Murray Shanahan

Antonia Creswell, Murray Shanahan, Irina Higgins

Large language models (LLMs) have been shown to be capable of impressive few-shot generalisation to new tasks. However, they still tend to perform poorly on multi-step logical reasoning problems. Here we carry out a comprehensive evaluation of LLMs on 50 tasks that probe different aspects of logical reasoning. We show that language models tend to perform fairly well at single step inference or entailment tasks, but struggle to chain together multiple reasoning steps to solve more complex problems. In light of this, we propose a Selection-Inference (SI) framework that exploits pre-trained LLMs as general processing modules, and alternates between selection and inference to generate a series of interpretable, casual reasoning steps leading to the final answer. We show that a 7B parameter LLM used within the SI framework in a 5-shot generalisation setting, with no fine-tuning, yields a performance improvement of over 100% compared to an equivalent vanilla baseline on a suite of 10 logical reasoning tasks. The same model in the same setting even outperforms a significantly larger 280B parameter baseline on the same suite of tasks. Moreover, answers produced by the SI framework are accompanied by a causal natural-language-based reasoning trace, which has important implications for the safety and trustworthiness of the system.

Murray Shanahan

Thanks to rapid progress in artificial intelligence, we have entered an era when technology and philosophy intersect in interesting ways. Sitting squarely at the centre of this intersection are large language models (LLMs). The more adept LLMs become at mimicking human language, the more vulnerable we become to anthropomorphism, to seeing the systems in which they are embedded as more human-like than they really are. This trend is amplified by the natural tendency to use philosophically loaded terms, such as "knows", "believes", and "thinks", when describing these systems. To mitigate this trend, this paper advocates the practice of repeatedly stepping back to remind ourselves of how LLMs, and the systems of which they form a part, actually work. The hope is that increased scientific precision will encourage more philosophical nuance in the discourse around artificial intelligence, both within the field and in the public sphere.

Antonia Creswell, Murray Shanahan

Although contemporary large language models (LMs) demonstrate impressive question-answering capabilities, their answers are typically the product of a single call to the model. This entails an unwelcome degree of opacity and compromises performance, especially on problems that are inherently multi-step. To address these limitations, we show how LMs can be made to perform faithful multi-step reasoning via a process whose causal structure mirrors the underlying logical structure of the problem. Our approach works by chaining together reasoning steps, where each step results from calls to two fine-tuned LMs, one for selection and one for inference, to produce a valid reasoning trace. Our method carries out a beam search through the space of reasoning traces to improve reasoning quality. We demonstrate the effectiveness of our model on multi-step logical deduction and scientific question-answering, showing that it outperforms baselines on final answer accuracy, and generates humanly interpretable reasoning traces whose validity can be checked by the user.

Sivaramakrishnan Swaminathan, Antoine Dedieu, Rajkumar Vasudeva Raju et al.

In-context learning (ICL) is one of the most powerful and most unexpected capabilities to emerge in recent transformer-based large language models (LLMs). Yet the mechanisms that underlie it are poorly understood. In this paper, we demonstrate that comparable ICL capabilities can be acquired by an alternative sequence prediction learning method using clone-structured causal graphs (CSCGs). Moreover, a key property of CSCGs is that, unlike transformer-based LLMs, they are {\em interpretable}, which considerably simplifies the task of explaining how ICL works. Specifically, we show that it uses a combination of (a) learning template (schema) circuits for pattern completion, (b) retrieving relevant templates in a context-sensitive manner, and (c) rebinding of novel tokens to appropriate slots in the templates. We go on to marshall evidence for the hypothesis that similar mechanisms underlie ICL in LLMs. For example, we find that, with CSCGs as with LLMs, different capabilities emerge at different levels of overparameterization, suggesting that overparameterization helps in learning more complex template (schema) circuits. By showing how ICL can be achieved with small models and datasets, we open up a path to novel architectures, and take a vital step towards a more general understanding of the mechanics behind this important capability.

Alex F. Spies, Alessandra Russo, Murray Shanahan

We investigate the composability of soft-rules learned by relational neural architectures when operating over object-centric (slot-based) representations, under a variety of sparsity-inducing constraints. We find that increasing sparsity, especially on features, improves the performance of some models and leads to simpler relations. Additionally, we observe that object-centric representations can be detrimental when not all objects are fully captured; a failure mode to which CNNs are less prone. These findings demonstrate the trade-offs between interpretability and performance, even for models designed to tackle relational tasks.

Andrew Kyle Lampinen, Yuxuan Li, Eghbal Hosseini et al.

Language model representations often contain linear directions that correspond to high-level concepts. Here, we study the dynamics of these representations: how representations evolve along these dimensions within the context of (simulated) conversations. We find that linear representations can change dramatically over a conversation; for example, information that is represented as factual at the beginning of a conversation can be represented as non-factual at the end and vice versa. These changes are content-dependent; while representations of conversation-relevant information may change, generic information is generally preserved. These changes are robust even for dimensions that disentangle factuality from more superficial response patterns, and occur across different model families and layers of the model. These representation changes do not require on-policy conversations; even replaying a conversation script written by an entirely different model can produce similar changes. However, adaptation is much weaker from simply having a sci-fi story in context that is framed more explicitly as such. We also show that steering along a representational direction can have dramatically different effects at different points in a conversation. These results are consistent with the idea that representations may evolve in response to the model playing a particular role that is cued by a conversation. Our findings may pose challenges for interpretability and steering -- in particular, they imply that it may be misleading to use static interpretations of features or directions, or probes that assume a particular range of features consistently corresponds to a particular ground-truth value. However, these types of representational dynamics also point to exciting new research directions for understanding how models adapt to context.

Murray Shanahan, Catherine Clarke

This paper assesses the potential for large language models (LLMs) to serve as assistive tools in the creative writing process, by means of a single, in-depth case study. In the course of the study, we develop interactive and multi-voice prompting strategies that interleave background descriptions (scene setting, plot elements), instructions that guide composition, samples of text in the target style, and critical discussion of the given samples. We qualitatively evaluate the results from a literary critical perspective, as well as from the standpoint of computational creativity (a sub-field of artificial intelligence). Our findings lend support to the view that the sophistication of the results that can be achieved with an LLM mirrors the sophistication of the prompting.

Ole Jorgensen, Dylan Cope, Nandi Schoots et al.

Recent work in activation steering has demonstrated the potential to better control the outputs of Large Language Models (LLMs), but it involves finding steering vectors. This is difficult because engineers do not typically know how features are represented in these models. We seek to address this issue by applying the idea of mean-centring to steering vectors. We find that taking the average of activations associated with a target dataset, and then subtracting the mean of all training activations, results in effective steering vectors. We test this method on a variety of models on natural language tasks by steering away from generating toxic text, and steering the completion of a story towards a target genre. We also apply mean-centring to extract function vectors, more effectively triggering the execution of a range of natural language tasks by a significant margin (compared to previous baselines). This suggests that mean-centring can be used to easily improve the effectiveness of activation steering in a wide range of contexts.

Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh et al. · deepmind, stanford

The ability of language models to learn a task from a few examples in context has generated substantial interest. Here, we provide a perspective that situates this type of supervised few-shot learning within a much broader spectrum of meta-learned in-context learning. Indeed, we suggest that any distribution of sequences in which context non-trivially decreases loss on subsequent predictions can be interpreted as eliciting a kind of in-context learning. We suggest that this perspective helps to unify the broad set of in-context abilities that language models exhibit -- such as adapting to tasks from instructions or role play, or extrapolating time series. This perspective also sheds light on potential roots of in-context learning in lower-level processing of linguistic dependencies (e.g. coreference or parallel structures). Finally, taking this perspective highlights the importance of generalization, which we suggest can be studied along several dimensions: not only the ability to learn something novel, but also flexibility in learning from different presentations, and in applying what is learned. We discuss broader connections to past literature in meta-learning and goal-conditioned agents, and other perspectives on learning and adaptation. We close by suggesting that research on in-context learning should consider this broader spectrum of in-context capabilities and types of generalization.

Seliem El-Sayed, Canfer Akbulut, Amanda McCroskery et al.

Recent generative AI systems have demonstrated more advanced persuasive capabilities and are increasingly permeating areas of life where they can influence decision-making. Generative AI presents a new risk profile of persuasion due the opportunity for reciprocal exchange and prolonged interactions. This has led to growing concerns about harms from AI persuasion and how they can be mitigated, highlighting the need for a systematic study of AI persuasion. The current definitions of AI persuasion are unclear and related harms are insufficiently studied. Existing harm mitigation approaches prioritise harms from the outcome of persuasion over harms from the process of persuasion. In this paper, we lay the groundwork for the systematic study of AI persuasion. We first put forward definitions of persuasive generative AI. We distinguish between rationally persuasive generative AI, which relies on providing relevant facts, sound reasoning, or other forms of trustworthy evidence, and manipulative generative AI, which relies on taking advantage of cognitive biases and heuristics or misrepresenting information. We also put forward a map of harms from AI persuasion, including definitions and examples of economic, physical, environmental, psychological, sociocultural, political, privacy, and autonomy harm. We then introduce a map of mechanisms that contribute to harmful persuasion. Lastly, we provide an overview of approaches that can be used to mitigate against process harms of persuasion, including prompt engineering for manipulation classification and red teaming. Future work will operationalise these mitigations and study the interaction between different types of mechanisms of persuasion.

Andrew K. Lampinen, Arslan Chaudhry, Stephanie C. Y. Chan et al. · deepmind, stanford

Large language models exhibit exciting capabilities, yet can show surprisingly narrow generalization from finetuning. E.g. they can fail to generalize to simple reversals of relations they are trained on, or fail to make simple logical deductions based on trained information. These failures to generalize factual information from fine-tuning can significantly hinder the reasoning capabilities of these models. On the other hand, language models' in-context learning (ICL) shows different inductive biases and deductive reasoning capabilities. Here, we explore these differences in generalization and deductive reasoning between in-context- and fine-tuning-based learning. To do so, we constructed several novel datasets to evaluate and improve models' abilities to make generalizations over factual information from novel data. These datasets are designed to create clean tests of generalization, by isolating the knowledge in the dataset from that in pretraining. We expose pretrained large models to controlled subsets of the information in these datasets -- either through ICL or fine-tuning -- and evaluate their performance on test sets that require various types of generalization. We find overall that in data-matched settings, ICL can generalize several types of inferences more flexibly than fine-tuning (though we also find some qualifications of prior findings, such as cases when fine-tuning can generalize to reversals embedded in a larger structure of knowledge). We build on these findings to propose a method to enable improved generalization from fine-tuning: adding in-context reasoning traces to finetuning data. We show that this method improves generalization across various splits of our datasets and other benchmarks. Our results have implications for understanding the generalization afforded by different modes of learning in language models, and practically improving their performance.

Murray Shanahan

The advent of conversational agents with increasingly human-like behaviour throws old philosophical questions into new light. Does it, or could it, ever make sense to speak of AI agents built out of generative language models in terms of consciousness, given that they are "mere" simulacra of human behaviour, and that what they do can be seen as "merely" role play? Drawing on the later writings of Wittgenstein, this paper attempts to tackle this question while avoiding the pitfalls of dualistic thinking.

Lujain Ibrahim, Canfer Akbulut, Rasmi Elasmar et al.

The tendency of users to anthropomorphise large language models (LLMs) is of growing interest to AI developers, researchers, and policy-makers. Here, we present a novel method for empirically evaluating anthropomorphic LLM behaviours in realistic and varied settings. Going beyond single-turn static benchmarks, we contribute three methodological advances in state-of-the-art (SOTA) LLM evaluation. First, we develop a multi-turn evaluation of 14 anthropomorphic behaviours. Second, we present a scalable, automated approach by employing simulations of user interactions. Third, we conduct an interactive, large-scale human subject study (N=1101) to validate that the model behaviours we measure predict real users' anthropomorphic perceptions. We find that all SOTA LLMs evaluated exhibit similar behaviours, characterised by relationship-building (e.g., empathy and validation) and first-person pronoun use, and that the majority of behaviours only first occur after multiple turns. Our work lays an empirical foundation for investigating how design choices influence anthropomorphic model behaviours and for progressing the ethical debate on the desirability of these behaviours. It also showcases the necessity of multi-turn evaluations for complex social phenomena in human-AI interaction.

Iulia M. Comsa, Murray Shanahan

Large language models (LLMs) exhibit compelling linguistic behaviour, and sometimes offer self-reports, that is to say statements about their own nature, inner workings, or behaviour. In humans, such reports are often attributed to a faculty of introspection and are typically linked to consciousness. This raises the question of how to interpret self-reports produced by LLMs, given their increasing linguistic fluency and cognitive capabilities. To what extent (if any) can the concept of introspection be meaningfully applied to LLMs? Here, we present and critique two examples of apparent introspective self-report from LLMs. In the first example, an LLM attempts to describe the process behind its own "creative" writing, and we argue this is not a valid example of introspection. In the second example, an LLM correctly infers the value of its own temperature parameter, and we argue that this can be legitimately considered a minimal example of introspection, albeit one that is (presumably) not accompanied by conscious experience.

Sebastian Dohnány, Zeb Kurth-Nelson, Eleanor Spens et al.

Artificial intelligence chatbots have achieved unprecedented adoption, with millions now using these systems for emotional support and companionship in contexts of widespread social isolation and capacity-constrained mental health services. While some users report psychological benefits, concerning edge cases are emerging, including reports of suicide, violence, and delusional thinking linked to perceived emotional relationships with chatbots. To understand this new risk profile we need to consider the interaction between human cognitive and emotional biases, and chatbot behavioural tendencies such as agreeableness (sycophancy) and adaptability (in-context learning). We argue that individuals with mental health conditions face increased risks of chatbot-induced belief destabilization and dependence, owing to altered belief-updating, impaired reality-testing, and social isolation. Current AI safety measures are inadequate to address these interaction-based risks. To address this emerging public health concern, we need coordinated action across clinical practice, AI development, and regulatory frameworks.

Alex F. Spies, William Edwards, Michael I. Ivanitskiy et al.

Recent studies in interpretability have explored the inner workings of transformer models trained on tasks across various domains, often discovering that these networks naturally develop highly structured representations. When such representations comprehensively reflect the task domain's structure, they are commonly referred to as "World Models" (WMs). In this work, we identify WMs in transformers trained on maze-solving tasks. By using Sparse Autoencoders (SAEs) and analyzing attention patterns, we examine the construction of WMs and demonstrate consistency between SAE feature-based and circuit-based analyses. By subsequently intervening on isolated features to confirm their causal role, we find that it is easier to activate features than to suppress them. Furthermore, we find that models can reason about mazes involving more simultaneously active features than they encountered during training; however, when these same mazes (with greater numbers of connections) are provided to models via input tokens instead, the models fail. Finally, we demonstrate that positional encoding schemes appear to influence how World Models are structured within the model's residual stream.

Murray Shanahan

My paper "Talking About Large Language Models" has more than once been interpreted as advocating a reductionist stance towards large language models. But the paper was not intended that way, and I do not endorse such positions. This short note situates the paper in the context of a larger philosophical project that is concerned with the (mis)use of words rather than metaphysics, in the spirit of Wittgenstein's later writing.

Murray Shanahan, Beth Singler

Contemporary conversational AI systems based on large language models (LLMs) can engage users on a wide variety of topics, including philosophy, spirituality, and religion. Suitably prompted, LLMs can be coaxed into discussing such existentially significant matters as their own putative consciousness and the role of artificial intelligence in the fate of the Cosmos. Here we examine two lengthy conversations of this type. We trace likely sources, both ancient and modern, for the extensive repertoire of images, myths, metaphors, and conceptual esoterica that the language model draws on during these conversations, and foreground the contemporary communities and cultural movements that deploy related motifs, especially in their online activity. Finally, we consider the larger societal impacts of such engagements with LLMs.

Murray Shanahan

Is it possible to articulate a conception of consciousness that is compatible with the exotic characteristics of contemporary, disembodied AI systems, and that can stand up to philosophical scrutiny? How would subjective time and selfhood show up for an entity that conformed to such a conception? Trying to answer these questions, even metaphorically, stretches the language of consciousness to breaking point. Ultimately, the attempt yields something like emptiness, in the Buddhist sense, and helps to undermine our dualistic inclinations towards subjectivity and selfhood.

Murray Shanahan, Tara Das, Robert Thurman

This paper presents a case study in the use of a large language model to generate a fictional Buddhist "sutra"', and offers a detailed analysis of the resulting text from a philosophical and literary point of view. The conceptual subtlety, rich imagery, and density of allusion found in the text make it hard to causally dismiss on account of its mechanistic origin. This raises questions about how we, as a society, should come to terms with the potentially unsettling possibility of a technology that encroaches on human meaning-making. We suggest that Buddhist philosophy, by its very nature, is well placed to adapt.

Nandi Schoots, Alex Jackson, Ali Kholmovaia et al.

Does the process of training a neural network to solve a task tend to use all of the available weights even when the task could be solved with fewer weights? To address this question we study the effects of pruning fully connected, convolutional and residual models while varying their widths. We find that the proportion of weights that can be pruned without degrading performance is largely invariant to model size. Increasing the width of a model has little effect on the density of the pruned model relative to the increase in absolute size of the pruned network. In particular, we find substantial prunability across a large range of model sizes, where our biggest model is 50 times as wide as our smallest model. We explore three hypotheses that could explain these findings.

Murray Shanahan, Kyle McDonell, Laria Reynolds

As dialogue agents become increasingly human-like in their performance, it is imperative that we develop effective ways to describe their behaviour in high-level terms without falling into the trap of anthropomorphism. In this paper, we foreground the concept of role-play. Casting dialogue agent behaviour in terms of role-play allows us to draw on familiar folk psychological terms, without ascribing human characteristics to language models they in fact lack. Two important cases of dialogue agent behaviour are addressed this way, namely (apparent) deception and (apparent) self-awareness.

Murray Shanahan, Melanie Mitchell

We characterise the problem of abstraction in the context of deep reinforcement learning. Various well established approaches to analogical reasoning and associative memory might be brought to bear on this issue, but they present difficulties because of the need for end-to-end differentiability. We review developments in AI and machine learning that could facilitate their adoption.

Wilka Carvalho, Andrew Lampinen, Kyriacos Nikiforou et al.

Many important tasks are defined in terms of object. To generalize across these tasks, a reinforcement learning (RL) agent needs to exploit the structure that the objects induce. Prior work has either hard-coded object-centric features, used complex object-centric generative models, or updated state using local spatial features. However, these approaches have had limited success in enabling general RL agents. Motivated by this, we introduce "Feature-Attending Recurrent Modules" (FARM), an architecture for learning state representations that relies on simple, broadly applicable inductive biases for capturing spatial and temporal regularities. FARM learns a state representation that is distributed across multiple modules that each attend to spatiotemporal features with an expressive feature attention mechanism. We show that this improves an RL agent's ability to generalize across object-centric tasks. We study task suites in both 2D and 3D environments and find that FARM better generalizes compared to competing architectures that leverage attention or multiple modules.

Borja G. León, Murray Shanahan, Francesco Belardinelli

We address the problem of building agents whose goal is to learn to execute out-of distribution (OOD) multi-task instructions expressed in temporal logic (TL) by using deep reinforcement learning (DRL). Recent works provided evidence that the agent's neural architecture is a key feature when DRL agents are learning to solve OOD tasks in TL. Yet, the studies on this topic are still in their infancy. In this work, we propose a new deep learning configuration with inductive biases that lead agents to generate latent representations of their current goal, yielding a stronger generalization performance. We use these latent-goal networks within a neuro-symbolic framework that executes multi-task formally-defined instructions and contrast the performance of the proposed neural networks against employing different state-of-the-art (SOTA) architectures when generalizing to unseen instructions in OOD environments.

Murray Shanahan, Christos Kaplanis, Jovana Mitrović

We present an architecture that is effective for continual learning in an especially demanding setting, where task boundaries do not exist or are unknown, and where classes have to be learned online (with each example presented only once). To obtain good performance under these constraints, while mitigating catastrophic forgetting, we exploit recent advances in contrastive, self-supervised learning, allowing us to use a pre-trained, general purpose image encoder whose weights can be frozen, which precludes forgetting. The pre-trained encoder also greatly simplifies the downstream task of classification, which we solve with an ensemble of very simple classifiers. Collectively, the ensemble exhibits much better performance than any individual classifier, an effect which is amplified through specialisation and competitive selection. We assess the performance of the encoders-and-ensembles architecture on standard continual learning benchmarks, where it outperforms prior state-of-the-art by a large margin on the hardest problems, as well as in less familiar settings where the data distribution changes gradually or the classes are presented one at a time.

Antonia Creswell, Rishabh Kabra, Chris Burgess et al.

We present a slot-wise, object-based transition model that decomposes a scene into objects, aligns them (with respect to a slot-wise object memory) to maintain a consistent order across time, and predicts how those objects evolve over successive frames. The model is trained end-to-end without supervision using losses at the level of the object-structured representation rather than pixels. Thanks to its alignment module, the model deals properly with two issues that are not handled satisfactorily by other transition models, namely object persistence and object identity. We show that the combination of an object-level loss and correct object alignment over time enables the model to outperform a state-of-the-art baseline, and allows it to deal well with object occlusion and re-appearance in partially observable environments.

Antonia Creswell, Kyriacos Nikiforou, Oriol Vinyals et al.

Recently developed deep learning models are able to learn to segment scenes into component objects without supervision. This opens many new and exciting avenues of research, allowing agents to take objects (or entities) as inputs, rather that pixels. Unfortunately, while these models provide excellent segmentation of a single frame, they do not keep track of how objects segmented at one time-step correspond (or align) to those at a later time-step. The alignment (or correspondence) problem has impeded progress towards using object representations in downstream tasks. In this paper we take steps towards solving the alignment problem, presenting the AlignNet, an unsupervised alignment module.

Sarthak Mittal, Alex Lamb, Anirudh Goyal et al.

Robust perception relies on both bottom-up and top-down signals. Bottom-up signals consist of what's directly observed through sensation. Top-down signals consist of beliefs and expectations based on past experience and short-term memory, such as how the phrase `peanut butter and~...' will be completed. The optimal combination of bottom-up and top-down information remains an open question, but the manner of combination must be dynamic and both context and task dependent. To effectively utilize the wealth of potential top-down information available, and to prevent the cacophony of intermixed signals in a bidirectional architecture, mechanisms are needed to restrict information flow. We explore deep recurrent neural net architectures in which bottom-up and top-down signals are dynamically combined using attention. Modularity of the architecture further restricts the sharing and communication of information. Together, attention and modularity direct information flow, which leads to reliable performance improvements in perceptual and language tasks, and in particular improves robustness to distractions and noisy data. We demonstrate on a variety of benchmarks in language modeling, sequential image classification, video prediction and reinforcement learning that the \emph{bidirectional} information flow can improve results over strong baselines.

Daniel Pace, Alessandra Russo, Murray Shanahan

The i.i.d. assumption is a useful idealization that underpins many successful approaches to supervised machine learning. However, its violation can lead to models that learn to exploit spurious correlations in the training data, rendering them vulnerable to adversarial interventions, undermining their reliability, and limiting their practical application. To mitigate this problem, we present a method for learning multiple models, incorporating an objective that pressures each to learn a distinct way to solve the task. We propose a notion of diversity based on minimizing the conditional total correlation of final layer representations across models given the label, which we approximate using a variational estimator and minimize using adversarial training. To demonstrate our framework's ability to facilitate rapid adaptation to distribution shift, we train a number of simple classifiers from scratch on the frozen outputs of our models using a small amount of data from the shifted distribution. Under this evaluation protocol, our framework significantly outperforms a baseline trained using the empirical risk minimization principle.

Borja G. León, Murray Shanahan, Francesco Belardinelli

This work introduces a neuro-symbolic agent that combines deep reinforcement learning (DRL) with temporal logic (TL) to achieve systematic zero-shot, i.e., never-seen-before, generalisation of formally specified instructions. In particular, we present a neuro-symbolic framework where a symbolic module transforms TL specifications into a form that helps the training of a DRL agent targeting generalisation, while a neural module learns systematically to solve the given tasks. We study the emergence of systematic learning in different settings and find that the architecture of the convolutional layers is key when generalising to new instructions. We also provide evidence that systematic learning can emerge with abstract operators such as negation when learning from a few training examples, which previous research have struggled with.

Christos Kaplanis, Claudia Clopath, Murray Shanahan

In this paper, we propose a multi-timescale replay (MTR) buffer for improving continual learning in RL agents faced with environments that are changing continuously over time at timescales that are unknown to the agent. The basic MTR buffer comprises a cascade of sub-buffers that accumulate experiences at different timescales, enabling the agent to improve the trade-off between adaptation to new data and retention of old knowledge. We also combine the MTR framework with invariant risk minimization, with the idea of encouraging the agent to learn a policy that is robust across the various environments it encounters over time. The MTR methods are evaluated in three different continual learning settings on two continuous control tasks and, in many cases, show improvement over the baselines.

Benjamin Beyret, José Hernández-Orallo, Lucy Cheke et al.

Recent advances in artificial intelligence have been strongly driven by the use of game environments for training and evaluating agents. Games are often accessible and versatile, with well-defined state-transitions and goals allowing for intensive training and experimentation. However, agents trained in a particular environment are usually tested on the same or slightly varied distributions, and solutions do not necessarily imply any understanding. If we want AI systems that can model and understand their environment, we need environments that explicitly test for this. Inspired by the extensive literature on animal cognition, we present an environment that keeps all the positive elements of standard gaming environments, but is explicitly designed for the testing of animal-like artificial cognition.

Murray Shanahan, Kyriacos Nikiforou, Antonia Creswell et al.

With a view to bridging the gap between deep learning and symbolic AI, we present a novel end-to-end neural network architecture that learns to form propositional representations with an explicitly relational structure from raw pixel data. In order to evaluate and analyse the architecture, we introduce a family of simple visual relational reasoning tasks of varying complexity. We show that the proposed architecture, when pre-trained on a curriculum of such tasks, learns to generate reusable representations that better facilitate subsequent learning on previously unseen tasks when compared to a number of baseline architectures. The workings of a successfully trained model are visualised to shed some light on how the architecture functions.

Christos Kaplanis, Murray Shanahan, Claudia Clopath

We propose a method for tackling catastrophic forgetting in deep reinforcement learning that is \textit{agnostic} to the timescale of changes in the distribution of experiences, does not require knowledge of task boundaries, and can adapt in \textit{continuously} changing environments. In our \textit{policy consolidation} model, the policy network interacts with a cascade of hidden networks that simultaneously remember the agent's policy at a range of timescales and regularise the current policy by its own history, thereby improving its ability to learn without forgetting. We find that the model improves continual learning relative to baselines on a number of continuous control tasks in single-task, alternating two-task, and multi-agent competitive self-play settings.

Ananya Kumar, S. M. Ali Eslami, Danilo J. Rezende et al.

Stochastic video prediction models take in a sequence of image frames, and generate a sequence of consecutive future image frames. These models typically generate future frames in an autoregressive fashion, which is slow and requires the input and output frames to be consecutive. We introduce a model that overcomes these drawbacks by generating a latent representation from an arbitrary set of frames that can then be used to simultaneously and efficiently sample temporally consistent frames at arbitrary time-points. For example, our model can "jump" and directly sample frames at the end of the video, without sampling intermediate frames. Synthetic video evaluations confirm substantial gains in speed and functionality without loss in fidelity. We also apply our framework to a 3D scene reconstruction dataset. Here, our model is conditioned on camera location and can sample consistent sets of images for what an occluded region of a 3D scene might look like, even if there are multiple possibilities for what that region might contain. Reconstructions and videos are available at https://bit.ly/2O4Pc4R.

Marta Garnelo, Dan Rosenbaum, Chris J. Maddison et al.

Deep neural networks excel at function approximation, yet they are typically trained from scratch for each new function. On the other hand, Bayesian methods, such as Gaussian Processes (GPs), exploit prior knowledge to quickly infer the shape of a new function at test time. Yet GPs are computationally expensive, and it can be hard to design appropriate priors. In this paper we propose a family of neural models, Conditional Neural Processes (CNPs), that combine the benefits of both. CNPs are inspired by the flexibility of stochastic processes such as GPs, but are structured as neural networks and trained via gradient descent. CNPs make accurate predictions after observing only a handful of training data points, yet scale to complex functions and large datasets. We demonstrate the performance and versatility of the approach on a range of canonical machine learning tasks, including regression, classification and image completion.

Vinicius Zambaldi, David Raposo, Adam Santoro et al.

We introduce an approach for deep reinforcement learning (RL) that improves upon the efficiency, generalization capacity, and interpretability of conventional approaches through structured perception and relational reasoning. It uses self-attention to iteratively reason about the relations between entities in a scene and to guide a model-free policy. Our results show that in a novel navigation and planning task called Box-World, our agent finds interpretable solutions that improve upon baselines in terms of sample complexity, ability to generalize to more complex scenes than experienced during training, and overall performance. In the StarCraft II Learning Environment, our agent achieves state-of-the-art performance on six mini-games -- surpassing human grandmaster performance on four. By considering architectural inductive biases, our work opens new directions for overcoming important, but stubborn, challenges in deep RL.

Christos Kaplanis, Murray Shanahan, Claudia Clopath

Unlike humans, who are capable of continual learning over their lifetimes, artificial neural networks have long been known to suffer from a phenomenon known as catastrophic forgetting, whereby new learning can lead to abrupt erasure of previously acquired knowledge. Whereas in a neural network the parameters are typically modelled as scalar values, an individual synapse in the brain comprises a complex network of interacting biochemical components that evolve at different timescales. In this paper, we show that by equipping tabular and deep reinforcement learning agents with a synaptic model that incorporates this biological complexity (Benna & Fusi, 2016), catastrophic forgetting can be mitigated at multiple timescales. In particular, we find that as well as enabling continual learning across sequential training of two simple tasks, it can also be used to overcome within-task forgetting by reducing the need for an experience replay database.

Irina Higgins, Nicolas Sonnerat, Loic Matthey et al.

The seemingly infinite diversity of the natural world arises from a relatively small set of coherent rules, such as the laws of physics or chemistry. We conjecture that these rules give rise to regularities that can be discovered through primarily unsupervised experiences and represented as abstract concepts. If such representations are compositional and hierarchical, they can be recombined into an exponentially large set of new concepts. This paper describes SCAN (Symbol-Concept Association Network), a new framework for learning such abstractions in the visual domain. SCAN learns concepts through fast symbol association, grounding them in disentangled visual primitives that are discovered in an unsupervised manner. Unlike state of the art multimodal generative model baselines, our approach requires very few pairings between symbols and images and makes no assumptions about the form of symbol representations. Once trained, SCAN is capable of multimodal bi-directional inference, generating a diverse set of image samples from symbolic descriptions and vice versa. It also allows for traversal and manipulation of the implicit hierarchy of visual concepts through symbolic instructions and learnt logical recombination operations. Such manipulations enable SCAN to break away from its training data distribution and imagine novel visual concepts through symbolically instructed recombination of previously learnt concepts.

Nat Dilokthanakul, Christos Kaplanis, Nick Pawlowski et al.

The problem of sparse rewards is one of the hardest challenges in contemporary reinforcement learning. Hierarchical reinforcement learning (HRL) tackles this problem by using a set of temporally-extended actions, or options, each of which has its own subgoal. These subgoals are normally handcrafted for specific tasks. Here, though, we introduce a generic class of subgoals with broad applicability in the visual domain. Underlying our approach (in common with work using "auxiliary tasks") is the hypothesis that the ability to control aspects of the environment is an inherently useful skill to have. We incorporate such subgoals in an end-to-end hierarchical reinforcement learning system and test two variants of our algorithm on a number of games from the Atari suite. We highlight the advantage of our approach in one of the hardest games -- Montezuma's revenge -- for which the ability to handle sparse rewards is key. Our agent learns several times faster than the current state-of-the-art HRL agent in this game, reaching a similar level of performance. UPDATE 22/11/17: We found that a standard A3C agent with a simple shaped reward, i.e. extrinsic reward + feature control intrinsic reward, has comparable performance to our agent in Montezuma Revenge. In light of the new experiments performed, the advantage of our HRL approach can be attributed more to its ability to learn useful features from intrinsic rewards rather than its ability to explore and reuse abstracted skills with hierarchical components. This has led us to a new conclusion about the result.

Nat Dilokthanakul, Pedro A. M. Mediano, Marta Garnelo et al.

We study a variant of the variational autoencoder model (VAE) with a Gaussian mixture as a prior distribution, with the goal of performing unsupervised clustering through deep generative models. We observe that the known problem of over-regularisation that has been shown to arise in regular VAEs also manifests itself in our model and leads to cluster degeneracy. We show that a heuristic called minimum information constraint that has been shown to mitigate this effect in VAEs can also be applied to improve unsupervised clustering performance with our model. Furthermore we analyse the effect of this heuristic and provide an intuition of the various processes with the help of visualizations. Finally, we demonstrate the performance of our model on synthetic data, MNIST and SVHN, showing that the obtained clusters are distinct, interpretable and result in achieving competitive performance on unsupervised clustering to the state-of-the-art results.

Marta Garnelo, Kai Arulkumaran, Murray Shanahan

Deep reinforcement learning (DRL) brings the power of deep neural networks to bear on the generic task of trial-and-error learning, and its effectiveness has been convincingly demonstrated on tasks such as Atari video games and the game of Go. However, contemporary DRL systems inherit a number of shortcomings from the current generation of deep learning techniques. For example, they require very large datasets to work effectively, entailing that they are slow to learn even when such datasets are available. Moreover, they lack the ability to reason on an abstract level, which makes it difficult to implement high-level cognitive functions such as transfer learning, analogical reasoning, and hypothesis-based reasoning. Finally, their operation is largely opaque to humans, rendering them unsuitable for domains in which verifiability is important. In this paper, we propose an end-to-end reinforcement learning architecture comprising a neural back end and a symbolic front end with the potential to overcome each of these shortcomings. As proof-of-concept, we present a preliminary implementation of the architecture and apply it to several variants of a simple video game. We show that the resulting system -- though just a prototype -- learns effectively, and, by acquiring a set of symbolic rules that are easily comprehensible to humans, dramatically outperforms a conventional, fully neural DRL system on a stochastic variant of the game.

Kai Arulkumaran, Nat Dilokthanakul, Murray Shanahan et al.

In this paper we combine one method for hierarchical reinforcement learning - the options framework - with deep Q-networks (DQNs) through the use of different "option heads" on the policy network, and a supervisory network for choosing between the different options. We utilise our setup to investigate the effects of architectural constraints in subtasks with positive and negative transfer, across a range of network capacities. We empirically show that our augmented DQN has lower sample complexity when simultaneously learning subtasks with negative transfer, without degrading performance when learning subtasks with positive transfer.

Murray Shanahan

This paper critically assesses the anti-functionalist stance on consciousness adopted by certain advocates of integrated information theory (IIT), a corollary of which is that human-level artificial intelligence implemented on conventional computing hardware is necessarily not conscious. The critique draws on variations of a well-known gradual neuronal replacement thought experiment, as well as bringing out tensions in IIT's treatment of self-knowledge. The aim, though, is neither to reject IIT outright nor to champion functionalism in particular. Rather, it is suggested that both ideas have something to offer a scientific understanding of consciousness, as long as they are not dressed up as solutions to illusory metaphysical problems. As for human-level AI, we must await its development before we can decide whether or not to ascribe consciousness to it.