Marek Grześ

Théophile Champion, Marek Grześ, Lisa Bonheme et al.

Active inference is a theory of perception, learning and decision making, which can be applied to neuroscience, robotics, and machine learning. Recently, reasearch has been taking place to scale up this framework using Monte-Carlo tree search and deep learning. The goal of this activity is to solve more complicated tasks using deep active inference. First, we review the existing literature, then, we progresively build a deep active inference agent. For two agents, we have experimented with five definitions of the expected free energy and three different action selection strategies. According to our experiments, the models able to solve the dSprites environment are the ones that maximise rewards. Finally, we compare the similarity of the representation learned by the layers of various agents using centered kernel alignment. Importantly, the agent maximising reward and the agent minimising expected free energy learn very similar representations except for the last layer of the critic network (reflecting the difference in learning objective), and the variance layers of the transition and encoder networks. We found that the reward maximising agent is a lot more certain than the agent minimising expected free energy. This is because the agent minimising expected free energy always picks the action down, and does not gather enough data for the other actions. In contrast, the agent maximising reward, keeps on selecting the actions left and right, enabling it to successfully solve the task. The only difference between those two agents is the epistemic value, which aims to make the outputs of the transition and encoder networks as close as possible. Thus, the agent minimising expected free energy picks a single action (down), and becomes an expert at predicting the future when selecting this action. This makes the KL divergence between the output of the transition and encoder networks small.

Fabricio Goes, Zisen Zhou, Piotr Sawicki et al.

This paper presents the Crowd Score, a novel method to assess the funniness of jokes using large language models (LLMs) as AI judges. Our method relies on inducing different personalities into the LLM and aggregating the votes of the AI judges into a single score to rate jokes. We validate the votes using an auditing technique that checks if the explanation for a particular vote is reasonable using the LLM. We tested our methodology on 52 jokes in a crowd of four AI voters with different humour types: affiliative, self-enhancing, aggressive and self-defeating. Our results show that few-shot prompting leads to better results than zero-shot for the voting question. Personality induction showed that aggressive and self-defeating voters are significantly more inclined to find more jokes funny of a set of aggressive/self-defeating jokes than the affiliative and self-enhancing voters. The Crowd Score follows the same trend as human judges by assigning higher scores to jokes that are also considered funnier by human judges. We believe that our methodology could be applied to other creative domains such as story, poetry, slogans, etc. It could both help the adoption of a flexible and accurate standard approach to compare different work in the CC community under a common metric and by minimizing human participation in assessing creative artefacts, it could accelerate the prototyping of creative artefacts and reduce the cost of hiring human participants to rate creative artefacts.

Lisa Bonheme, Marek Grzes

The ability of Variational Autoencoders (VAEs) to learn disentangled representations has made them popular for practical applications. However, their behaviour is not yet fully understood. For example, the questions of when they can provide disentangled representations, or suffer from posterior collapse are still areas of active research. Despite this, there are no layerwise comparisons of the representations learned by VAEs, which would further our understanding of these models. In this paper, we thus look into the internal behaviour of VAEs using representational similarity techniques. Specifically, using the CKA and Procrustes similarities, we found that the encoders' representations are learned long before the decoders', and this behaviour is independent of hyperparameters, learning objectives, and datasets. Moreover, the encoders' representations in all but the mean and variance layers are similar across hyperparameters and learning objectives.

Théophile Champion, Marek Grześ, Howard Bowman

Active inference is a state-of-the-art framework for modelling the brain that explains a wide range of mechanisms such as habit formation, dopaminergic discharge and curiosity. Recently, two versions of branching time active inference (BTAI) based on Monte-Carlo tree search have been developed to handle the exponential (space and time) complexity class that occurs when computing the prior over all possible policies up to the time horizon. However, those two versions of BTAI still suffer from an exponential complexity class w.r.t the number of observed and latent variables being modelled. In the present paper, we resolve this limitation by first allowing the modelling of several observations, each of them having its own likelihood mapping. Similarly, we allow each latent state to have its own transition mapping. The inference algorithm then exploits the factorisation of the likelihood and transition mappings to accelerate the computation of the posterior. Those two optimisations were tested on the dSprites environment in which the metadata of the dSprites dataset was used as input to the model instead of the dSprites images. On this task, $BTAI_{VMP}$ (Champion et al., 2022b,a) was able to solve 96.9\% of the task in 5.1 seconds, and $BTAI_{BF}$ (Champion et al., 2021a) was able to solve 98.6\% of the task in 17.5 seconds. Our new approach ($BTAI_{3MF}$) outperformed both of its predecessors by solving the task completly (100\%) in only 2.559 seconds. Finally, $BTAI_{3MF}$ has been implemented in a flexible and easy to use (python) package, and we developed a graphical user interface to enable the inspection of the model's beliefs, planning process and behaviour.

Lisa Bonheme, Marek Grzes

When training a variational autoencoder (VAE) on a given dataset, determining the optimal number of latent variables is mostly done by grid search: a costly process in terms of computational time and carbon footprint. In this paper, we explore the intrinsic dimension estimation (IDE) of the data and latent representations learned by VAEs. We show that the discrepancies between the IDE of the mean and sampled representations of a VAE after only a few steps of training reveal the presence of passive variables in the latent space, which, in well-behaved VAEs, indicates a superfluous number of dimensions. Using this property, we propose FONDUE: an algorithm which quickly finds the number of latent dimensions after which the mean and sampled representations start to diverge (i.e., when passive variables are introduced), providing a principled method for selecting the number of latent dimensions for VAEs and autoencoders.

Lisa Bonheme, Marek Grzes

Variational autoencoders (VAEs) are used for transfer learning across various research domains such as music generation or medical image analysis. However, there is no principled way to assess before transfer which components to retrain or whether transfer learning is likely to help on a target task. We propose to explore this question through the lens of representational similarity. Specifically, using Centred Kernel Alignment (CKA) to evaluate the similarity of VAEs trained on different datasets, we show that encoders' representations are generic but decoders' specific. Based on these insights, we discuss the implications for selecting which components of a VAE to retrain and propose a method to visually assess whether transfer learning is likely to help on classification tasks.

Peter Clapham, Lisa Bonheme, Marek Grzes

Variational Autoencoders (VAEs) often exhibit a polarised regime in which latent variables separate into active, passive, and mixed subsets. Existing criteria for identifying active dimensions depend on a Gaussian prior, limiting their applicability to variational models and specific priors. We propose a simple information-theoretic classification of the polarised regime based on the entropy of the mean representation. We show theoretically how this entropy couples to KL minimisation through entropy--variance bounds, and we relate the resulting criterion to Bonheme's active/passive conditions. We also clarify a key limitation: entropy of the mean alone cannot reliably distinguish active from mixed dimensions without additional signals from the variance representation. Empirically, we evaluate the entropy criterion on $β$-VAEs, identifiable VAEs, Least-Volume Autoencoders, and L2-regularised autoencoders, and find that it consistently recovers a polarised regime when such a regime is present across the model classes studied. Finally, we show that passive dimensions can yield small but consistent improvements on downstream tasks when latent codes are appropriately normalised, suggesting that collapse is often a matter of scale rather than absolute information removal.

Théophile Champion, Howard Bowman, Dimitrije Marković et al.

Active inference is a leading theory of perception, learning and decision making, which can be applied to neuroscience, robotics, psychology, and machine learning. Active inference is based on the expected free energy, which is mostly justified by the intuitive plausibility of its formulations, e.g., the risk plus ambiguity and information gain / pragmatic value formulations. This paper seek to formalize the problem of deriving these formulations from a single root expected free energy definition, i.e., the unification problem. Then, we study two settings, each one having its own root expected free energy definition. In the first setting, no justification for the expected free energy has been proposed to date, but all the formulations can be recovered from it. However, in this setting, the agent cannot have arbitrary prior preferences over observations. Indeed, only a limited class of prior preferences over observations is compatible with the likelihood mapping of the generative model. In the second setting, a justification of the root expected free energy definition is known, but this setting only accounts for two formulations, i.e., the risk over states plus ambiguity and entropy plus expected energy formulations.

Piotr Sawicki, Marek Grześ, Dan Brown et al.

This study adapts the Consensual Assessment Technique (CAT) for Large Language Models (LLMs), introducing a novel methodology for poetry evaluation. Using a 90-poem dataset with a ground truth based on publication venue, we demonstrate that this approach allows LLMs to significantly surpass the performance of non-expert human judges. Our method, which leverages forced-choice ranking within small, randomized batches, enabled Claude-3-Opus to achieve a Spearman's Rank Correlation of 0.87 with the ground truth, dramatically outperforming the best human non-expert evaluation (SRC = 0.38). The LLM assessments also exhibited high inter-rater reliability, underscoring the methodology's robustness. These findings establish that LLMs, when guided by a comparative framework, can be effective and reliable tools for assessing poetry, paving the way for their broader application in other creative domains.

Théophile Champion, Marek Grześ, Howard Bowman

Model-based reinforcement learning refers to a set of approaches capable of sample-efficient decision making, which create an explicit model of the environment. This model can subsequently be used for learning optimal policies. In this paper, we propose a temporal Gaussian Mixture Model composed of a perception model and a transition model. The perception model extracts discrete (latent) states from continuous observations using a variational Gaussian mixture likelihood. Importantly, our model constantly monitors the collected data searching for new Gaussian components, i.e., the perception model performs a form of structure learning (Smith et al., 2020; Friston et al., 2018; Neacsu et al., 2022) as it learns the number of Gaussian components in the mixture. Additionally, the transition model learns the temporal transition between consecutive time steps by taking advantage of the Dirichlet-categorical conjugacy. Both the perception and transition models are able to forget part of the data points, while integrating the information they provide within the prior, which ensure fast variational inference. Finally, decision making is performed with a variant of Q-learning which is able to learn Q-values from beliefs over states. Empirically, we have demonstrated the model's ability to learn the structure of several mazes: the model discovered the number of states and the transition probabilities between these states. Moreover, using its learned Q-values, the agent was able to successfully navigate from the starting position to the maze's exit.

Piotr Sawicki, Marek Grzes, Fabricio Goes et al.

This study examines the ability of GPT-3.5, GPT-3.5-turbo (ChatGPT) and GPT-4 models to generate poems in the style of specific authors using zero-shot and many-shot prompts (which use the maximum context length of 8192 tokens). We assess the performance of models that are not fine-tuned for generating poetry in the style of specific authors, via automated evaluation. Our findings indicate that without fine-tuning, even when provided with the maximum number of 17 poem examples (8192 tokens) in the prompt, these models do not generate poetry in the desired style.

Théophile Champion, Marek Grześ, Howard Bowman

Branching Time Active Inference (Champion et al., 2021b,a) is a framework proposing to look at planning as a form of Bayesian model expansion. Its root can be found in Active Inference (Friston et al., 2016; Da Costa et al., 2020; Champion et al., 2021c), a neuroscientific framework widely used for brain modelling, as well as in Monte Carlo Tree Search (Browne et al., 2012), a method broadly applied in the Reinforcement Learning literature. Up to now, the inference of the latent variables was carried out by taking advantage of the flexibility offered by Variational Message Passing (Winn and Bishop, 2005), an iterative process that can be understood as sending messages along the edges of a factor graph (Forney, 2001). In this paper, we harness the efficiency of an alternative method for inference called Bayesian Filtering (Fox et al., 2003), which does not require the iteration of the update equations until convergence of the Variational Free Energy. Instead, this scheme alternates between two phases: integration of evidence and prediction of future states. Both of those phases can be performed efficiently and this provides a seventy times speed up over the state-of-the-art.

Théophile Champion, Howard Bowman, Marek Grześ

Active inference is a state-of-the-art framework for modelling the brain that explains a wide range of mechanisms such as habit formation, dopaminergic discharge and curiosity. However, recent implementations suffer from an exponential complexity class when computing the prior over all the possible policies up to the time horizon. Fountas et al (2020) used Monte Carlo tree search to address this problem, leading to very good results in two different tasks. Additionally, Champion et al (2021a) proposed a tree search approach based on (temporal) structure learning. This was enabled by the development of a variational message passing approach to active inference, which enables compositional construction of Bayesian networks for active inference. However, this message passing tree search approach, which we call branching-time active inference (BTAI), has never been tested empirically. In this paper, we present an experimental study of BTAI in the context of a maze solving agent. In this context, we show that both improved prior preferences and deeper search help mitigate the vulnerability to local minima. Then, we compare BTAI to standard active inference (AcI) on a graph navigation task. We show that for small graphs, both BTAI and AcI successfully solve the task. For larger graphs, AcI exhibits an exponential (space) complexity class, making the approach intractable. However, BTAI explores the space of policies more efficiently, successfully scaling to larger graphs. Then, BTAI was compared to the POMCP algorithm on the frozen lake environment. The experiments suggest that BTAI and the POMCP algorithm accumulate a similar amount of reward. Also, we describe when BTAI receives more rewards than the POMCP agent, and when the opposite is true. Finally, we compared BTAI to the approach of Fountas et al (2020) on the dSprites dataset, and we discussed the pros and cons of each approach.

Théophile Champion, Lancelot Da Costa, Howard Bowman et al.

Over the last 10 to 15 years, active inference has helped to explain various brain mechanisms from habit formation to dopaminergic discharge and even modelling curiosity. However, the current implementations suffer from an exponential (space and time) complexity class when computing the prior over all the possible policies up to the time-horizon. Fountas et al (2020) used Monte Carlo tree search to address this problem, leading to impressive results in two different tasks. In this paper, we present an alternative framework that aims to unify tree search and active inference by casting planning as a structure learning problem. Two tree search algorithms are then presented. The first propagates the expected free energy forward in time (i.e., towards the leaves), while the second propagates it backward (i.e., towards the root). Then, we demonstrate that forward and backward propagations are related to active inference and sophisticated inference, respectively, thereby clarifying the differences between those two planning strategies.

Lisa Bonheme, Marek Grzes

The ability of Variational Autoencoders to learn disentangled representations has made them appealing for practical applications. However, their mean representations, which are generally used for downstream tasks, have recently been shown to be more correlated than their sampled counterpart, on which disentanglement is usually measured. In this paper, we refine this observation through the lens of selective posterior collapse, which states that only a subset of the learned representations, the active variables, is encoding useful information while the rest (the passive variables) is discarded. We first extend the existing definition to multiple data examples and show that active variables are equally disentangled in mean and sampled representations. Based on this extension and the pre-trained models from disentanglement lib, we then isolate the passive variables and show that they are responsible for the discrepancies between mean and sampled representations. Specifically, passive variables exhibit high correlation scores with other variables in mean representations while being fully uncorrelated in sampled ones. We thus conclude that despite what their higher correlation might suggest, mean representations are still good candidates for downstream tasks applications. However, it may be beneficial to remove their passive variables, especially when used with models sensitive to correlated features.

Théophile Champion, Marek Grześ, Howard Bowman

Active inference is a state-of-the-art framework in neuroscience that offers a unified theory of brain function. It is also proposed as a framework for planning in AI. Unfortunately, the complex mathematics required to create new models -- can impede application of active inference in neuroscience and AI research. This paper addresses this problem by providing a complete mathematical treatment of the active inference framework -- in discrete time and state spaces -- and the derivation of the update equations for any new model. We leverage the theoretical connection between active inference and variational message passing as describe by John Winn and Christopher M. Bishop in 2005. Since, variational message passing is a well-defined methodology for deriving Bayesian belief update equations, this paper opens the door to advanced generative models for active inference. We show that using a fully factorized variational distribution simplifies the expected free energy -- that furnishes priors over policies -- so that agents seek unambiguous states. Finally, we consider future extensions that support deep tree searches for sequential policy optimisation -- based upon structure learning and belief propagation.

Jack Shannon, Marek Grzes

Neural networks allow Q-learning reinforcement learning agents such as deep Q-networks (DQN) to approximate complex mappings from state spaces to value functions. However, this also brings drawbacks when compared to other function approximators such as tile coding or their generalisations, radial basis functions (RBF) because they introduce instability due to the side effect of globalised updates present in neural networks. This instability does not even vanish in neural networks that do not have any hidden layers. In this paper, we show that simple modifications to the structure of the neural network can improve stability of DQN learning when a multi-layer perceptron is used for function approximation.

Farhana Ferdousi Liza, Marek Grzes

Neural language models do not scale well when the vocabulary is large. Noise-contrastive estimation (NCE) is a sampling-based method that allows for fast learning with large vocabularies. Although NCE has shown promising performance in neural machine translation, it was considered to be an unsuccessful approach for language modelling. A sufficient investigation of the hyperparameters in the NCE-based neural language models was also missing. In this paper, we showed that NCE can be a successful approach in neural language modelling when the hyperparameters of a neural network are tuned appropriately. We introduced the 'search-then-converge' learning rate schedule for NCE and designed a heuristic that specifies how to use this schedule. The impact of the other important hyperparameters, such as the dropout rate and the weight initialisation range, was also demonstrated. We showed that appropriate tuning of NCE-based neural language models outperforms the state-of-the-art single-model methods on a popular benchmark.

Marek Grzes, Jesse Hoey, Shehroz Khan et al.

Assistive systems for persons with cognitive disabilities (e.g. dementia) are difficult to build due to the wide range of different approaches people can take to accomplishing the same task, and the significant uncertainties that arise from both the unpredictability of client's behaviours and from noise in sensor readings. Partially observable Markov decision process (POMDP) models have been used successfully as the reasoning engine behind such assistive systems for small multi-step tasks such as hand washing. POMDP models are a powerful, yet flexible framework for modelling assistance that can deal with uncertainty and utility. Unfortunately, POMDPs usually require a very labour intensive, manual procedure for their definition and construction. Our previous work has described a knowledge driven method for automatically generating POMDP activity recognition and context sensitive prompting systems for complex tasks. We call the resulting POMDP a SNAP (SyNdetic Assistance Process). The spreadsheet-like result of the analysis does not correspond to the POMDP model directly and the translation to a formal POMDP representation is required. To date, this translation had to be performed manually by a trained POMDP expert. In this paper, we formalise and automate this translation process using a probabilistic relational model (PRM) encoded in a relational database. We demonstrate the method by eliciting three assistance tasks from non-experts. We validate the resulting POMDP models using case-based simulations to show that they are reasonable for the domains. We also show a complete case study of a designer specifying one database, including an evaluation in a real-life experiment with a human actor.