TABLE 5.
Percentage of test data patterns correctly identified or rejected as “unknown” by each of two criteria for 20 “known” species (on which the network had been trained) and for 14 “novel” speciesa
| Species | % Test data patternsb
|
|||
|---|---|---|---|---|
| Criterion 1
|
Criterion 2
|
|||
| Correctly identified | Rejected as unknown | Correctly identified | Rejected as unknown | |
| Known | ||||
| Alexandrium tamarensis | 84.8 | 14.8 | 87.2 | 12.8 |
| Aphanizomenon sp. | 66.5 | 31.5 | 87.0 | 11.2 |
| Chlorella vulgaris | 94.0 | 2.2 | 93.5 | 4.2 |
| Chroomonas salina | 86.8 | 13.2 | 91.2 | 8.8 |
| Chrysochromulina camella | 89.2 | 3.0 | 86.2 | 7.2 |
| Emiliania huxleyii | 96.2 | 2.5 | 92.2 | 7.0 |
| Gymnodinium simplex | 90.8 | 1.2 | 86.2 | 6.2 |
| Gyrodinium aureolum | 95.0 | 1.8 | 88.2 | 9.5 |
| Halosphaera russellii | 82.0 | 3.0 | 80.5 | 7.2 |
| Heterocapsa triquetra | 92.5 | 3.5 | 87.2 | 10.2 |
| Microcystis aeruginosa | 89.8 | 10.0 | 91.5 | 8.5 |
| Nitschia palea | 87.5 | 11.2 | 83.5 | 15.5 |
| Oscillatoria sp. | 94.2 | 4.5 | 95.5 | 3.0 |
| Phaeocystis globosa | 89.0 | 1.2 | 84.2 | 8.0 |
| Porphyridium pupureum | 89.8 | 10.2 | 93.5 | 6.5 |
| Pyramimonas obovata | 89.5 | 2.0 | 84.5 | 9.2 |
| Skeletonema costatum | 87.8 | 10.0 | 87.2 | 10.5 |
| Staurastrum sp. | 79.5 | 20.5 | 87.8 | 12.2 |
| Pseudopedinella | 67.5 | 2.8 | 62.5 | 9.8 |
| Thalassiosira rotula | 82.8 | 16.5 | 85.5 | 14.0 |
| Avg | 86.76 | 8.28 | 86.76 | 9.08 |
| Novel | ||||
| Anabaena flos-aquae | 97.2 | 87.0 | ||
| Chlamydomonas | 73.8 | 80.8 | ||
| Chlorella salina | 5.0 | 30.0 | ||
| Cryptomonas baltica | 100.0 | 100.0 | ||
| Cryptomonas calceiformis | 100.0 | 100.0 | ||
| Dunaliella tertiolecta | 1.5 | 69.2 | ||
| Ochromonas sp. | 34.8 | 70.8 | ||
| Oscillatoria redeckii | 100.0 | 100.0 | ||
| Prymnesium parvum | 6.5 | 59.8 | ||
| Rhodomonas sp. | 100.0 | 100.0 | ||
| Selenastrum capricornutum | 3.0 | 27.5 | ||
| Scenedesmus quadricauda | 93.5 | 97.0 | ||
| Scenedesmus subspicatum | 11.0 | 26.8 | ||
| Tetraselmis rubens | 12.8 | 54.0 | ||
| Avg | 52.79 | 71.64 | ||
a
The network used 40 non-radially symmetric Gaussian basis functions (width parameter = 1.25), the centers of which were positioned by using the Kohonen LVQ algorithm.
b
Criterion 1, rejection if sum of HLN outputs is <θ, for θ = 0.7; criterion 2, rejection if maximum HLN output is <θ, for θ = 0.4. For comparison, rejection thresholds were selected that gave comparable overall proportions of correctly recognized patterns for the known species.