Peter J Bentley | University College London (original) (raw)

Papers by Peter J Bentley

arXiv (Cornell University), Sep 30, 2019

Maintaining the ability to fire sparsely is crucial for information encoding in neural networks. ... more Maintaining the ability to fire sparsely is crucial for information encoding in neural networks. Additionally, spiking homeostasis is vital for spiking neural networks with changing numbers of weights and neurons. We discuss a range of network stabilisation approaches, inspired by homeostatic synaptic plasticity mechanisms reported in the brain. These include weight scaling, and weight change as a function of the network's spiking activity. We tested normalisation of the sum of weights for all neurons, and by neuron type. We examined how this approach affects firing rate and performance on clustering of time-series data in the form of moving geometric shapes. We found that neuron type-specific normalisation is a promising approach for preventing weight drift in spiking neural networks, thus enabling longer training cycles. It can be adapted for networks with architectural plasticity.

arXiv (Cornell University), May 8, 2019

We propose an end-to-end deep learning learning model for graph classification and representation... more We propose an end-to-end deep learning learning model for graph classification and representation learning that is invariant to permutation of the nodes of the input graphs. We address the challenge of learning a fixed size graph representation for graphs of varying dimensions through a differentiable node attention pooling mechanism. In addition to a theoretical proof of its invariance to permutation, we provide empirical evidence demonstrating the statistically significant gain in accuracy when faced with an isomorphic graph classification task given only a small number of training examples. We analyse the effect of four different matrices to facilitate the local message passing mechanism by which graph convolutions are performed vs. a matrix parametrised by a learned parameter pair able to transition smoothly between the former. Finally, we show that our model achieves competitive classification performance with existing techniques on a set of molecule datasets.

2021 International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), Oct 7, 2021

Machine Learning has the potential to discover new correlations between energy usage in apartment... more Machine Learning has the potential to discover new correlations between energy usage in apartments and variables such as seasonality, apartment location, size, efficiency and details of those staying in the apartments, thus helping apartments to become more sustainable and helping those who stay in them to use less energy. The biggest impedance to creating such ML tools is lack of viable data-without the data, the tools cannot be created-yet it is not feasible to wait for several years' worth of good data before creating the tools. Here we present a solution to this problem: the use of a digital twin to generate synthetic data. This approach is viable even when there is no existing data, but when expert knowledge about the relationship between systems exist. To achieve this, we develop a new agent-based synthetic data generator (ASDG) and explore a case study with a corporate housing and luxury alternate accommodation marketplace called TheSqua.re. We show that unlimited quantities of realistic data can be automatically generated, including data for different scenarios, and that it can be used by Machine Learning to discover the underlying correlations.

In this work we present IHDNs: an original model of computation for the simulation of interacting... more In this work we present IHDNs: an original model of computation for the simulation of interacting, dynamic, multi-scale systems. We show that a novel message passing mechanism that operates across layers of abstraction in hierarchical dynamic networks is effective in expressing the complex dependencies of living systems. Using a conventional computational model of cell evolution in cancerous tumour growth for comparison, we demonstrate the validity of IHDNs in emulating the behaviour of lifelike systems, as well as the additional capabilities in enabling Neo4j Cypher patternmatching queries, demonstrated here in the analysis of evolutionary cell heritage.

Fractal proteins are a new evolvable method of mapping genotype to phenotype through a developmen... more Fractal proteins are a new evolvable method of mapping genotype to phenotype through a developmental process, where genes are expressed into proteins comprised of subsets of the Mandelbrot Set. The resulting network of gene and protein interactions can be designed by evolution to produce specific patterns, that in turn can be used to solve problems. This chapter introduces the fractal development algorithm in detail and describes the use of fractal gene regulatory networks for learning a robot path through a series of obstacles. The results indicate the ability of this system to learn regularities in solutions and automatically create and use modules.

Spiking neural networks, thanks to their sensitivity to the timing of the inputs, are a promising... more Spiking neural networks, thanks to their sensitivity to the timing of the inputs, are a promising tool for unsupervised processing of spatio-temporal data. However, they do not perform as well as the traditional machine learning approaches and their real-world applications are still limited. Various supervised and reinforcement learning methods for optimising spiking neural networks have been proposed, but more recently the evolutionary approach regained attention as a tool for training neural networks. Here, we describe a simple evolutionary approach for optimising spiking neural networks. This is the first published use of evolutionary algorithm to develop hyperparameters for fully unsupervised spike-timingdependent learning for pattern clustering using spiking neural networks. Our results show that combining evolution and unsupervised learning leads to faster convergence on the optimal solutions, better stability of fit solutions and higher fitness of the whole population than using each approach separately.

Scientific Reports

Extreme polarization of opinions fuels many of the problems facing our societies today, from issu... more Extreme polarization of opinions fuels many of the problems facing our societies today, from issues on human rights to the environment. Social media provides the vehicle for these opinions and enables the spread of ideas faster than ever before. Previous computational models have suggested that significant external events can induce extreme polarization. We introduce the Social Opinion Amplification Model (SOAM) to investigate an alternative hypothesis: that opinion amplification can result in extreme polarization. SOAM models effects such as sensationalism, hype, or “fake news” as people express amplified versions of their actual opinions, motivated by the desire to gain a greater following. We show for the first time that this simple idea results in extreme polarization, especially when the degree of amplification is small. We further show that such extreme polarization can be prevented by two methods: preventing individuals from amplifying more than five times, or through consist...

arXiv (Cornell University), Jul 4, 2022

Real-world design problems are a messy combination of constraints, objectives, and features. Expl... more Real-world design problems are a messy combination of constraints, objectives, and features. Exploring these problem spaces can be defined as a Multi-Criteria Exploration (MCX) problem, whose goals are to produce a set of diverse solutions with high performance across many objectives, while avoiding low performance across any objectives. Quality-Diversity algorithms produce the needed design variation, but typically consider only a single objective. We present a new ranking, T-DominO, specifically designed to handle multiple objectives in MCX problems. T-DominO ranks individuals relative to other solutions in the archive, favoring individuals with balanced performance over those which excel at a few objectives at the cost of the others. Keeping only a single balanced solution in each MAP-Elites bin maintains the visual accessibility of the archive-a strong asset for design exploration. We illustrate our approach on a set of easily understood benchmarks, and showcase its potential in a many-objective real-world architecture case study.

When someone goes against their own ideas and instead follows the ideas of others, they conform. ... more When someone goes against their own ideas and instead follows the ideas of others, they conform. In this work, we investigate the tendency of different personality types to conform. We use an agent-based model to simulate interactions between people with different personalities and goals. We look at how likely agents are to retain their original goal when they interact with other agents that have different goals. We found there are significant differences in tendency to conform between different personality types.

Proceedings of the International Conference on Health Informatics, 2013

In this paper we describe a methodology for heart sound classification and results obtained at PA... more In this paper we describe a methodology for heart sound classification and results obtained at PASCAL Classifying Heart Sounds Challenge. The results of competing methodologies are shown. The approach has two steps: segmentation and classification of heart sounds. We also describe the data collection procedure.

Fractal proteins are an evolvable method of mapping genotype to phenotype through a developmental... more Fractal proteins are an evolvable method of mapping genotype to phenotype through a developmental process, where genes are expressed into proteins comprised of subsets of the Mandelbrot Set. The resulting network of gene and protein interactions can be designed by evolution to produce specific patterns that in turn can be used to solve problems. In this paper, adaptive developmental programs, capable of developing different solutions in response to different signals from an environment, are investigated. Experiments show that such methods are highly effective in producing robot controllers that generate different movements in response to sensor inputs.

Springer eBooks, 2009

Bio-inspired processes are involved more and more in today's technologies, yet their modelling an... more Bio-inspired processes are involved more and more in today's technologies, yet their modelling and implementation tend to be taken away from their original concept because of the limitations of the classical computation paradigm. To address this, systemic computation (SC), a model of interacting systems with natural characteristics, followed by a modelling platform with a bio-inspired system implementation were introduced. In this paper, we investigate the impact of local knowledge and asynchronous computation: significant natural properties of biological neural networks (NN) and naturally handled by SC. We present here a bio-inspired model of artificial NN, focussing on agent interactions, and show that exploiting these built-in properties, which come for free, enables neural structure flexibility without reducing performance.

Autonomous Robots, Mar 1, 2006

Autonomous adaptation in robots has become recognised as crucial for devices deployed in remote o... more Autonomous adaptation in robots has become recognised as crucial for devices deployed in remote or inhospitable environments. The aim of this work is to investigate autonomous robot adaptation, focussing on damage recovery and adaptation to unknown environments. An embodied evolutionary algorithm is introduced and its capabilities demonstrated with experimental results. This algorithm is shown to be able to control the motion of a robot snake effectively; this same algorithm inherently recovers the snake's motion after damage. Another experiment shows that the algorithm is capable of contorting a shape-changing antenna in such a way as to minimise the affect of background noise on it, thus allowing the antenna to achieve a better signal.

The fundamental challenge faced by any visual system within natural environments is the ambiguity... more The fundamental challenge faced by any visual system within natural environments is the ambiguity caused by the fact that light that falls on the system's sensors conflates multiple attributes of the physical world. Understanding the computational principles by which natural systems overcome this challenge and generate useful behaviour remains the key objective in neuroscience and machine vision research. In this paper we introduce Mosaic World, an artificial life model that maintains the essential characteristics of natural visual ecologies, and which is populated by virtual agents thatthrough 'natural' selection-come to resolve stimulus ambiguity by adapting the functional structure of their visual networks according to the statistical structure of their ecological experience. Mosaic World therefore presents us with an important tool for exploring the computational principles by which vision can overcome stimulus ambiguity and usefully guide behaviour.

Springer eBooks, Sep 27, 2008

As a step towards creating evolutionary developmental neural networks on FPGAs, a bio-inspired ce... more As a step towards creating evolutionary developmental neural networks on FPGAs, a bio-inspired cellular structure suitable for online routing of axons and dendrites on FPGAs based on a new digital spiking neuron model (introduced previously by the authors) is proposed here. This structure is designed to allow changing the routing of the dendrites and axons and formation/elimination of synapses on the fly by dynamic partial reconfiguration of the LUTs. The feasibility and techniques for implementing this structure on a Xilinx Virtex-5 FPGA are also studied.

Lecture Notes in Computer Science, 2003

Robots that can recover from damage did not exist outside science fiction. Here we describe a sel... more Robots that can recover from damage did not exist outside science fiction. Here we describe a self-adaptive snake robot that uses shape memory alloy as muscles and an evolutionary algorithm as a method of adaptive control. Experiments demonstrate that if some of the robot's muscles are deliberately damaged, evolution is able to find new sequences of muscle activations that compensate, thus enabling the robot to recover its ability to move. 2 Background 2.1 Self-Repairing and Shape Memory Alloy Robots Although evolutionary design [1] is common, little work is evident in the field of selfrepairing robotics. Most current research seems to be limited to theory and future

This paper investigates whether replacing non-modular artificial neural network brains of visual ... more This paper investigates whether replacing non-modular artificial neural network brains of visual agents with modular brains improves their ability to solve difficult tasks, specifically, survive in a changing environment. A set of experiments was conducted and confirmed that agents with modular brains are in fact better. Further analysis of the evolved modules characterised their function and determined their mechanism of operation. The results indicate that the greater survival ability is obtained due to functional specialisation of the evolved modules; good agents do well because their modules are more specialised at tasks such as reproduction and consumption. Overall, the more specialised the modules, the fitter the agents.

Lecture Notes in Computer Science, 2005

This paper investigates evolvability of artificial neural networks within an artificial life envi... more This paper investigates evolvability of artificial neural networks within an artificial life environment. Five different structural mutations are investigated, including adaptive evolution, structure duplication, and incremental changes. The total evolvability indicator, E total , and the evolvability function through time, are calculated in each instance, in addition to other functional attributes of the system. The results indicate that incremental modifications to networks, and incorporating an adaptive element into the evolution process itself, significantly increases neural network evolvability within open-ended artificial life simulations.