Question? Leave a message!




Perception and Memory in Cognitive Systems

Perception and Memory in Cognitive Systems
Intelligent Control and Cognitive Systems Perception and Memory in Cognitive Systems Joanna J. Bryson University of Bath, United KingdomAbout the Course(work) Course was designed as MSc level. • One of its deliverables is engaging you • with research (both reading writing.) Informed public, citizen science. • Exam a concession to finalyear • dissertations. …so, a bit about research ( coursework.) •Time Management and Degree Outcomes Most of you are writing dissertations. 100 hours/course / 12 weeks/course (8.3 hours/week 2 hours/week) 3 week/coursework 19 hours/coursework. 5 writing up 14 hours to hackDoing Research First you need to get your hands dirty. Learn about the problem domain. Check to see whether you have an approach that might work. Publishable projects are normally preceded by pilot projects.What’s Worth Doing At this point, should run with something you have. Normally – look at literature for controversies you might take a side on. Test which side is right. Google Scholar – who cited a paper you’re interested in, and whyWriting Up Research Any paper can have only one point. Point is in the Intro Conclusion. Results prove the conclusion; Approach Discussion explain the Results. Pick a point / claim you think will be most promising to talk about, then examine it in detail..Which Point What surprised you What did you need to learn to get your robot working Is there something your robot does better than the others Could you teach others to do it Look for “tricks” / lessons learned; think about the Brooks contribution.Intelligent Control and Cognitive Systems brings you... Perception and Memory in Cognitive Systems Joanna J. Bryson University of Bath, United KingdomSensing vs Perception First week: Sensing – what information • comes in. This week: Perception – what you think is • going on. Perception includes expectations. • Necessary for disambiguating noisy and • impoverished sensory information.What Would Nature Touch input Do Ear input Eye inputBasic Natural Input CPU Actually whole thing • computes senses. Processing Signal takes time to • (accumulation) propagate, one cell may get two messages from same axon. Many different types • configurations of nerve cells. OutputHow Eyes WorkHow (Vertebrate) Eyes Work Lens focuses world on back of eye. • Rods cones (receptors) respond to light • falling on them. Bipolar cells combine information, detect • edges gradients. Ganglian cells aggregate bipolar cells. •Ganglian cells detect changes in colour, brightness. Send this information to the back of the brain.Excitation, Inhibition Feature Detection c.f. Mann linkInformation Projected and Accumulated Visual cortex has retinotopic maps responding to different edges, motions.Associative Cortices Features • aggregated into objects. Still in maps, e.g. • for pose. Output of a pose sensitive Different systems cell. • for what versus where.Back to Perception Touch input Ear input Eye input➢ ➢ Somatasensory Motor Cortex Cortex ➢Brain Expectations Up until the Thalamus ➞ Retina connection, there are as many axons going towards the eye as away from it.Perception Requires Knowledge Sensing alone is impossible to discriminate. • A whole lot of what the brain does is look • for regularities (cooccurrences), then represent them. Nerve connections positively reinforced • when both sides fire in sequence. The Brain’s Job Pattern Recognition Changing actions / developing skills, • Discovering concepts / categories for • contexts to apply actions, Optimising representations. •Cognitive Architecture: ModularityArchetypical RealTime AI Architecture Perception / action modules (bottom layer) • Reactive (dynamic) plans to arbitrate • between them. (Maybe) planner or at least goal arbitration • at the top. e.g. 3T, PRS, SoarGoal Acquisition (smell) Behaviour Patterns Vertebrate (Fish) Sense/Act Modules Brain as 3Layer Architecture Output (spinal cord)The Obvious Brain Modules are Functional Longterm storage and episodic category memory learning construction (neocortex use / cerebrum) / RAM (hippocampus) routing (thalamus) smoothing (cerebellum) autonomic (midbrain)Modules in Cognitive Systems Many AI cognitive architectures separate • episodic or working memory from long term memory. Few have the process emphasis of brain • e.g. smoothing, category learning, autonomic systems. Exceptions: Ymir (Thórisson 1996). Brooks’ sensing→action modules are • almost ubiquitous in AI.NonModular AI Some researchers try to • solve all AI with a single learning function. Can get amazing skills • together this way. Complete systems (e.g. • driverless cars) need more structure. Atkeson, Schaal students 19972014Modularity in Cognitive Science Fodor (1983) describes two kinds of • modules: Vertical (sensing or motor skills) • Horizontal (crosstask skills like language, • reasoning.) Brooks’ (1986) are sort of super vertical • (sense and act; cf. Minsky’s agents, Society of Mind 1985).Perception in Nature Perception can be seen as both horizontally • and vertically divided. Horizontal: specialist mapping regions. • Vertical: Cone of perceptual processing • leading to single “decision” cell coordinating descending cone of motor activation.Vertical Modules Distributed • across sensory motor pre motor cortex – names are dated (Graziano 2010) Speciestypical • behaviour (again mapped); multi modal stimuli.Cone or Column Individual neurons must be • agnostic, can’t know whether they are the winner while processing. Winning candidates shift • continuously with stimuli, posture. Local competitors inhibit • each other (winnertakeall).Retina again: Bipolar and Ganglian cells not only gather receptor information but locally inhibit.Models of Cortical Maps Built with Localised Excitation and Inhibition Mexican hat function + Winner take all c.f. Willshaw, Hinton.References Mann’s Book (linked, on line). • Carlson, Physiology of Behaviour (many • editions in many years, great text book.) Tom Mitchell, Machine Learning, 1997. • Both in Library •Q: Why put the visual cortex in the rear More timecritical • processing happens in midbrain. V1 also processes • Braille. Light travels faster • than sound: facilitates ‘reptilian’ coincidence detection. vision
Website URL
Comment