Behavior has often been described as the ultimate output of the brain; in neuroscience biological functioning is described in terms of behavior. Since behavior is relevant for every day life of humans, everybody is - if not an expert - autodidactic in this field. For parameters (for instance physiological) for which we have no specialized cortical fields to interpret information we develop special techniques, equipment and algorithms to translate the measurements into terms of biological relevance. For behavior we use our primate brain even in case of other species. Therefore, behavioral definitions often reflect the interpretation and the presumed intention of the behaving individuals and lack the objectivity of other parameters for which special tools have been developed. Despite the efforts of ethologists and behaviorists to introduce clearly defined objective procedures to score and analyze behavior, shortcomings in terms of subjective descriptions and registrations are still present. The distinction between behavior and the contraction of a particular muscle is also vague. As our perception gains sensitivity due to the use of technology, the definition and interpretation of animal activity becomes more complicated. Another limiting factor for behavioral studies is the design of the experimental circumstances during which behavior is observed. Often those circumstances are chosen in such a way that the activity of the animal answers only a particular question. Some tests have been developed to measure predominantly anxiety, cognition, etc. It is tempting to use a one-parameter test which yields clear answers. This has resulted in the registration of numbers of visits to a particular area, time needed to reach a target, distances of simple movements, positions, etc. Moreover, measuring such parameters is easily automated [6, 7].

However, originally behavior is more recognized by changes in shape. Our sensory abilities are better suited to detect changes in shape than quantitative measures such as speed, distances, precise positions, etc.; the combination of both aspects of animal activity should yield the optimal ethogram. Therefore, and to avoid the inter-observer unreliability associated with human observation, an instrument (EthoVision) was developed which makes use of image features describing locomotion and posture as well as the expertise obtained so far in choosing the relevant sets of muscle movements which are known as functional behaviors [1, 4, 5]. EthoVision has the possibility to derive a number of features from series of sequentially digitized images which provide information to discriminate between various behavioral elements. Subsequently, pattern recognition techniques are applied to allow an objective description of functional behaviors based on the features extracted from the image sequences and manually registered behavioral states. We use two approaches for classification: statistical classifiers and neural networks. The statistical approach is described by Van Lochem et al. [2]. In this paper we focus on the use of neural networks.

Neural networks have the ability to process fuzzy data and their output can be modified depending on the input used as examples, a training set of data. We have explored the possibility of a three layer network. The input consisted of features derived from digitized images of the rat’s posture, position, speed, center of gravity, etc. The network was trained by images which had been scored frame-by-frame by experienced observers. Parameters essential for the performance are: nature of the input, i.e. digitized images of one or two camera positions, the number of features extracted from those images, the nature of the network, the reliability of the input, the incorporation of temporal relationships between elements, the architecture of the network. The level of resolution of such a tool does not depend on imaging parameters; sound but also physiological measures can be used as input as well. Especially non-invasively collected parameters such as body temperature may yield additional information about the state of the animal essential for an enhanced distinction of otherwise indistinguishable behaviors. In a pilot study a back propagation network which was trained with the features of repeated frames of digitized images, the reliability in scoring 7 behaviors was superior compared to human observers [3]. Our first results with such a network will be presented and the perspectives will be discussed.

This research is financially supported by Eureka project no. EU88011.

References

1. Lochem, P.B.A. van; Buma, M.O.S. (1998). Video tracking: improved methods for identification of animals with color markers. This volume.
2. Lochem, P.B.A. van; Buma, M.O.S.; Rousseau, J.B.I.; Noldus, L.P.J.J. (1998). Automatic recognition of behavioral patterns of rats using video imaging and statistical classification. This volume.
3. Rousseau, J.B.I.; van Lochem, P.B.A.; Melder, W.; Costa-Florencio, C.; Gispen, W.H.; Spruijt, B.M. (1998). Classification of rat behavior by a neural network. This volume.
4. Rousseau, J.B.I.; Spruijt, B.M.; Gispen, W.H. (1996). Automated observation of multiple individually identified rats. Measuring Behavior ’96 (Utrecht, 16-18 October 1996), 86-87.
5. Smit, J.; Rousseau, J.B.I.; van Lochem, P.B.A.; Plakke, R. (1996). Automatic recognition of behavioral patterns in rodents using digital imaging. Measuring Behavior ’96 (Utrecht, 16-18 October 1996), 95-96.
6. Spruijt, B.M.; Hol, T.; Rousseau, J.B.I. (1992). Approach, avoidance, and contact behavior of individually recognized animals automatically quantified with an imaging technique. Physiology and Behavior, 51, 747-752.
7. Spruijt, B.M.; Rousseau, J.B.I. (1996). Consequences of the ongoing automation of the observation and analysis of animal behaviour. Measuring Behavior ’96 (Utrecht, 16-18 October 1996), 100-101.