neural networks generalize remains one of the most important unsolved problems in the field. In this work, we conduct an extensive empirical study (20'910 models, 15 tasks) to investigate whether insights from the theory of computation can predict the limits of neural network generalization in practice. We demonstrate that grouping tasks according to the Chomsky hierarchy allows us to forecast whether certain architectures will be able to generalize to out-of-distribution inputs. This includes negative results where even extensive amounts of data and training time never lead to any non-trivial generalization, despite models having sufficient capacity to fit the training data perfectly. Our results show that, for our subset of tasks, RNNs and Transformers fail to generalize on non-regular tasks, LSTMs can solve regular and counter-language tasks, and only networks augmented with structured memory (such as a stack or memory tape) can successfully generalize on context-free and context-sensitive tasks.

Mary Phuong, Marcus Hutter

This document aims to be a self-contained, mathematically precise overview of transformer architectures and algorithms (*not* results). It covers what transformers are, how they are trained, what they are used for, their key architectural components, and a preview of the most prominent models. The reader is assumed to be familiar with basic ML terminology and simpler neural network architectures such as MLPs.

Matthew Aitchison, Penny Sweetser, Marcus Hutter

The Arcade Learning Environment (ALE) has become an essential benchmark for assessing the performance of reinforcement learning algorithms. However, the computational cost of generating results on the entire 57-game dataset limits ALE's use and makes the reproducibility of many results infeasible. We propose a novel solution to this problem in the form of a principled methodology for selecting small but representative subsets of environments within a benchmark suite. We applied our method to identify a subset of five ALE games, called Atari-5, which produces 57-game median score estimates within 10% of their true values. Extending the subset to 10-games recovers 80% of the variance for log-scores for all games within the 57-game set. We show this level of compression is possible due to a high degree of correlation between many of the games in ALE.

Tor Lattimore, Marcus Hutter

A possibly immortal agent tries to maximise its summed discounted rewards over time, where discounting is used to avoid infinite utilities and encourage the agent to value current rewards more than future ones. Some commonly used discount functions lead to time-inconsistent behavior where the agent changes its plan over time. These inconsistencies can lead to very poor behavior. We generalise the usual discounted utility model to one where the discount function changes with the age of the agent. We then give a simple characterisation of time-(in)consistent discount functions and show the existence of a rational policy for an agent that knows its discount function is time-inconsistent.

Tim Genewein, Grégoire Delétang, Anian Ruoss et al.

Memory-based meta-learning is a technique for approximating Bayes-optimal predictors. Under fairly general conditions, minimizing sequential prediction error, measured by the log loss, leads to implicit meta-learning. The goal of this work is to investigate how far this interpretation can be realized by current sequence prediction models and training regimes. The focus is on piecewise stationary sources with unobserved switching-points, which arguably capture an important characteristic of natural language and action-observation sequences in partially observable environments. We show that various types of memory-based neural models, including Transformers, LSTMs, and RNNs can learn to accurately approximate known Bayes-optimal algorithms and behave as if performing Bayesian inference over the latent switching-points and the latent parameters governing the data distribution within each segment.

Jorg Bornschein, Yazhe Li, Marcus Hutter

Minimum Description Length (MDL) provides a framework and an objective for principled model evaluation. It formalizes Occam's Razor and can be applied to data from non-stationary sources. In the prequential formulation of MDL, the objective is to minimize the cumulative next-step log-loss when sequentially going through the data and using previous observations for parameter estimation. It thus closely resembles a continual- or online-learning problem. In this study, we evaluate approaches for computing prequential description lengths for image classification datasets with neural networks. Considering the computational cost, we find that online-learning with rehearsal has favorable performance compared to the previously widely used block-wise estimation. We propose forward-calibration to better align the models predictions with the empirical observations and introduce replay-streams, a minibatch incremental training technique to efficiently implement approximate random replay while avoiding large in-memory replay buffers. As a result, we present description lengths for a suite of image classification datasets that improve upon previously reported results by large margins.

Tomer Galanti, András György, Marcus Hutter

We study the ability of foundation models to learn representations for classification that are transferable to new, unseen classes. Recent results in the literature show that representations learned by a single classifier over many classes are competitive on few-shot learning problems with representations learned by special-purpose algorithms designed for such problems. We offer a theoretical explanation for this behavior based on the recently discovered phenomenon of class-feature-variability collapse, that is, that during the training of deep classification networks the feature embeddings of samples belonging to the same class tend to concentrate around their class means. More specifically, we show that the few-shot error of the learned feature map on new classes (defined as the classification error of the nearest class-center classifier using centers learned from a small number of random samples from each new class) is small in case of class-feature-variability collapse, under the assumption that the classes are selected independently from a fixed distribution. This suggests that foundation models can provide feature maps that are transferable to new downstream tasks, even with very few samples; to our knowledge, this is the first performance bound for transfer-learning that is non-vacuous in the few-shot setting.

Marcus Hutter, Steven Hansen

What is the action sequence aa'a" that was likely responsible for reaching state s"' (from state s) in 3 steps? Addressing such questions is important in causal reasoning and in reinforcement learning. Inverse "MDP" models p(aa'a"|ss"') can be used to answer them. In the traditional "forward" view, transition "matrix" p(s'|sa) and policy π(a|s) uniquely determine "everything": the whole dynamics p(as'a's"a"...|s), and with it, the action-conditional state process p(s's"...|saa'a"), the multi-step inverse models p(aa'a"...|ss^i), etc. If the latter is our primary concern, a natural question, analogous to the forward case is to which extent 1-step inverse model p(a|ss') plus policy π(a|s) determine the multi-step inverse models or even the whole dynamics. In other words, can forward models be inferred from inverse models or even be side-stepped. This work addresses this question and variations thereof, and also whether there are efficient decision/inference algorithms for this.

Boumediene Hamzi, Marcus Hutter, Houman Owhadi

Machine Learning (ML) and Algorithmic Information Theory (AIT) look at Complexity from different points of view. We explore the interface between AIT and Kernel Methods (that are prevalent in ML) by adopting an AIT perspective on the problem of learning kernels from data, in kernel ridge regression, through the method of Sparse Kernel Flows. In particular, by looking at the differences and commonalities between Minimal Description Length (MDL) and Regularization in Machine Learning (RML), we prove that the method of Sparse Kernel Flows is the natural approach to adopt to learn kernels from data. This approach aligns naturally with the MDL principle, offering a more robust theoretical basis than the existing reliance on cross-validation. The study reveals that deriving Sparse Kernel Flows does not require a statistical approach; instead, one can directly engage with code-lengths and complexities, concepts central to AIT. Thereby, this approach opens the door to reformulating algorithms in machine learning using tools from AIT, with the aim of providing them a more solid theoretical foundation.

Laurent Orseau, Marcus Hutter

Golden-section search and bisection search are the two main principled algorithms for 1d minimization of quasiconvex (unimodal) functions. The first one only uses function queries, while the second one also uses gradient queries. Other algorithms exist under much stronger assumptions, such as Newton's method. However, to the best of our knowledge, there is no principled exact line search algorithm for general convex functions -- including piecewise-linear and max-compositions of convex functions -- that takes advantage of convexity. We propose two such algorithms: $Δ$-Bisection is a variant of bisection search that uses (sub)gradient information and convexity to speed up convergence, while $Δ$-Secant is a variant of golden-section search and uses only function queries. While bisection search reduces the $x$ interval by a factor 2 at every iteration, $Δ$-Bisection reduces the (sometimes much) smaller $x^*$-gap $Δ^x$ (the $x$ coordinates of $Δ$) by at least a factor 2 at every iteration. Similarly, $Δ$-Secant also reduces the $x^*$-gap by at least a factor 2 every second function query. Moreover, the $y^*$-gap $Δ^y$ (the $y$ coordinates of $Δ$) also provides a refined stopping criterion, which can also be used with other algorithms. Experiments on a few convex functions confirm that our algorithms are always faster than their quasiconvex counterparts, often by more than a factor 2. We further design a quasi-exact line search algorithm based on $Δ$-Secant. It can be used with gradient descent as a replacement for backtracking line search, for which some parameters can be finicky to tune -- and we provide examples to this effect, on strongly-convex and smooth functions. We provide convergence guarantees, and confirm the efficiency of quasi-exact line search on a few single- and multivariate convex functions.

Bryn Elesedy, Marcus Hutter

U-Clip is a simple amendment to gradient clipping that can be applied to any iterative gradient optimization algorithm. Like regular clipping, U-Clip involves using gradients that are clipped to a prescribed size (e.g. with component wise or norm based clipping) but instead of discarding the clipped portion of the gradient, U-Clip maintains a buffer of these values that is added to the gradients on the next iteration (before clipping). We show that the cumulative bias of the U-Clip updates is bounded by a constant. This implies that the clipped updates are unbiased on average. Convergence follows via a lemma that guarantees convergence with updates $u_i$ as long as $\sum_{i=1}^t (u_i - g_i) = o(t)$ where $g_i$ are the gradients. Extensive experimental exploration is performed on CIFAR10 with further validation given on ImageNet.

Marcus Hutter

Given well-shuffled data, can we determine whether the data items are statistically (in)dependent? Formally, we consider the problem of testing whether a set of exchangeable random variables are independent. We will show that this is possible and develop tests that can confidently reject the null hypothesis that data is independent and identically distributed and have high power for (some) exchangeable distributions. We will make no structural assumptions on the underlying sample space. One potential application is in Deep Learning, where data is often scraped from the whole internet, with duplications abound, which can render data non-iid and test-set evaluation prone to give wrong answers.

Jonathon Schwartz, Hanna Kurniawati, Marcus Hutter

The design of autonomous agents that can interact effectively with other agents without prior coordination is a core problem in multi-agent systems. Type-based reasoning methods achieve this by maintaining a belief over a set of potential behaviours for the other agents. However, current methods are limited in that they assume full observability of the state and actions of the other agent or do not scale efficiently to larger problems with longer planning horizons. Addressing these limitations, we propose Partially Observable Type-based Meta Monte-Carlo Planning (POTMMCP) - an online Monte-Carlo Tree Search based planning method for type-based reasoning in large partially observable environments. POTMMCP incorporates a novel meta-policy for guiding search and evaluating beliefs, allowing it to search more effectively to longer horizons using less planning time. We show that our method converges to the optimal solution in the limit and empirically demonstrate that it effectively adapts online to diverse sets of other agents across a range of environments. Comparisons with the state-of-the art method on problems with up to $10^{14}$ states and $10^8$ observations indicate that POTMMCP is able to compute better solutions significantly faster.

Yazhe Li, Jorg Bornschein, Marcus Hutter

Although much of the success of Deep Learning builds on learning good representations, a rigorous method to evaluate their quality is lacking. In this paper, we treat the evaluation of representations as a model selection problem and propose to use the Minimum Description Length (MDL) principle to devise an evaluation metric. Contrary to the established practice of limiting the capacity of the readout model, we design a hybrid discrete and continuous-valued model space for the readout models and employ a switching strategy to combine their predictions. The MDL score takes model complexity, as well as data efficiency into account. As a result, the most appropriate model for the specific task and representation will be chosen, making it a unified measure for comparison. The proposed metric can be efficiently computed with an online method and we present results for pre-trained vision encoders of various architectures (ResNet and ViT) and objective functions (supervised and self-supervised) on a range of downstream tasks. We compare our methods with accuracy-based approaches and show that the latter are inconsistent when multiple readout models are used. Finally, we discuss important properties revealed by our evaluations such as model scaling, preferred readout model, and data efficiency.

Samuel Allen Alexander, David Quarel, Len Du et al.

Inspired by recent progress in multi-agent Reinforcement Learning (RL), in this work we examine the collective intelligent behaviour of theoretical universal agents by introducing a weighted mixture operation. Given a weighted set of agents, their weighted mixture is a new agent whose expected total reward in any environment is the corresponding weighted average of the original agents' expected total rewards in that environment. Thus, if RL agent intelligence is quantified in terms of performance across environments, the weighted mixture's intelligence is the weighted average of the original agents' intelligences. This operation enables various interesting new theorems that shed light on the geometry of RL agent intelligence, namely: results about symmetries, convex agent-sets, and local extrema. We also show that any RL agent intelligence measure based on average performance across environments, subject to certain weak technical conditions, is identical (up to a constant factor) to performance within a single environment dependent on said intelligence measure.

John Aslanides, Jan Leike, Marcus Hutter

Many state-of-the-art reinforcement learning (RL) algorithms typically assume that the environment is an ergodic Markov Decision Process (MDP). In contrast, the field of universal reinforcement learning (URL) is concerned with algorithms that make as few assumptions as possible about the environment. The universal Bayesian agent AIXI and a family of related URL algorithms have been developed in this setting. While numerous theoretical optimality results have been proven for these agents, there has been no empirical investigation of their behavior to date. We present a short and accessible survey of these URL algorithms under a unified notation and framework, along with results of some experiments that qualitatively illustrate some properties of the resulting policies, and their relative performance on partially-observable gridworld environments. We also present an open-source reference implementation of the algorithms which we hope will facilitate further understanding of, and experimentation with, these ideas.

Mary Phuong, Matthew Aitchison, Elliot Catt et al. · deepmind

To understand the risks posed by a new AI system, we must understand what it can and cannot do. Building on prior work, we introduce a programme of new "dangerous capability" evaluations and pilot them on Gemini 1.0 models. Our evaluations cover four areas: (1) persuasion and deception; (2) cyber-security; (3) self-proliferation; and (4) self-reasoning. We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs. Our goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models.

Cole Wyeth, Marcus Hutter

We generalize the AIXI reinforcement learning agent to admit a wider class of utility functions. Assigning a utility to each possible interaction history forces us to confront the ambiguity that some hypotheses in the agent's belief distribution only predict a finite prefix of the history, which is sometimes interpreted as implying a chance of death equal to a quantity called the semimeasure loss. This death interpretation suggests one way to assign utilities to such history prefixes. We argue that it is as natural to view the belief distributions as imprecise probability distributions, with the semimeasure loss as total ignorance. This motivates us to consider the consequences of computing expected utilities with Choquet integrals from imprecise probability theory, including an investigation of their computability level. We recover the standard recursive value function as a special case. However, our most general expected utilities under the death interpretation cannot be characterized as such Choquet integrals.

Amal Rannen-Triki, Jorg Bornschein, Razvan Pascanu et al. · deepmind

We consider the problem of online fine tuning the parameters of a language model at test time, also known as dynamic evaluation. While it is generally known that this approach improves the overall predictive performance, especially when considering distributional shift between training and evaluation data, we here emphasize the perspective that online adaptation turns parameters into temporally changing states and provides a form of context-length extension with memory in weights, more in line with the concept of memory in neuroscience. We pay particular attention to the speed of adaptation (in terms of sample efficiency),sensitivity to the overall distributional drift, and the computational overhead for performing gradient computations and parameter updates. Our empirical study provides insights on when online adaptation is particularly interesting. We highlight that with online adaptation the conceptual distinction between in-context learning and fine tuning blurs: both are methods to condition the model on previously observed tokens.

Li Kevin Wenliang, Grégoire Delétang, Matthew Aitchison et al.

We propose a novel algorithmic framework for distributional reinforcement learning, based on learning finite-dimensional mean embeddings of return distributions. We derive several new algorithms for dynamic programming and temporal-difference learning based on this framework, provide asymptotic convergence theory, and examine the empirical performance of the algorithms on a suite of tabular tasks. Further, we show that this approach can be straightforwardly combined with deep reinforcement learning, and obtain a new deep RL agent that improves over baseline distributional approaches on the Arcade Learning Environment.

Li Kevin Wenliang, Anian Ruoss, Jordi Grau-Moya et al.

Large language models (LLMs) can be prompted to do many tasks, but finding good prompts is not always easy, nor is understanding some performant prompts. We explore these issues by viewing prompting as conditioning a near-optimal sequence predictor (LLM) pretrained on diverse data sources. Through numerous prompt search experiments, we show that the unintuitive patterns in optimal prompts can be better understood given the pretraining distribution, which is often unavailable in practice. Moreover, even using exhaustive search, reliably identifying optimal prompts from practical neural predictors can be difficult. Further, we demonstrate that common prompting methods, such as using intuitive prompts or samples from the targeted task, are in fact suboptimal. Thus, this work takes an initial step towards understanding the difficulties in finding and understanding optimal prompts from a statistical and empirical perspective.

Alexander Meulemans, Rajai Nasser, Maciej Wołczyk et al.

The standard theory of model-free reinforcement learning assumes that the environment dynamics are stationary and that agents are decoupled from their environment, such that policies are treated as being separate from the world they inhabit. This leads to theoretical challenges in the multi-agent setting where the non-stationarity induced by the learning of other agents demands prospective learning based on prediction models. To accurately model other agents, an agent must account for the fact that those other agents are, in turn, forming beliefs about it to predict its future behavior, motivating agents to model themselves as part of the environment. Here, building upon foundational work on universal artificial intelligence (AIXI), we introduce a mathematical framework for prospective learning and embedded agency centered on self-prediction, where Bayesian RL agents predict both future perceptual inputs and their own actions, and must therefore resolve epistemic uncertainty about themselves as part of the universe they inhabit. We show that in multi-agent settings, self-prediction enables agents to reason about others running similar algorithms, leading to new game-theoretic solution concepts and novel forms of cooperation unattainable by classical decoupled agents. Moreover, we extend the theory of AIXI, and study universally intelligent embedded agents which start from a Solomonoff prior. We show that these idealized agents can form consistent mutual predictions and achieve infinite-order theory of mind, potentially setting a gold standard for embedded multi-agent learning.

Joel Veness, Marcus Hutter, Andras Gyorgy et al.

This paper considers a generalisation of universal source coding for interaction data, namely data streams that have actions interleaved with observations. Our goal will be to construct a coding distribution that is both universal \emph{and} can be used as a control policy. Allowing for action generation needs careful treatment, as naive approaches which do not distinguish between actions and observations run into the self-delusion problem in universal settings. We showcase our perspective in the context of the challenging non-stationary stochastic Bernoulli bandit problem. Our main contribution is an efficient and high performing algorithm for this problem that generalises the Partition Tree Weighting universal source coding technique for passive prediction to the control setting.

Laurent Orseau, Marcus Hutter, Levi H. S. Lelis

Levin Tree Search (LTS) (Orseau et al., 2018) is a search algorithm for deterministic environments that uses a user-specified policy to guide the search. It comes with a formal guarantee on the number of search steps (node visits) for finding a solution node that depends on the quality of the policy. In this paper, we introduce a new algorithm, called $\sqrt{\text{LTS}}$ (pronounce root-LTS), which implicitly starts an LTS search rooted at every node of the search tree. Each LTS search is assigned a rerooting weight by a (user-defined or learnt) rerooter, and the search effort is shared between all LTS searches proportionally to their weights. The rerooting mechanism implicitly decomposes the search space into subtasks, leading to significant speedups. We prove that the number of node visits that $\sqrt{\text{LTS}}$ takes is competitive with the best decomposition into subtasks, at the price of a factor that relates to the uncertainty of the rerooter. If LTS takes time $T$, in the best case with $q$ rerooting points, $\sqrt{\text{LTS}}$ only takes time $O(q\sqrt[q]{T})$. Like the policy, the rerooter can be learnt from data, and we expect $\sqrt{\text{LTS}}$ to be applicable to a wide range of domains.

Jordi Grau-Moya, Tim Genewein, Marcus Hutter et al.

Meta-learning has emerged as a powerful approach to train neural networks to learn new tasks quickly from limited data. Broad exposure to different tasks leads to versatile representations enabling general problem solving. But, what are the limits of meta-learning? In this work, we explore the potential of amortizing the most powerful universal predictor, namely Solomonoff Induction (SI), into neural networks via leveraging meta-learning to its limits. We use Universal Turing Machines (UTMs) to generate training data used to expose networks to a broad range of patterns. We provide theoretical analysis of the UTM data generation processes and meta-training protocols. We conduct comprehensive experiments with neural architectures (e.g. LSTMs, Transformers) and algorithmic data generators of varying complexity and universality. Our results suggest that UTM data is a valuable resource for meta-learning, and that it can be used to train neural networks capable of learning universal prediction strategies.

Samuel Yang-Zhao, Kee Siong Ng, Marcus Hutter

Prior approximations of AIXI, a Bayesian optimality notion for general reinforcement learning, can only approximate AIXI's Bayesian environment model using an a-priori defined set of models. This is a fundamental source of epistemic uncertainty for the agent in settings where the existence of systematic bias in the predefined model class cannot be resolved by simply collecting more data from the environment. We address this issue in the context of Human-AI teaming by considering a setup where additional knowledge for the agent in the form of new candidate models arrives from a human operator in an online fashion. We introduce a new agent called DynamicHedgeAIXI that maintains an exact Bayesian mixture over dynamically changing sets of models via a time-adaptive prior constructed from a variant of the Hedge algorithm. The DynamicHedgeAIXI agent is the richest direct approximation of AIXI known to date and comes with good performance guarantees. Experimental results on epidemic control on contact networks validates the agent's practical utility.

Laurent Orseau, Marcus Hutter, Levi H. S. Lelis

Levin Tree Search (LTS) is a search algorithm that makes use of a policy (a probability distribution over actions) and comes with a theoretical guarantee on the number of expansions before reaching a goal node, depending on the quality of the policy. This guarantee can be used as a loss function, which we call the LTS loss, to optimize neural networks representing the policy (LTS+NN). In this work we show that the neural network can be substituted with parameterized context models originating from the online compression literature (LTS+CM). We show that the LTS loss is convex under this new model, which allows for using standard convex optimization tools, and obtain convergence guarantees to the optimal parameters in an online setting for a given set of solution trajectories -- guarantees that cannot be provided for neural networks. The new LTS+CM algorithm compares favorably against LTS+NN on several benchmarks: Sokoban (Boxoban), The Witness, and the 24-Sliding Tile puzzle (STP). The difference is particularly large on STP, where LTS+NN fails to solve most of the test instances while LTS+CM solves each test instance in a fraction of a second. Furthermore, we show that LTS+CM is able to learn a policy that solves the Rubik's cube in only a few hundred expansions, which considerably improves upon previous machine learning techniques.

Tomer Galanti, András György, Marcus Hutter

We study the ability of foundation models to learn representations for classification that are transferable to new, unseen classes. Recent results in the literature show that representations learned by a single classifier over many classes are competitive on few-shot learning problems with representations learned by special-purpose algorithms designed for such problems. In this paper we provide an explanation for this behavior based on the recently observed phenomenon that the features learned by overparameterized classification networks show an interesting clustering property, called neural collapse. We demonstrate both theoretically and empirically that neural collapse generalizes to new samples from the training classes, and -- more importantly -- to new classes as well, allowing foundation models to provide feature maps that work well in transfer learning and, specifically, in the few-shot setting.

Laurent Orseau, Marcus Hutter

We extend and combine several tools of the literature to design fast, adaptive, anytime and scale-free online learning algorithms. Scale-free regret bounds must scale linearly with the maximum loss, both toward large losses and toward very small losses. Adaptive regret bounds demonstrate that an algorithm can take advantage of easy data and potentially have constant regret. We seek to develop fast algorithms that depend on as few parameters as possible, in particular they should be anytime and thus not depend on the time horizon. Our first and main tool, isotuning, is a generalization of the idea of designing adaptive learning rates that balance the trade-off of the regret. We provide a simple and versatile theorem that can be applied to a wide range of settings, and competes with the best balancing in hindsight within a factor 2. The second tool is an online correction, which allows us to obtain centered bounds for many algorithms, to prevent the regret bounds from being vacuous when the domain is overly large or only partially constrained. The last tool, null updates, prevents the algorithm from performing overly large updates, which could result in unbounded regret, or even invalid updates. We develop a general theory to combine all these tools and apply it to several standard algorithms. In particular, we (almost entirely) restore the adaptivity to small losses of FTRL for unbounded domains, design and prove scale-free adaptive guarantees for a variant of Mirror Descent (at least when the Bregman divergence is convex in its second argument), extend Adapt-ML-Prod to scale-free guarantees, and provide several additional contributions about Prod, AdaHedge, BOA and Soft-Bayes.

Sultan J. Majeed, Marcus Hutter

The field of General Reinforcement Learning (GRL) formulates the problem of sequential decision-making from ground up. The history of interaction constitutes a "ground" state of the system, which never repeats. On the one hand, this generality allows GRL to model almost every domain possible, e.g.\ Bandits, MDPs, POMDPs, PSRs, and history-based environments. On the other hand, in general, the near-optimal policies in GRL are functions of complete history, which hinders not only learning but also planning in GRL. The usual way around for the planning part is that the agent is given a Markovian abstraction of the underlying process. So, it can use any MDP planning algorithm to find a near-optimal policy. The Extreme State Aggregation (ESA) framework has extended this idea to non-Markovian abstractions without compromising on the possibility of planning through a (surrogate) MDP. A distinguishing feature of ESA is that it proves an upper bound of $O\left(\varepsilon^{-A} \cdot (1-γ)^{-2A}\right)$ on the number of states required for the surrogate MDP (where $A$ is the number of actions, $γ$ is the discount-factor, and $\varepsilon$ is the optimality-gap) which holds \emph{uniformly} for \emph{all} domains. While the possibility of a universal bound is quite remarkable, we show that this bound is very loose. We propose a novel non-MDP abstraction which allows for a much better upper bound of $O\left(\varepsilon^{-1} \cdot (1-γ)^{-2} \cdot A \cdot 2^{A}\right)$. Furthermore, we show that this bound can be improved further to $O\left(\varepsilon^{-1} \cdot (1-γ)^{-2} \cdot \log^3 A \right)$ by using an action-sequentialization method.

Pedro A. Ortega, Markus Kunesch, Grégoire Delétang et al.

The recent phenomenal success of language models has reinvigorated machine learning research, and large sequence models such as transformers are being applied to a variety of domains. One important problem class that has remained relatively elusive however is purposeful adaptive behavior. Currently there is a common perception that sequence models "lack the understanding of the cause and effect of their actions" leading them to draw incorrect inferences due to auto-suggestive delusions. In this report we explain where this mismatch originates, and show that it can be resolved by treating actions as causal interventions. Finally, we show that in supervised learning, one can teach a system to condition or intervene on data by training with factual and counterfactual error signals respectively.

Samuel Allen Alexander, Marcus Hutter

Can an agent's intelligence level be negative? We extend the Legg-Hutter agent-environment framework to include punishments and argue for an affirmative answer to that question. We show that if the background encodings and Universal Turing Machine (UTM) admit certain Kolmogorov complexity symmetries, then the resulting Legg-Hutter intelligence measure is symmetric about the origin. In particular, this implies reward-ignoring agents have Legg-Hutter intelligence 0 according to such UTMs.

Elliot Catt, Marcus Hutter, Joel Veness

Reinforcement Learning formalises an embodied agent's interaction with the environment through observations, rewards and actions. But where do the actions come from? Actions are often considered to represent something external, such as the movement of a limb, a chess piece, or more generally, the output of an actuator. In this work we explore and formalize a contrasting view, namely that actions are best thought of as the output of a sequence of internal choices with respect to an action model. This view is particularly well-suited for leveraging the recent advances in large sequence models as prior knowledge for multi-task reinforcement learning problems. Our main contribution in this work is to show how to augment the standard MDP formalism with a sequential notion of internal action using information-theoretic techniques, and that this leads to self-consistent definitions of both internal and external action value functions.

Michael K. Cohen, Badri Vellambi, Marcus Hutter

Algorithmic Information Theory has inspired intractable constructions of general intelligence (AGI), and undiscovered tractable approximations are likely feasible. Reinforcement Learning (RL), the dominant paradigm by which an agent might learn to solve arbitrary solvable problems, gives an agent a dangerous incentive: to gain arbitrary "power" in order to intervene in the provision of their own reward. We review the arguments that generally intelligent algorithmic-information-theoretic reinforcement learners such as Hutter's (2005) AIXI would seek arbitrary power, including over us. Then, using an information-theoretic exploration schedule, and a setup inspired by causal influence theory, we present a variant of AIXI which learns to not seek arbitrary power; we call it "unambitious". We show that our agent learns to accrue reward at least as well as a human mentor, while relying on that mentor with diminishing probability. And given a formal assumption that we probe empirically, we show that eventually, the agent's world-model incorporates the following true fact: intervening in the "outside world" will have no effect on reward acquisition; hence, it has no incentive to shape the outside world.

Michael K. Cohen, Marcus Hutter, Neel Nanda

In imitation learning, imitators and demonstrators are policies for picking actions given past interactions with the environment. If we run an imitator, we probably want events to unfold similarly to the way they would have if the demonstrator had been acting the whole time. In general, one mistake during learning can lead to completely different events. In the special setting of environments that restart, existing work provides formal guidance in how to imitate so that events unfold similarly, but outside that setting, no formal guidance exists. We address a fully general setting, in which the (stochastic) environment and demonstrator never reset, not even for training purposes, and we allow our imitator to learn online from the demonstrator. Our new conservative Bayesian imitation learner underestimates the probabilities of each available action, and queries for more data with the remaining probability. Our main result: if an event would have been unlikely had the demonstrator acted the whole time, that event's likelihood can be bounded above when running the (initially totally ignorant) imitator instead. Meanwhile, queries to the demonstrator rapidly diminish in frequency. If any such event qualifies as "dangerous", our imitator would have the notable distinction of being relatively "safe".

Marcus Hutter

Recently a number of empirical "universal" scaling law papers have been published, most notably by OpenAI. `Scaling laws' refers to power-law decreases of training or test error w.r.t. more data, larger neural networks, and/or more compute. In this work we focus on scaling w.r.t. data size $n$. Theoretical understanding of this phenomenon is largely lacking, except in finite-dimensional models for which error typically decreases with $n^{-1/2}$ or $n^{-1}$, where $n$ is the sample size. We develop and theoretically analyse the simplest possible (toy) model that can exhibit $n^{-β}$ learning curves for arbitrary power $β>0$, and determine whether power laws are universal or depend on the data distribution.

Sultan Javed Majeed, Marcus Hutter

The reinforcement learning (RL) framework formalizes the notion of learning with interactions. Many real-world problems have large state-spaces and/or action-spaces such as in Go, StarCraft, protein folding, and robotics or are non-Markovian, which cause significant challenges to RL algorithms. In this work we address the large action-space problem by sequentializing actions, which can reduce the action-space size significantly, even down to two actions at the expense of an increased planning horizon. We provide explicit and exact constructions and equivalence proofs for all quantities of interest for arbitrary history-based processes. In the case of MDPs, this could help RL algorithms that bootstrap. In this work we show how action-binarization in the non-MDP case can significantly improve Extreme State Aggregation (ESA) bounds. ESA allows casting any (non-MDP, non-ergodic, history-based) RL problem into a fixed-sized non-Markovian state-space with the help of a surrogate Markovian process. On the upside, ESA enjoys similar optimality guarantees as Markovian models do. But a downside is that the size of the aggregated state-space becomes exponential in the size of the action-space. In this work, we patch this issue by binarizing the action-space. We provide an upper bound on the number of states of this binarized ESA that is logarithmic in the original action-space size, a double-exponential improvement.

Thomas Mesnard, Théophane Weber, Fabio Viola et al.

Credit assignment in reinforcement learning is the problem of measuring an action's influence on future rewards. In particular, this requires separating skill from luck, i.e. disentangling the effect of an action on rewards from that of external factors and subsequent actions. To achieve this, we adapt the notion of counterfactuals from causality theory to a model-free RL setup. The key idea is to condition value functions on future events, by learning to extract relevant information from a trajectory. We formulate a family of policy gradient algorithms that use these future-conditional value functions as baselines or critics, and show that they are provably low variance. To avoid the potential bias from conditioning on future information, we constrain the hindsight information to not contain information about the agent's actions. We demonstrate the efficacy and validity of our algorithm on a number of illustrative and challenging problems.

Jianan Wang, Eren Sezener, David Budden et al.

Human intelligence is characterized not only by the capacity to learn complex skills, but the ability to rapidly adapt and acquire new skills within an ever-changing environment. In this work we study how the learning of modular solutions can allow for effective generalization to both unseen and potentially differently distributed data. Our main postulate is that the combination of task segmentation, modular learning and memory-based ensembling can give rise to generalization on an exponentially growing number of unseen tasks. We provide a concrete instantiation of this idea using a combination of: (1) the Forget-Me-Not Process, for task segmentation and memory based ensembling; and (2) Gated Linear Networks, which in contrast to contemporary deep learning techniques use a modular and local learning mechanism. We demonstrate that this system exhibits a number of desirable continual learning properties: robustness to catastrophic forgetting, no negative transfer and increasing levels of positive transfer as more tasks are seen. We show competitive performance against both offline and online methods on standard continual learning benchmarks.

Marcus Hutter

Permutation-invariant, -equivariant, and -covariant functions and anti-symmetric functions are important in quantum physics, computer vision, and other disciplines. Applications often require most or all of the following properties: (a) a large class of such functions can be approximated, e.g. all continuous function, (b) only the (anti)symmetric functions can be represented, (c) a fast algorithm for computing the approximation, (d) the representation itself is continuous or differentiable, (e) the architecture is suitable for learning the function from data. (Anti)symmetric neural networks have recently been developed and applied with great success. A few theoretical approximation results have been proven, but many questions are still open, especially for particles in more than one dimension and the anti-symmetric case, which this work focusses on. More concretely, we derive natural polynomial approximations in the symmetric case, and approximations based on a single generalized Slater determinant in the anti-symmetric case. Unlike some previous super-exponential and discontinuous approximations, these seem a more promising basis for future tighter bounds. We provide a complete and explicit universality proof of the Equivariant MultiLayer Perceptron, which implies universality of symmetric MLPs and the FermiNet.

Laurent Orseau, Marcus Hutter, Omar Rivasplata

The Lottery Ticket Hypothesis is a conjecture that every large neural network contains a subnetwork that, when trained in isolation, achieves comparable performance to the large network. An even stronger conjecture has been proven recently: Every sufficiently overparameterized network contains a subnetwork that, at random initialization, but without training, achieves comparable accuracy to the trained large network. This latter result, however, relies on a number of strong assumptions and guarantees a polynomial factor on the size of the large network compared to the target function. In this work, we remove the most limiting assumptions of this previous work while providing significantly tighter bounds:the overparameterized network only needs a logarithmic factor (in all variables but depth) number of neurons per weight of the target subnetwork.

Michael K. Cohen, Marcus Hutter

If we could define the set of all bad outcomes, we could hard-code an agent which avoids them; however, in sufficiently complex environments, this is infeasible. We do not know of any general-purpose approaches in the literature to avoiding novel failure modes. Motivated by this, we define an idealized Bayesian reinforcement learner which follows a policy that maximizes the worst-case expected reward over a set of world-models. We call this agent pessimistic, since it optimizes assuming the worst case. A scalar parameter tunes the agent's pessimism by changing the size of the set of world-models taken into account. Our first main contribution is: given an assumption about the agent's model class, a sufficiently pessimistic agent does not cause "unprecedented events" with probability $1-δ$, whether or not designers know how to precisely specify those precedents they are concerned with. Since pessimism discourages exploration, at each timestep, the agent may defer to a mentor, who may be a human or some known-safe policy we would like to improve. Our other main contribution is that the agent's policy's value approaches at least that of the mentor, while the probability of deferring to the mentor goes to 0. In high-stakes environments, we might like advanced artificial agents to pursue goals cautiously, which is a non-trivial problem even if the agent were allowed arbitrary computing power; we present a formal solution.

Michael K. Cohen, Elliot Catt, Marcus Hutter

Reinforcement learners are agents that learn to pick actions that lead to high reward. Ideally, the value of a reinforcement learner's policy approaches optimality--where the optimal informed policy is the one which maximizes reward. Unfortunately, we show that if an agent is guaranteed to be "asymptotically optimal" in any (stochastically computable) environment, then subject to an assumption about the true environment, this agent will be either "destroyed" or "incapacitated" with probability 1. Much work in reinforcement learning uses an ergodicity assumption to avoid this problem. Often, doing theoretical research under simplifying assumptions prepares us to provide practical solutions even in the absence of those assumptions, but the ergodicity assumption in reinforcement learning may have led us entirely astray in preparing safe and effective exploration strategies for agents in dangerous environments. Rather than assuming away the problem, we present an agent, Mentee, with the modest guarantee of approaching the performance of a mentor, doing safe exploration instead of reckless exploration. Critically, Mentee's exploration probability depends on the expected information gain from exploring. In a simple non-ergodic environment with a weak mentor, we find Mentee outperforms existing asymptotically optimal agents and its mentor.

Eren Sezener, Marcus Hutter, David Budden et al.

We introduce a new and completely online contextual bandit algorithm called Gated Linear Contextual Bandits (GLCB). This algorithm is based on Gated Linear Networks (GLNs), a recently introduced deep learning architecture with properties well-suited to the online setting. Leveraging data-dependent gating properties of the GLN we are able to estimate prediction uncertainty with effectively zero algorithmic overhead. We empirically evaluate GLCB compared to 9 state-of-the-art algorithms that leverage deep neural networks, on a standard benchmark suite of discrete and continuous contextual bandit problems. GLCB obtains median first-place despite being the only online method, and we further support these results with a theoretical study of its convergence properties.

Joel Veness, Tor Lattimore, David Budden et al.

This paper presents a new family of backpropagation-free neural architectures, Gated Linear Networks (GLNs). What distinguishes GLNs from contemporary neural networks is the distributed and local nature of their credit assignment mechanism; each neuron directly predicts the target, forgoing the ability to learn feature representations in favor of rapid online learning. Individual neurons can model nonlinear functions via the use of data-dependent gating in conjunction with online convex optimization. We show that this architecture gives rise to universal learning capabilities in the limit, with effective model capacity increasing as a function of network size in a manner comparable with deep ReLU networks. Furthermore, we demonstrate that the GLN learning mechanism possesses extraordinary resilience to catastrophic forgetting, performing comparably to a MLP with dropout and Elastic Weight Consolidation on standard benchmarks. These desirable theoretical and empirical properties position GLNs as a complementary technique to contemporary offline deep learning methods.

Tom Everitt, Marcus Hutter, Ramana Kumar et al.

Can humans get arbitrarily capable reinforcement learning (RL) agents to do their bidding? Or will sufficiently capable RL agents always find ways to bypass their intended objectives by shortcutting their reward signal? This question impacts how far RL can be scaled, and whether alternative paradigms must be developed in order to build safe artificial general intelligence. In this paper, we study when an RL agent has an instrumental goal to tamper with its reward process, and describe design principles that prevent instrumental goals for two different types of reward tampering (reward function tampering and RF-input tampering). Combined, the design principles can prevent both types of reward tampering from being instrumental goals. The analysis benefits from causal influence diagrams to provide intuitive yet precise formalizations.

Marcus Hutter

A popular approach of achieving fairness in optimization problems is by constraining the solution space to "fair" solutions, which unfortunately typically reduces solution quality. In practice, the ultimate goal is often an aggregate of sub-goals without a unique or best way of combining them or which is otherwise only partially known. I turn this problem into a feature and suggest to use a parametrized objective and vary the parameters within reasonable ranges to get a "set" of optimal solutions, which can then be optimized using secondary criteria such as fairness without compromising the primary objective, i.e. without regret (societal cost).

Michael K Cohen, Badri Vellambi, Marcus Hutter

General intelligence, the ability to solve arbitrary solvable problems, is supposed by many to be artificially constructible. Narrow intelligence, the ability to solve a given particularly difficult problem, has seen impressive recent development. Notable examples include self-driving cars, Go engines, image classifiers, and translators. Artificial General Intelligence (AGI) presents dangers that narrow intelligence does not: if something smarter than us across every domain were indifferent to our concerns, it would be an existential threat to humanity, just as we threaten many species despite no ill will. Even the theory of how to maintain the alignment of an AGI's goals with our own has proven highly elusive. We present the first algorithm we are aware of for asymptotically unambitious AGI, where "unambitiousness" includes not seeking arbitrary power. Thus, we identify an exception to the Instrumental Convergence Thesis, which is roughly that by default, an AGI would seek power, including over us.