Sheep recognize familiar and unfamiliar human faces from two-dimensional images

Abstract

One of the most important human social skills is the ability to recognize faces. Humans recognize familiar faces easily, and can learn to identify unfamiliar faces from repeatedly presented images. Sheep are social animals that can recognize other sheep as well as familiar humans. Little is known, however, about their holistic face-processing abilities. In this study, we trained eight sheep (Ovis aries) to recognize the faces of four celebrities from photographic portraits displayed on computer screens. After training, the sheep chose the 'learned-familiar' faces rather than the unfamiliar faces significantly above chance. We then tested whether the sheep could recognize the four celebrity faces if they were presented in different perspectives. This ability has previously been shown only in humans. Sheep successfully recognized the four celebrity faces from tilted images. Interestingly, there was a drop in performance with the tilted images (from 79.22 ± 7.5% to 66.5 ± 4.1%) of a magnitude similar to that seen when humans perform this task. Finally, we asked whether sheep could recognize a very familiar handler from photographs. Sheep identified the handler in 71.8 ± 2.3% of the trials without pretraining. Together these data show that sheep have advanced face-recognition abilities, comparable with those of humans and non-human primates.

Keywords: cognitive testing; learning; sheep.

Figure 1.

The operant system. All sheep stay in the waiting pen until brought separately in to the pre-testing area (1). They enter the one-way ambulatory circuit (solid arrow) via the entry corridor (2). The sheep self-activate each trial by passing an infra-red sensor in the corridor that leads to the testing area (3). The visual stimuli are presented on two screens in the testing area. An infra-red sensor above each screen captures the selection behaviour of the animal. Activation of the sensor either initiates the dispensing of a food reward into the feed trough or generates an error signal. After receiving a reward (or error signal), the sheep proceeds through a one-way gate to the beginning of the one-way ambulatory circuit. After the session is completed, the sheep is taken back into the pre-testing area (2). From there, the sheep is released into the resting pen. (Adapted from [39].)

Figure 2.

Examples of images that were used for each of the Training stages 1–3. In all stages, a photograph of one of the four celebrities (Emma Watson shown) is the S+. All faces at this stage are shown front-on. In Training 1, the S− is a black screen; in Training 2, S− is an object randomly selected from a set of 62 objects; in Training 3, S− is an unfamiliar face selected randomly from a set of 36 unfamiliar faces. In all Training stages, side (left or right) of presentation of the S+ was pseudorandomized.

Figure 3.

Examples of photographs shown during the test-probe. In the test-probe, a photograph of a learned-familiar face (in this case, Emma Watson shown in the left-hand column) was presented not only face-on, but also in two new perspectives, tilted to the right or to the left side. The learned-familiar faces (S+) were presented against novel unfamiliar faces (S−) in either front-on or tilted perspective. S− images were matched for sex and ethnicity but not hairstyle. The unfamiliar faces were randomly picked from a pool of 48 images different from those used in the training stages. Note that during the task, the images of the four celebrities were each presented an equal number of times, in pseudorandom order, with side (left or right) also pseudorandomized.

Figure 4.

The effect of training on choice performance for selecting one of the four celebrity faces. At all training stages, S+ is one of the celebrity faces shown. S− was a black screen (Training 1) an object (Training 2) or an unfamiliar face (Training 3). Data are shown as the mean ± s.e.m. Where error bars are not visible, they are obscured by the symbols. Chance (50%) is shown as a dashed line.

Figure 5.

Choice performance for learned-familiar faces of the four celebrities presented in new perspectives, and for the presentation of the face portrait of the handler. Data are shown as mean ± s.e.m. Chance (50%) is shown as dashed line.