Using perceptrons to explore the reorientation task
Section snippets
The reorientation task
The ability to orient and navigate in space is critical for the survival of humans and animals. Studies of navigation in indoor environments have found that humans and other animals can use available external cues to determine direction (Cheng & Newcombe, 2005). Such cues can include the overall shape of the environment (geometric cues), as well as other available landmarks or local elements that might also be placed in the environment (feature cues). Geometric cues are presumed to be
Perceptrons and the reorientation task
An artificial neural network is a system of simple, interconnected processing units that learns to generate a desired response to a presented stimulus by adjusting the weights of the connections between its processors (Bechtel and Abrahamsen, 2002, Dawson, 2004, Dawson, 2005, Rumelhart and McClelland, 1986). Typically, artificial neural networks are composed of three categories of processors: input units, hidden units, and output units. The stimulus is encoded as a pattern of activity in a set
Simulation 1: wall color cues in rectangular arenas
Geometric properties are those that are described by relational properties (Gallistel, 1990). For example, the geometric information that describes Location 4 in Fig. 1 is “a 90° corner with the long wall on the left and the short wall on the right”. However, when geometric cues are all that are available in rectangular arenas like the one in Fig. 1, target locations are ambiguous because more than one location can possess identical geometric properties. In order to uniquely identify a location
Simulation 2: reorientation in kite-shaped arenas
Reorientation has been studied in arenas of various shapes and sizes. In Simulation 2, we consider kite-shaped arenas that can been used (Graham et al., 2006, Pearce et al., 2004, Pearce et al., 2006), in combination with manipulation of wall color features, to explore some of the predictions that emerge from the theory that arena shape is processed by a geometric module (Cheng, 1986, Gallistel, 1990). Fig. 3) illustrates some example kite-shaped arenas that were explored in this simulation.
Simulation 3: distinct objects in corners of rectangular arenas
Another approach to providing feature information in the reorientation task is to provide distinctive objects to distinguish the target location from others (Kelly et al., 1998). This third simulation adopted this paradigm. During training, a three-featured “object” was placed at each of the four arena locations in a rectangular arena. After training, the behavior of the perceptron was observed in additional arenas in which the objects were moved to different locations, or were removed.
Simulation 4: unique, but less salient, feature cues
When objects (often colored and patterned panels) are used to provide feature cues in the reorientation task, the goal is usually to provide a unique visual marker at each arena location. With a perceptron we can easily manipulate the salience of featural cues during training, as is shown in Simulation 4. It is identical to Simulation 3, with the exception that the objects at each location have only one distinctive feature, and share the other two features with every other landmark in the arena.
Simulation 5: moving landmark features
In Simulations 3 and 4, each of the four arena locations in Fig. 1 were associated with objects that were defined by at least one unique feature, and these cues were put in conflict with geometric cues by moving objects in their entirety. One advantage of our modeling approach is that alternative transformations of object features can be performed. For instance, imagine that agents are not processing landmarks as whole objects, but are rather learning about the information carried by local
Simulation 6: effects of arena size with feature cues present
Many researchers have explored the effect of changing arena size on the reorientation task (Chiandetti et al., 2007, Hermer and Spelke, 1994, Learmonth et al., 2002, Learmonth et al., 2001, Learmonth et al., 2008, Ratliff and Newcombe, 2008, Sovrano et al., 2002, Sovrano et al., 2003, Sovrano et al., 2007, Sovrano and Vallortigara, 2006, Vallortigara et al., 2005). A common finding, replicated in studies on many different species, is that when agents are trained in a larger arena, feature cues
General discussion
The current paper explored how a particularly simple type of artificial neural network, the perceptron (Rosenblatt, 1958, Rosenblatt, 1962) could accomplish spatial reorientation. The simulations reported above indicate that perceptrons can generate some of the basic findings associated with the reorientation task (Cheng, 2005, Cheng and Newcombe, 2005): rotational error in rectangular arenas that do not include feature cues, the ability to uniquely identify target locations when featural cues
References (74)
A purely geometric module in the rat’s spatial representation
Cognition
(1986)Whither geometry? Troubles of the geometric module
Trends in Cognitive Sciences
(2008)- et al.
Toddlers’ use of metric information and landmarks to reorient
Journal of Experimental Child Psychology
(2001) The choice axiom after 20 years
Journal of Mathematical Psychology
(1977)Modeling the effects of enclosure size on geometry learning
Behavioural Processes
(2009)Production systems: Models of control structures
Physical symbol systems
Cognitive Science
(1980)Power and limits of reactive agents
Neurocomputing
(2002)- et al.
Modularity and spatial reorientation in a simple mind: encoding of geometric and nongeometric properties of a spatial environment by fish
Cognition
(2002) - et al.
Animals’ use of landmarks and metric information to reorient: effects of the size of the experimental space
Cognition
(2005)
Ants learn geometry and features
Current Biology
Comparing black-capped (Poecile atricapillus) and mountain chickadees (Poecile gambeli): Use of geometric and featural information in a spatial orientation task
Animal Cognition
Connectionism and the mind: Parallel processing, dynamics, and evolution in networks
Vehicles: Explorations in synthetic psychology
Cambrian intelligence. The early history of the new AI
Growing in circles – rearing environment alters spatial navigation in fish
Psychological Science
Spatial cognition and the brain
Year in Cognitive Neuroscience
Reflections on geometry and navigation
Connection Science
Is there a geometric module for spatial orientation? Squaring theory and evidence
Psychonomic Bulletin and Review
Bayesian integration of spatial information
Psychological Bulletin
The information content of panoramic images II: View-based navigation in nonrectangular experimental arenas
Journal of Experimental Psychology-Animal Behavior Processes
Spatial reorientation: The effects of space size on the encoding of landmark and geometry information
Animal Cognition
Being there: putting brain, body, and world together again
Natural-born Cyborgs
Understanding cognitive science
Minds and machines: Connectionism and psychological modeling
Connectionism: A hands-on approach
Connectionism and classical conditioning
Comparative Cognition and Behavior Reviews
Simple artificial networks that match probability and exploit and explore when confronting a multiarmed bandit
Ieee Transactions on Neural Networks
Traditional perceptrons do not produce the overexpectation effect
Neural Information Processing – Letters and Reviews
Interpreting the internal structure of a connectionist model of the balance scale task
Brain and Mind
Neural networks and animal behavior
The modularity of mind
The organization of learning
From conditioning to category learning – An adaptive network model
Journal of Experimental Psychology-General
Gateway to memory: an introduction to neural network modeling of the hippocampus and learning
Spatial learning based on the shape of the environment is influenced by properties of the objects forming the shape
Journal of Experimental Psychology-Animal Behavior Processes
Cited by (29)
Standing on shoulders of a giant: Marcia Spetch's contributions to the study of spatial reorientation
2019, Behavioural ProcessesEnvironment size and the use of feature and geometric cues for reorientation
2013, Acta PsychologicaCitation Excerpt :With respect to orientation tasks, these accounts suggest that the spatial cue(s) with the greatest amount of associative strength will exert the most influence on spatial behavior. Specifically, the correct location is suggested to be composed of independent cues (e.g., object, color, long wall left, short wall right, 90° angle, etc.), and these independent cues all acquire positive associative strength when this location is rewarded during training trials (Dawson et al., 2010; Miller, 2009; Miller & Shettleworth, 2007; Ponticorvo & Miglino, 2010). Above chance performance to the two geometrically equivalent locations, in the absence of the distinct training feature, is suggested to result from equivalent associative strengths of the trained and rotationally equivalent corners because each location contains an identical number of rewarded cues (i.e., long wall left, short wall right, 90° angle).
Potentiation and overshadowing of shape by wall color in a kite-shaped maze using rats in a foraging task
2011, Learning and MotivationCitation Excerpt :One puzzling feature of the results of Experiment 1 (and those of Graham et al. and Pearce et al., 2006) was the failure of the shape-only rats to show evidence on test trials of having learned about the shape of the environment. Using Rescorla and Wagner-based models like those advanced by Miller and Shettleworth (2006) and Dawson et al. (2010), it could be assumed that wall color is always a major factor, even when it is not reliable, as would have been the case for the shape-only group in all these experiments. For shape-only rats each wall color was paired with the correct corner on half the training trials and, as a result, should have gained some associative strength.
A reinforcement learning approach to model interactions between landmarks and geometric cues during spatial learning
2010, Brain ResearchCitation Excerpt :Rather, we show that the proposed model possesses properties that are similar to feature enhancement in the associative model by Miller and Shettleworth (2007), and hence may explain a similar array of data. The advantage of our model compared to other similar models (Miller and Shettleworth, 2007; Miller and Shettleworth, 2008; Dawson et al., 2010) is that geometric cues in the model are encoded implicitly by the locale strategy and so the modeler is not required to explicitly insert into the model such parameters as ‘background cues’, ‘correct/incorrect geometry’, wall lengths, etc. Moreover, the model architecture can be mapped on the biological network implicated in behavior, works in realistic time scale and generates trajectories of simulated animals during learning instead of providing rather abstract predictions of behavior in terms of choice probabilities.
Geometry, Features, and Panoramic Views: Ants in Rectangular Arenas
2011, Journal of Experimental Psychology: Animal Behavior Processes