Questions raised by Seth
Note: The chapter makes use of hypothetical concepts such as quus and grue
which are defined such that they might be mistaken for plus and
green, respectivelty.
1. How are we are to explain evolved agent *misbehaviour*?
e.g., a robot guard dog fails to achieve its task, because it
mistakes a child for a burglar and attacks it
- What counts as misbehaviour, malfunction, misidentification?
- What role did the activity in the guardbot's neural net play?
- Was this role (mis)representational?
- What if it were just a Braitenberg vehicle?
- Would there still be (mis)representation going on inside?
- What (neuron? wire?) is representing what (child?) as what
(burglar?) to whom (dogbot? another neuron?)?
- Could a Beer-style dynamical-systems/cybernetic analysis in terms
of agent-environment dynamics do a better job?
- Might such an analysis lose its ability to distinguish behaviour
from misbehaviour, since it's all just dynamics?
- How does *normativity* apply here?
2. How general purpose can we expect evolved agent mechanisms to be?
How can we determine the function of:
- innate fixed mechanisms
- innate plastic mechanisms (that vary somewhat randomly)
- learned behaviours that remain fixed thereafter
- learned behaviours that are then plastic
- the learning mechanisms themselves
3. How can we understand behavioural novelty?
Can we call the following normal agent behaviour?
- first time any agent ever does A (blink, bark, face now at 11am)
- first time any agent ever does B (scaring opponent by barking)
- first time agent X ever does species-typical behaviour C
- first time agent X ever does culture-specific behaviour D
- first time any agent has learned to consistently do E to get food
- first time any agent has found that injecting F feels good
Kenny Smith: The cultural evolution of linguistic structure
Abstract: Language is a culturally-transmitted system for relating sound and meaning --- individuals acquire their linguistic competence on the basis of the linguistic behaviour of others. This cultural transmission can lead to the emergence of structured language. The Language Evolution and Computation Research Unit has pioneered the application of techniques from machine learning and evolutionary computation to the modelling of this kind of cultural evolution. In this talk I will focus on the cultural evolution of compositionality --- in a compositional system, such as language, the structure of underlying semantic representations is reflected in the structure of externalized signals. I will demonstrate, with the use of a computational model, that the poverty of the stimulus available to language learners leads to the emergence of compositional language, through purely cultural processes. The poverty of the linguistic stimulus forces language to be generalisable, and compositionality is language's adaptation to this pressure.
Paul Vogt: Adaptive grounding of compositional languages
Abstract: This talk will present some novel work on modelling language evolution. The focus of this work is on the emergence of compositional structures grounded through the interactions of a population of communicating agents with their environment. The model is based on previous work of grounding lexicons by robotic agents and recent work done at LEC on iterated learning with respect to the emergence of syntactic structures. I will show how the compositional structures that emerge very much resemble the regularities found in the agents' world.
Slow And Fast Inhibition And H-Current Interact To Create A Theta Rhythm In CA1 In Vitro
Gillies et al. have shown the existence of atropine-resistant theta frequency oscillations in slices of the CA1 hippocampal area in the presence of metabotropic glutamate agonists and total blockade of AMPA receptors. We present a biophysically-inspired mathematical model that successfully reproduced experimental findings. This model focuses on the activity of O-LM (O), cells producing slow IPSPs, and other inhibitory neurons (I), each modeled as a single compartment. In addition to standard Hodgkin-Huxley currents, persistent Na and a hyperpolarization-activated (Ih) current were used for the O cells; blockade of Ih has been shown to destroy the rhythmicity both experimentally and in simulations. We explain by means of numerical and analytical techniques the mechanism by which coherent theta oscillations are created, due to the interaction of the I and O cells via the fast and slow inhibition; we emphasize the effect that I cells exert on O cells due to the presence of Ih. In an attempt to deeply understand the dynamics of the network, we perform a "geometric asymptotic analysis". In particular, we show that for a single O cell the interspike interval may be divided in subintervals inside which the dynamics can be described by lower dimensional systems with slower currents as modulators. We try to exploit this in order to explain the synchronization properties in larger networks. But this is work in progress.... This is a joint work Nancy Kopell (1), M. Gillies (2), M. Whittington (2), Corey Acker (3) and John White (3).
(1)Department of Mathematics and Center for Biodynamics, Boston University.
(2) School of Biomedical Sciences, University of Leeds.
(3) Department of Biomedical Engineering and Center for Biodynamics, Boston University.
Spontaneous activity in the developing nervous system: form and function
Neural activity, driven by visual experience, plays an important role in the development of the mammalian visual system. Some early aspects of visual system development depend on neural activity and yet occur before vision is possible. These early developmental processes are thought to be driven by spontaneously generated patterns of activity. In the first part of the talk, I will summarise some recent experimental findings concerning these spontaneous activity patterns. In the second part, I will show how modelling studies help us understand how this activity might play an instructive role in the Hebbian-based modification of synaptic connections in the nervous system.
BioSystems Reading Group: Discussion Questions.
John Cartlidge.