QUICK SEARCH:   [advanced] Author: Keyword(s):
Home     Help     Feedback     Subscriptions     Archive     Search     Table of Contents

This Article Summary Full Text Full Text (PDF) Alert me when this article is cited Alert me if a correction is posted Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Schuster, S. Articles by Amtsfeld, S. Search for Related Content PubMed PubMed Citation Articles by Schuster, S. Articles by Amtsfeld, S.

Template-matching describes visual pattern-recognition tasks in the weakly electric fish Gnathonemus petersii

Stefan Schuster^* and Silke Amtsfeld

Institut für Biologie I, Hauptstrasse 1, Albert-Ludwigs-Universität Freiburg, D-79104 Freiburg, Germany

View larger version (8K): [in a new window]   Fig. 1. Top view of an experimental tank used to train Gnathonemus petersii to discriminate visual patterns shown at a fixed distance. The fish rests in a daytime shelter whose front end is blocked by a movable transparent Plexiglas screen. Through this screen, the fish views two patterns printed on white cards that are fixed in a pattern holder on the front screen of the tank. The fish therefore views the visual patterns from a fixed vantage point. When the screen is lifted, a trained fish swims straight towards one of the two patterns, where a partition prevents it from seeing the other pattern.

View larger version (9K): [in a new window]   Fig. 2. Effects of ambient light intensity (A) and pattern size (B) on visual pattern discrimination in Gnathonemus petersii. (A) Fish A–C were trained at 9.6 lx (grey bar) to the patterns shown in Fig. 3. The percentage of correct choices in unrewarded tests with the training patterns was determined at the daytime light intensities indicated on the abscissa. Each chosen daytime light level was kept constant for several days until all tests were finished. Testing started each day not earlier than 2 h after light onset in the morning. After a change to a new daytime light level, testing was omitted for one day. Fish A (filled triangles), 100 tests at each of two light levels; fish B (open squares), 200 tests at 9.6 lx, 50 tests each at other light levels; fish C (filled circles), 100 tests at each light level. (B) Tests were performed at standard light intensity (9.6 lx) with fish B (filled and open squares) and fish C (filled circles) in which size-reduced versions of the respective training patterns were shown. Training patterns were as shown in Fig. 3, except for the second series (filled squares) with fish B in which training patterns were as shown in Fig. 6 (top row). Note that, in the first series with fish B (open squares), the two patterns were not of the same size; the abscissa indicates the size (visual angle subtended at the retina) of the larger of the two patterns. Filled circles, 100 tests at each size; open squares, 200 tests at training size, 50 tests at reduced size; filled squares, 100 tests at training size, 50 tests at reduced size.

View larger version (13K): [in a new window]   Fig. 3. Visual pattern discrimination and transfer tests in Gnathonemus petersii. Fish A–C could be successfully trained to discriminate the patterns shown in the top row. In each case, approach towards the disk was rewarded (indicated by ‘+’). To analyze the variables used by the fish to discriminate these patterns, a series of unrewarded transfer tests was conducted (Testing). In these, fish were tested with patterns that differed from the training patterns. These presentations were interspersed with rewarded presentations of the training patterns (at a ratio of approximately 1:3). For each fish, the patterns used in the transfer tests are shown below the respective training patterns, and the absolute numbers of choices are given below each pattern. The scale bar indicates 10° of visual angle.

View larger version (12K): [in a new window]   Fig. 4. A template-matching system may assess the similarity of an actual image to a stored template on the basis of the amount of overlap L (area in light grey), the total area T of the template that remains unmatched (three patches marked dark grey) and the total area M of the actual image that remains unmatched (three patches marked black).

View larger version (9K): [in a new window]   Fig. 5. Outline for a critical test to falsify the template-matching hypothesis. If a fish trained to discriminate a disk from an upright triangle has stored a template of the disk and selects other patterns by matching them with this template, then it could be made to fail in discriminating a disk and a triangle if they are appropriately rescaled. If both are of the same size and both fit within the hypothetical template, then the quantities L, T and M (see Fig. 4) that quantify the quality of the matching are the same for both patterns. Consequently, the fish should fail to discriminate them. This conclusion holds irrespective of the particular way in which L, T and M are combined to assess the match.

View larger version (13K): [in a new window]   Fig. 6. Attempts to falsify the template-matching hypothesis. Fish B and C were both trained to discriminate a disk (rewarded figure, ‘+’) from an upright triangle (‘Training’). After unrewarded tests with the training patterns revealed that both fish had learned the task (‘Testing’; absolute numbers of choices by fish B are given below the respective patterns with the absolute numbers of choices by fish C in parentheses), transfer tests were conducted in which size-reduced versions of the disk and the triangle were shown that should be indistinguishable to a template-matching system. In these tests, both fish chose randomly (numbers of choices given below figures). To analyze whether the loss of preference for the disk was because the size-reduced test patterns were too small to be discriminated, both fish were subsequently trained to discriminate the test patterns, and both successfully learned this discrimination. The sizes of the respective patterns (diameter of disk and side length of triangle) are indicated.

View larger version (13K): [in a new window]   Fig. 7. Outline for (A) and results of (B) a preference-reversal experiment to falsify the template-matching hypothesis. (A) A fish that uses template-matching to discriminate a rewarded disk from an unrewarded triangle should reverse its preference and select the triangle in tests in which the sizes of the disk and triangle are reduced such that the triangle better matches the template (indicated in grey) of the disk that was shown during the training. (B) A corresponding experiment with fish B. The fish was first trained to discriminate a disk (rewarded, ‘+’) from a triangle (unrewarded, ‘–’) of large size, as indicated. After unrewarded tests had revealed that the fish had learned the task (absolute number of choices given below each figure), the fish was tested with two size-reduced versions of the training patterns. Sizes were chosen (i) such that a template-matching system should classify the triangle as more similar to the original disk and (ii) such that the smaller disk would still be large enough to be perceived (see Figs 2B and 6 bottom). Although trained to a circle, the fish preferred the triangle. For definitions of L, M and T, see Fig. 4.

View larger version (10K): [in a new window]   Fig. 8. Discrimination of figures that contain more than one element. (A) Fish A could be trained to discriminate a pattern card with only one centrally placed black disk (rewarded, ‘+’) from a card with two equally sized disks (unrewarded, ‘–’). The numbers of choices of the two pattern cards in 100 unrewarded tests are given below the patterns. In transfer tests in which the single disk was larger in size or placed off-centre, the fish still preferred the card with a single disk. (B) The results of these transfer tests conform with a template-matching mechanism in which the fish (i) treated the double figure as one figure that is laid on the template such that the centres coincide and (ii) quantified the match by the relative overlap L/(L+T+M). The values of L, T and M, to assess the quality of the match (see Fig. 4), are given for both test pairs. Disk diameters are indicated. For definitions of L, M and T, see Fig. 4. A, area of stored template.