Perception - Object Recognition – Neural Feature Detectors PDF

Preview

Summary

Recognizing objects ...

Description

Object Recognition – Pandemonium & Neural Feature Detectors Object recognition: Recognising than an object/pattern is a member of a particular category, we must compare a representation of the object/pattern derived from the sensory input with some stored definition or prototype Templates?     



Could do the job – but are not enough Represent simple patterns Template has to match the stimulus for recognition to occur How many templates? o Parsimony – they are infinite Flexibility o Changes in objects e.g. rotations o Occlusions (obstructions) Lettvin, Maturana, McCulloch & Pitts (1959) ‘What the Frog’s eye tells the Frog’s Brain’ o Templates may be okay for a frog but not for us o Encephalisation (the amount of brain mass related to an animal's total body mass)

Hierarchical models (Complex feature analysis) Pandemonium (Selfridge, 1959): Is a data-driven bottom up system recognition model based on feature analysis – objects are recognised from an analysis of their component Pandemonium is composed of four types of recognition units (demons):  





Image demon – e.g. a letter. Record initial image which is formed in the retina Feature demons – recognise little features which become active if the little feature is present in the image. Looks for particular characteristics Cognitive demons – meaningful templates which become active if feature demons are present Decision demon – decide which is the best match for the image. (Find out the best match/alternative rather than having a strict template. Need a good bunch of features – more efficient way of object recognition than templates)

Model requires:    

Feature detectors Different levels of demons Increasing complexity at higher Fewer number of demons at higher levels

Arguments for pandemonium – power: 

With a finite set of feature detectors, pandemonium can recognise a potentially infinite number of objects

Introduction to Cognitive Psychology – Perception: Lectures 5 & 6

   

Will recognise letters regardless to changes in size, orientation and other distortions Kind of mini templates joined together Rely on levels of analysis (like simple templates at low levels) Each level carries out a more complex analysis before a decision is made

Evidence in favour of pandemonium: 

People’s performance in identification studies o Neisser (1964): In letter identification tasks, people will mistakenly call ‘Z’ a ‘V’ or ‘W’, but never an ‘O’ or a ‘G’

Evidence against pandemonium: 

People’s performance in identification studies o Navon (1977): Local to global analysis but… o If a large letter (global) is made from small letters (local) o Identification of large letter unaffected by small latter o Identification of small letter slower if large letter is different

Conclusion:  

Does not account for context effects Top-down or conceptually driven processing: o Processing which is influenced by the context and higher-level knowledge o Intuitive examples: The young or old woman illusion

BUT can neural feature detectors replace the demons in the pandemonium model?  

They seem to increase in complexity at different levels and decrease in number at higher levels like in pandemonium There are different levels in the visual system just like pandemonium

Neural feature detectors levels: 



Retinal ganglion cells o Leave eye through optic nerve o Synapse in the lateral geniculate nuclei (LGN) of thalamus  Synapse in the visual cortex in occipital lobe V1 (primary visual cortex o Synapse with extra-striate regions e.g. V4 (colour), MT (motion)

Neural feature detectors receptive fields: 

The region of the sensory surface that when stimulated causes a change in a neuron’s response e.g. skin, retina, tongue

Neural feature detectors: 

RBCs o o o

Circular receptive field Antagonistic concentric centre surround organisation Is a visual neuron but doesn’t detect light, detects contrast instead  + in the middle where firing rate goes up if light is shined and – on the outside where firing rate goes down if light is shined  Both positive and negative charge cancel one another out  Optimal stimulus is darkness and the contrast of light against dark

Introduction to Cognitive Psychology – Perception: Lectures 5 & 6





Is like a feature demon in the sense that it is a neural feature detector Herman-hering grid LGN have the same receptive field profiles as RGCs Detect black or whit spots & Wiesel (1979) - Simple cells (in V1) In the cortex Rectangular shape Antagonistic centre-surround More complex than RGC & LGN  Selective for orientation o Larger RF than simple cells o Respond to ‘bar’ anywhere in receptive field o Hypercomplex cells in V1: End stopped and corners o Simple, complex and hypercomplex cells detect edges, bars and lines Grandmother cell o Perrett, Rolls & Caan (1982) o Face selective neurons o Detects grandmothers

o o o o Hubel o o o o

4 objections to neural feature detectors:     

De-emphasise the role of context (young and old woman illusion) Object recognition does not depend on list of features of a stimulus – it’s their relationship Still recognise objects when features occluded Global processing preceds local (Navon, 1977) Perception must involve a richer form of description than identifying features – criticises the notion of feature detectors driven by physiology (single cell recordings) – we should be trying to explain vision much more by psychology than physiology o Scene analysis – the nervous system has to arrive at a representation that describes something of the spatiotemporal visual scene

Introduction to Cognitive Psychology – Perception: Lectures 5 & 6...