AV视频

Centre for Cognitive Science (COGS)

Autumn 2023

Autumn 2023

Tuesdays 16:00-17:30

DateSeminarVenue

Oct 17

The Emperor’s New Markov Blankets
Dr Manuel Baltieri
Tokyo/Sussex

Abstract: The free energy principle, an influential framework in computational neuroscience and theoretical neurobiology, starts from the assumption that living systems ensure adaptive exchanges with their environment by minimizing the objective function of variational free energy. Following this premise, it claims to deliver a promising integration of the life sciences. In recent work, Markov Blankets, one of the central constructs of the free energy principle, have been applied to resolve debates central to philosophy (such as demarcating the boundaries of the mind). The aim of this paper is twofold. First, we trace the development of Markov blankets starting from their standard application in Bayesian networks, via variational inference, to their use in the literature on active inference. We then identify a persistent confusion in the literature between the formal use of Markov blankets as an epistemic tool for Bayesian inference, and their novel metaphysical use in the free energy framework to demarcate the physical boundary between an agent and its environment. Consequently, we propose to distinguish between ‘Pearl blankets’ to refer to the original epistemic use of Markov blankets and ‘Friston blankets’ to refer to the new metaphysical construct. Second, we use this distinction to critically assess claims resting on the application of Markov blankets to philosophical problems. We suggest that this literature would do well in differentiating between two different research programs: ‘inference with a model’ and ‘inference within a model’. Only the latter is capable of doing metaphysical work with Markov blankets, but requires additional philosophical premises and cannot be justified by an appeal to the success of the mathematical framework alone.

Fulton Building, FUL-G15

Passcode: 952915

postponed

Computational Cognitive Archaeology
Dr Axel Constant
Sussex

Abstract: Archaeology, as a discipline, offers methods for analysing material remains to understand the activity of humans having produced those remains – the humans of the past. Archaeological reasoning is in turn an attempt at inferring such activity based on those artefacts and material remains. This process of inference is known as archaeological interpretation. How easy it is to make a “sound” archaeological interpretation about a phenomenon will determine where in the “archaeological ladder of inference” that phenomenon is located. At the bottom of the ladder lies techniques (e.g., inferring cooking techniques from remains of pots). At the top of the ladder are religious institutions (e.g., inferring features of spiritual life based on pots). A sound interpretation must be (i) pertinent in the sense of being about remains that are relevant to the phenomenon under investigation (e.g., it is pertinent to try to infer cooking technique from pots, but perhaps not from figurines), (ii) grounded in a theory of how the phenomenon in question occurs in humans (e.g., a theory about nutritional needs of humans), and (iii) warranted by the way one uses data to test the claims that relate to the inference (e.g., testing the capacity of a certain material to retain heat). Cognitive Archaeology is an attempt at performing archaeological inferences about the mind and cognition of humans of the past, which is a phenomenon situated somewhere towards the top of the ladder of inference. How can we make sound inference about ways of perceiving, acting, attending, and learning in humans of the past based on artefacts and decorative pattern? In this talk, I explore how one could use agent-based modelling to ground and warrant archaeological inferences about past minds. More specifically, I focus on the way one could study with active inference attentional features characteristic of a population exposed to prehistoric pots reflecting the main styles and historical social transformation from Neolithic to Roman times.

tba

Passcode: 015571

Oct 31

Creativity in Literature: A Predictive-Processing Approach
Dr Karin Kukonnen
Oslo

Abstract: On the empty page, anything is possible. How do writers then invent a new story or new ways of telling old stories? How do they design probabilities for their readers, when they themselves do not yet know what their text will look like? How do they reinvent themselves from one book to the next? In this talk, I will propose a model for understanding creativity in literary writing, based on the predictive processing framework and 4E cognition. Creativity in literary writing, I propose, needs to be understood as a process that (1) combines spontaneous and deliberate modes of agency (see also Dietrich 2015); that (2) embeds these modes of agency in embodied practices; and that (3) develops the expertise in these practices that writers have accumulated across their career. What distinguishes writing practices that we call ‘creative’ from everyday writing, I argue, lies how authors use them to maximise contingency in the initial phase of the writing process before then translating them into the (illusion of) necessity that form gives to the literary text, namely, what I have called its “probability design” (Kukkonen 2020). Contingency, practice and form are hence the three key terms along which I propose to conceptualise creativity within the predictive processing framework.

Fulton Building - FUL-104

Passcode: 781188

Nov 14

Feedback Instruments and Multistable Musicianship
Dr Chris Keifer
Sussex

Abstract:The musical application of feedback systems in instrument design and performance is one of the oldest practices in electronic music, but it's only recently that we're finding a better understanding of how these instruments work and what their appeal is creatively. The AHRC Feedback Musicianship Network brought musicians and researchers together to develop a research agenda in the field; the presentation will explore some of the main questions that were raised during the network meetings. In particular, it will question how musical instrument design and practice can benefit from conceptualisation within cybernetics and complexity theory. The session will finish with some live demonstrations and opportunities to play feedback instruments.

Sussex Digital Humanities Lab, Silverstone SB211

Passcode: 125298

Nov 28

Investigating AI safety using behavioural economics and behavioural psychology
Dr Steve Phelps
UCL

Abstract:AI Alignment is often presented as an interaction between a single designer and an artificial agent in which the designer attempts to ensure the agent's behavior is consistent with its purpose, and risks arise solely because of conflicts caused by misalignment between the utility function of the designer and the resulting internal utility function of the agent. Particularly with the advent of real-world agents instantiated with large-language models (LLMs), which are pre-trained, we argue this does not capture the essential aspects of AI safety because the real world consists of many designers and many agents (both artificial and human) with heterogeneous values, and therefore AI safety is inherently an aspect of Multi-Agent Systems (MAS) research. We examine two sources of conflict in MAS: principal-Agent Problems, and Social Dilemmas. In the former, conflict arises because of information asymmetry together with inherent misalignment between the utility of the agent and its (many) users, and in the latter because individual agents may have an incentive to disrupt social-level objectives for their own individual advantage. Analysing the response to incentives and strategic behaviour of LLM agents theoretically can be challenging, since we lack prior knowledge of their internal processing, but the same issue also applies to people. In contrast to theoretical economics, behavioural economics performs experiments on human subjects, and has yielded many insights into how real people resolve such conflicts in the real world. We can apply almost exactly the same experimental protocols to studying LLM agents. Taking an empirical approach to AI safety, we investigated the capacity of LLMs to operationalise natural language descriptions of cooperative, competitive, altruistic, and self-interested behavior in social dilemmas. Using a within-subject experimental design, we instantiated LLM-generated agents with various prompts that conveyed different cooperative stances using GPT-3.5 and GPT-4 models. We then assessed the agents' level of cooperation in the Dictator Game and iterated Prisoner's Dilemma and found that the more advanced GPT-4 model demonstrates a better understanding of human normative behavior in these settings. In this talk, I will present these findings, discuss their implications for AI safety, and remaining open questions.

Fulton Building, FUL-G15

Passcode: 987556

Thursday

Dec 14

Naturally Minded: Mental Causation, Virtual Machines, and Maps
Dr Simon Bowes
Sussex

Abstract:Naturally Minded book cover

I will be summarising parts of my recently published book that are relevant to current debates in the philosophy of cognitive science. This book is an empirically informed investigation of the philosophical problem of mental causation, and a philosophical investigation of the status of cognitive scientific generalisations. If there are mental causes which can be classified in a way useful for predicting and explaining, then they are natural kinds. First, we develop an account of natural kinds that accommodates the cognitive. Second, we show how statements using these are not reducible to statements about physical kinds, involving biological and social facts. Finally, Virtual Machine Functionalism is defended as the correct account of the relationship between cognition and the material world.

There is a link to reviews and the e-version of the book available through the library:

Arts A 108

Passcode: 335753

Contact COGS

For suggestions for speakers, contact Simon Bowes

For publicity and questions regarding the website, contact Simon Bowes.

Please mention COGS and COGS seminars to all potentially interested newcomers to the university.

A good way to keep informed about COGS Seminars is to be a member of COGS.  Any member of the university may join COGS and the COGS mailing list.  Please contact Simon Bowes if you would like ot be added.

Follow us on Twitter: