Font Size:
Significance of Visual Features for Object Recognition in Pigeons
Last modified: 2011-06-08
Abstract
!Abstract
In my opinion the question whether animals do possess concept or don't is still very valid. This vaguely defined and broadly discussed ability of the mind is considered absent in the animal kingdom. Many researchers consider the attribution to animals inadequate. Maybe concept attribution should only revolve around human subjects. Starting with [1] researchers tried to prove the contradiction to this belief.
It is very important at this point to define the the term “concept.” It can be either a 1) combination of perceptual features that is not satisfactory nor required to identify a object or 2) A perceptual symbol (representation, a link or a word) located in the brain upon which all higher cognitive function work on or 3) A multi-modal non-symbolic and grounded mental model of the perceived object that is useful for either predicting object actions, modal facts or object affordances. At the time of writing I prefer the third definition.
In 1976, Herrnstein successfully trained pigeons to discriminate pictures with humans from non-human images. This was later done with trees, fishes and as for my experiment, snails also. This purely perceptual task solved by pigeons does not prove their ability to acquire concepts while looking at a two-dimensional representation of a real world object. Nor does it prove that their actions are based on higher mental properties - concepts. But we know for sure that pigeons readily learn this task based on some visual features extracted from the image. The issue is that to solve this generalization obligation it might be sufficient to react only to a very simple perceptual vector. The other possible solution to any test is taking a reference back to the learning set based on their perceptual similarity.
In a series of experiments Aust&Huber decided to redo the experiment [1] and shed additional light into the question at stake. The Complementary Information Procedure (CIP) described in [2] trains animals on parts of the present concept, and tests them on its complements. In case of [2] animals were trained to recognize hands of humans and were later tested on their recognition of human heads. This approach minimizes the possibility to generalize solely based on perceptual features (shape, texture, color) and shows that pigeons peck with an increased rate to novel complementary parts of a known concept.
Supported by a suite of empirical studies I am willing to supplement to the promising results done by my colleagues. In the combined experiments with the unpublished work by Grabner (2010) I am going to apply the CIP to a formerly unfamiliar natural concept, namely snails. In the experiment by Grabner subjects were divided into two groups, no-shell and no-head. The ongoing experiment works only with one group of pigeons that are all trained on complete snails. The results are going to provide a baseline for the preliminary experiment. The CIP should not work on unfamiliar concepts because the subjects never have experienced the unseen part. Yet another research question is to extract the main features contained in the positive image that are later-on used by pigeons for generalization. This implies the Gray-Scale (color), Blur (local information and texture) and Scramble (global information, shape) tests.
As already mentioned the CIP only works if the complementary parts are recognized as components of the embracing concept. It is therefore important to see whether the Head, the Skin and the Shell are considered good predictors of correct generalization.
Six subjects are being used in the experiment. Four out of the set are homing pigeons and two are of a local Austrian breed called “Strasser”. All subjects are naive to the task and are experienced with the Go/No-Go Procedure and the skinner box. Subjects are trained on the snail-positive non-snail-negative discrimination task. Images of different snails in many possible environments and view angles are presented to the pigeons. All subjects are assigned to a single experimental group and are trained with the same procedure.
!References
[1]Herrnstein et al, “Natural Concepts in Pigeons”, Journal of Experimental Psychology: Animal Behavior Processes, 1976, 2(4)
[2]Aust&Huber, “Target defining features in a 'people-present / people-absent' discrimination task by pigeons”, Animal Learning & Behavior, 2002, 30 (2), 165-176
[3]Weisman, “Determining When Birds Perceive Correspondence Between Pictures and Objects: A Critique”, Comparative Cognition & Behaviour Reviews, 2010, Vol 5, 117-131
In my opinion the question whether animals do possess concept or don't is still very valid. This vaguely defined and broadly discussed ability of the mind is considered absent in the animal kingdom. Many researchers consider the attribution to animals inadequate. Maybe concept attribution should only revolve around human subjects. Starting with [1] researchers tried to prove the contradiction to this belief.
It is very important at this point to define the the term “concept.” It can be either a 1) combination of perceptual features that is not satisfactory nor required to identify a object or 2) A perceptual symbol (representation, a link or a word) located in the brain upon which all higher cognitive function work on or 3) A multi-modal non-symbolic and grounded mental model of the perceived object that is useful for either predicting object actions, modal facts or object affordances. At the time of writing I prefer the third definition.
In 1976, Herrnstein successfully trained pigeons to discriminate pictures with humans from non-human images. This was later done with trees, fishes and as for my experiment, snails also. This purely perceptual task solved by pigeons does not prove their ability to acquire concepts while looking at a two-dimensional representation of a real world object. Nor does it prove that their actions are based on higher mental properties - concepts. But we know for sure that pigeons readily learn this task based on some visual features extracted from the image. The issue is that to solve this generalization obligation it might be sufficient to react only to a very simple perceptual vector. The other possible solution to any test is taking a reference back to the learning set based on their perceptual similarity.
In a series of experiments Aust&Huber decided to redo the experiment [1] and shed additional light into the question at stake. The Complementary Information Procedure (CIP) described in [2] trains animals on parts of the present concept, and tests them on its complements. In case of [2] animals were trained to recognize hands of humans and were later tested on their recognition of human heads. This approach minimizes the possibility to generalize solely based on perceptual features (shape, texture, color) and shows that pigeons peck with an increased rate to novel complementary parts of a known concept.
Supported by a suite of empirical studies I am willing to supplement to the promising results done by my colleagues. In the combined experiments with the unpublished work by Grabner (2010) I am going to apply the CIP to a formerly unfamiliar natural concept, namely snails. In the experiment by Grabner subjects were divided into two groups, no-shell and no-head. The ongoing experiment works only with one group of pigeons that are all trained on complete snails. The results are going to provide a baseline for the preliminary experiment. The CIP should not work on unfamiliar concepts because the subjects never have experienced the unseen part. Yet another research question is to extract the main features contained in the positive image that are later-on used by pigeons for generalization. This implies the Gray-Scale (color), Blur (local information and texture) and Scramble (global information, shape) tests.
As already mentioned the CIP only works if the complementary parts are recognized as components of the embracing concept. It is therefore important to see whether the Head, the Skin and the Shell are considered good predictors of correct generalization.
Six subjects are being used in the experiment. Four out of the set are homing pigeons and two are of a local Austrian breed called “Strasser”. All subjects are naive to the task and are experienced with the Go/No-Go Procedure and the skinner box. Subjects are trained on the snail-positive non-snail-negative discrimination task. Images of different snails in many possible environments and view angles are presented to the pigeons. All subjects are assigned to a single experimental group and are trained with the same procedure.
!References
[1]Herrnstein et al, “Natural Concepts in Pigeons”, Journal of Experimental Psychology: Animal Behavior Processes, 1976, 2(4)
[2]Aust&Huber, “Target defining features in a 'people-present / people-absent' discrimination task by pigeons”, Animal Learning & Behavior, 2002, 30 (2), 165-176
[3]Weisman, “Determining When Birds Perceive Correspondence Between Pictures and Objects: A Critique”, Comparative Cognition & Behaviour Reviews, 2010, Vol 5, 117-131