The study used eleven pigeons (Columba livia) of undetermined sex, as subjects. The birds were housed in individual cages, in a colony room maintained at 20°C, with a light-dark cycle of 12 h. During data collection, the pigeons were kept at their free-feeding weight, having unrestricted access to a mixture of grit, wheat, sunflower seeds, peas and corn. Water intake was controlled such that access to water was removed for 12 h prior to an experimental session.

All experimental, animal handling and housing procedures were carried out in accordance with the University of Otago’s Code of Ethical Conduct for the Manipulation of Animals and approved by the Animal Ethics Committee of the same entity.

The experiment was performed with head-restrained and body-restrained pigeons. A custom-made metal frame was built to accommodate the fixation of the animal and the administration of water. The apparatus contained a water receptacle that, connected to a pump, filled and emptied automatically (Figure 1C). The pump was controlled using custom code, written in MATLAB, jointly with the OTBR (Otto and Rose, 2023) and Psychophysics toolboxes (Brainard, 1997).

There were also two monitors for stimuli display: they were laterally placed at 74 cm, on either side of the pigeons’ eyes, with an azimuth that centered the images on the optic axis (Nalbach et al., 1990) of the pigeon. The monitors were gamma corrected.

Head-restrained and body-restrained pigeons passively viewed a prepared set of videos that displayed pigeons, or a control shape, performing the following behaviors: courtship, eating, flying and walking. The sequence of videos was pseudorandomized, guaranteeing that the same behavioral category wasn’t displayed consecutively more than three times. Each video was anteceded by a six second inter-trial interval, and was shown six times per session, three times on the left monitor and three times on the right monitor. In total, there were 432 trials, and a session took approximately 75 min to complete. A random 10% of the trials were followed by 2.4 s of access to water (Figure 1B).

The stimulus set of this experiment is composed of videos, recorded in the lab (GoPro Hero 4 session), with durations ranging from 2.2 to 4 s, depicting four different behaviors: courtship, eating, flying and walking (Figure 1). Within each category, four different videos were created to exemplify the same behavior, such that there is some variation in the way it is represented. Although all videos were recorded in the same aviary enclosure, which gave them a very similar background and illumination, they were recorded so that the same action was performed in different places on screen. Occasionally, different camera angles were used.

For each video, we created a corresponding control video in which the background was preserved but the motion that characterizes the targeted behavior was performed by a 2D shape, called a greeble (Gauthier et al., 1998). The greebles were edited to be the same size as the pigeons and were given the dominant color of a background subtracted frame of one of the pigeon videos. That is, the color of the greeble was the average color of the pigeon. All control videos were edited using Adobe After Effects (Adobe Inc., 2023).

Additionally, the stimulus set included four auxiliary videos, showing two pigeons engaging in no specific behavior. The two-pigeon stimulus set was included to disambiguate whether neuronal responses to courtship behavior represented courtship behavior or the presence of two animals on screen.

Furthermore, each video had a version where the natural noise, produced with the recorded behavior, was preserved, and a version in which the video was silent. Hence the developed stimulus set consisted of 72 videos, out of which 36 were visually unique.

An important aspect to consider when comparing the neural responses to stimuli ascribed to different categories is whether the observed differences are driven by low-level stimuli features, such as luminance and spatial frequency (Clark et al., 2022a,b). Therefore, we averaged the luminance values across all frames of all pigeon videos. Equalization was performed separately for the background and the subject. Subsequently, the average values were applied to the control (greeble) videos. The spatial frequency was equalized across all videos (pigeons and controls) but without changing the individual orientation (angle of the light and dark bars, obtained from the fast Fourier transform of the image) of each frame (Figure 2; Willenbockel et al., 2010). We chose this type of equalization to preserve the frame’s natural look and minimize any visual aberrations when they were parsed together to compile the video. Each frame was converted to the HSV color space and the low-level feature transformations were applied to the value layer (Dal Ben, 2023). The correction of luminance and spatial frequency were done using MATLAB (Mathworks Inc., Natick, MA United States) and the SHINE toolbox (Willenbockel et al., 2010, with some code adapted from the SHINE_color toolbox; Dal Ben, 2023). All videos were captured and displayed at 30 frames per second (a subset of the stimuli is provided in Supplementary material).

Lastly, we quantified how visually dissimilar each video was from its control, on a frame-by-frame basis. This way, we created an expectation of which moments of the video could drive an accentuated difference in neuronal responding, purely by the difference the pixel space occupied by the pigeon and the greeble on the image (further details and results in Supplementary material: frame-by-frame correlation).

We performed stereotactic surgery to implant one microdrive, with a probe of eight 25 μm formvar-coated nichrome wires, per animal. For NCL [anteroposterior (AP), 5.5; mediolateral (ML), (±) 7.5], two pigeons (K4, X27) were implanted on the left hemisphere and one (K6) on the right hemisphere. For MVL [AP 10.5; ML (±) 6], two pigeons (K8, JIM) were implanted on the left hemisphere and one (PAM) on the right and. For the visual Wulst [AP 11; ML (±) 3], two pigeons (C1, O4) were implanted on the left hemisphere and three (O9, O8, X15) on right hemisphere. The coordinates for implantation followed histological and inactivation studies (Karten and Hodos, 1967; Kröner and Güntürkün, 1999).

The birds were anesthetized with a mixture of Ketamine (30 mg/kg) and Xylazine (6 mg/kg), administered intramuscularly. A topical anesthetic (10% Xylocaine) was applied to the scalp and an incision was made to expose the skull. Subsequently, a small craniotomy was made over the target structure, and the dura matter was removed. For each bird, a microdrive was installed over one of the target regions: NCL, MVL or Wulst. The tips of the wire bundles were lowered until the beginning of the target region during surgery. To support the implant, stainless steel screws were inserted into the skull and one of them also served as a ground screw. The microdrive was secured to the skull and to the screws with dental acrylic (Jet Acrylic, Lang Dental). Around the microdrive, we added a short metal post, that was used to attach the animal to the frame of the experimental apparatus. In the days following the surgery, the animals were given analgesics and allowed to recover. On the first 3 days of recovery, for pain control, the animals received an intramuscular injection of Carprieve (4 mg/kg) and the wound margin was sprayed with 10% Xylocaine.

We used probes of eight wires to record extracellular single-neuron activity. Before each recording session, the probe was checked for neuronal signals: pairs of two wires were tested to assess which combination of recording wire and local reference yielded the best signal isolation. If the single-to-noise ratio of the isolated unit was at least 2:1, the unit was considered for recording. The subject was not submitted to any task related visual stimulation, during this process.

Sampling at 20,000 Hz, the signal was amplified using a Grass P511K amplifier (Grass Instruments, Quincy, MA, United States) and a 5 Hz notch filter was applied. The signal was digitized using a CED (Cambridge Electronic Design, Cambridge, United Kingdom) interface. Jointly with Spike2 software, spike sorting was done using the system’s template matching capabilities.

The experimental paradigm ran on a separate computer that sent triggers of relevant events to the CED system, which allowed the alignment of the visualization paradigm with the neural data.

After each recording session, the probe was advanced 42 μm and the animal was returned to its home cage. In case it wasn’t possible to isolate a unit that satisfied the inclusion criteria, the probe was advanced 21 μm and the session didn’t take place.

In this experiment, we analyzed single-cell activity in three regions of the pigeon brain (Wulst, MVL, NCL), during the passive observation of videos, depicting pigeons performing different behaviors and their respective controls. We were interested in quantifying neuronal responses, of each one of the regions, to video features shared amongst multiple stimuli. For instance, we wanted to assess neural modulation congruent with the presentation of a video where a specific behavior is performed by a pigeon. It is possible to argue that the features we expect to drive neuronal responses are nested within each video. Considering the nested structure of the stimuli and the repeated measurements of the same neuron, we used several generalized linear mixed models (GLMMs) to assess selectivity, at the single-neuron and population level. When interpreting the models, statistical significance was considered at a level of α = 0.05.

For all neurons, only the trials where the videos were shown on the screen contralateral to the pigeon’s implanted hemisphere were considered for analysis, due to the almost complete decussation of the optic nerve in this species and its impact in the ascending visual projections (Güntürkün and Hahmann, 1999). Therefore, for a given neuron, we recorded its activity to the same visual stimulus six times: three presentations of the video with sound and three without sound. Data analysis was performed in R (R Core Team, 2024) [packages: glmmTMB, DHARMa, (Hartig, 2024) emmeans (Lenth, 2025)] and MATLAB (The MathWorks Inc., 2024).

For each unit, spike counts were binned in 200 ms bins. To determine how the data of a recorded region was distributed, we fitted two null models: one with the poison distribution and the log link function, and another with the negative binomial distribution and log link function (Salinas Ruíz et al., 2023).

The model fitted with the negative binomial distribution had a better fit to the data than the Poisson model (see Results). Subsequently, we used the DHARMa package in R for residual diagnostics and zero-inflation testing. The test results informed the decision to fit all the models used in the study with the negative binomial distribution and the log link function (R package glmmTMB).

At the end of the experiment the pigeons were euthanized with CO₂ by placing them in a 450 mm (Length) × 340 mm (Width) × 300 mm (Depth) chamber, with a CO₂ delivery rate of 16 L/min, an amount that equated to displacing 35% of the chamber volume per minute. The CO₂ was left on for 5 min, and for at least 2 min beyond the point where the animal stopped breathing. The animal was then immediately perfused with 10% formalin in physiological saline. The brains were removed from the skull and kept in 10% formalin for at least 5 days, followed by sucrose formalin (10% formalin, 30% sucrose), and allowed to sink twice. The brains were frozen and sectioned at 40 μm sections with every 10th section mounted and stained with thionin.

All electrode tracks were within the borders of the targeted NCL, MVL, and Wulst regions as defined by Karten and Hodos (1967) and Kröner and Güntürkün (1999). For NCL, the intended track positions were AP +5.5 and ML ± 7.5. The track positions for two NCL birds (K6, X27) were as intended. The track position for K4 was AP +6.0, ML +7.0, differing from the intended AP and ML positions by 0.5 mm.

For MVL, the intended track positions were AP +10.5 and ML ± 6.0. The track positions for two MVL birds (JIM, PAM) were as intended. The position for K8 were AP +10.5, ML ± 5.5, differing only from the intended ML position by 0.5 mm.

For Wulst the intended track positions were AP +11.0 and ML ± 3.0. The track positions for three Wulst birds (C1, O4, X15) were as intended. The position for the O8 and O9 were AP +10.75, ML ± 3.0, differing only from the intended AP positions by 0.25 mm (Figure 3).

In this experiment, we recorded neural activity in three regions of the pigeon brain (Wulst, MVL and NCL) during the passive visualization of videos depicting conspecifics or control shapes performing the following behaviors: courtship, eating, flying and walking. Half of the video presentations, of the pigeon videos as well as the control shapes, included the sound of the pigeon videos.

We recorded and analyzed 103 units in MVL, 123 in NCL and 213 units in Wulst. Data analysis began with the assessment of the data distribution of the recorded neuronal populations. For all regions, the negative binomial distribution had a better fit to the data (lower AIC) than the Poisson distribution (log likelihood ratio test between models for the MVL data: χ²(1) = 22042,p < 0.0001; NCL: χ²(1) < 22853,p < 0.0001; Wulst: χ²(1) = 9808.9,p < 0.0001). Furthermore, the negative binomial null models were not zero-inflated (MVL: ratio = 1.0067, p = 0.79; NCL: ratio = 1.0183, p = 0.38; Wulst: ratio = 1.0014, p = 0.88) and the residuals fitted the assumptions of the distribution.

We first assessed if the spike counts of each neuron were significantly modulated, compared to baseline, for the fixed effects of behavior (courtship, eating, flying and walking, and non-interactive two pigeons), actor (the video depicted a pigeon or a greeble) and the interaction between them. Additionally, we tested the interaction between the presence of sound and the actor on screen (GLMM per neuron: m1, described in Methods).

In MVL, the activity of 65 out of 103 recorded neurons (63%) was task modulated. In this dataset, 15 units (23%) were categorically tuned to the presence of the actor: 7 (11%) units displayed the most amount of modulation when a pigeon was on screen; 5 (8%) were selective for the greebles, and 3 (5%) units had an excited response for the pigeon videos but an inhibited one for the greeble videos. Additionally, 2 of the units that coded for the pigeon were sound modulated: firing rates differed significantly between the trials in which the pigeon videos were displayed with sound on, compared to the same videos without sound, and the greeble videos with and without sound (Figure 4).

Moreover, 5 (8%) units were significantly modulated by one specific behavior, for both pigeons and greebles. 21 units (32%) were modulated by a specific behavior, performed only by one actor. We also found one unit that was selective for more than one behavior and 7 units that were predominately selective for the two pigeon controls. The remaining units were visually selective but not modulated by the content of the videos. Between the behaviorally modulated units and the exclusively visually selective units, 12 units presented a significant interaction between actor and sound and differentiated between the sound on and sound off presentations of the pigeon or greeble videos.

We applied the same categorical analysis (model m1) to the 213 units recorded in Wulst. We found 91 task modulated units. The activity observed was best explained by the presence of a pigeon on screen in 13 units (14%). Other 11 units (12%) fulfilled the criteria of having a significant interaction between the greeble and at least 2 behaviors or exclusively having a significant main effect for the actor greeble. 9 units (10%) were classified as behavior selective. Furthermore, 28 units (31%) were selective for a specific behavior performed only by one actor. We also found 7 (8%) units predominantly modulated by the two-pigeon stimulus set and 3 (3%) that responded to two behaviors. The remaining units had a generalized visual response to all stimulus categories. In this region, there were also 4 (4%) sound modulated units, with one of them being selective for the actor pigeon.

Lastly, we recorded 123 units in NCL. 60 (49%) neurons were task modulated. This dataset had 5 (8%) units in which the presence of a pigeon on screen drove the most amount of neural modulation, and other 5 (8%) units that were modulated by the greebles. We also found 9 (15%) units that encoded one specific behavior for both actors. There were also 3 (5%) units that encoded two behaviors and 3 (5%) other units that encoded two behaviors, each one performed by a different actor. 15 (25%) units encoded a specific behavior performed only by one actor. The remaining units have a significant main effect for both the pigeon and the greeble and no further differentiation amongst the behavioral conditions. That is, they are modulated by the presence of a stimulus in a way that is not specific to the content of the video. They were interpreted as visually selective. Regarding sound modulation, we found 7 (12%) units that were sound modulated, but none belonged to the ones labeled as pigeon selective.

The proportion of units (Figure 5) classified as actor selective did not differ between the three recorded regions [χ²(2) = 2.34, p = 0.3], and the same was the case for the proportion of units classified as behavior selective [χ²(2) = 1.54, p = 0.46]. However, there were differences regarding the proportion of sound modulated units: MVL had a higher proportion of units classified as sound modulated compared to Wulst [χ²(1) = 17.7, adjustedp-value = 0.0001]. No significant differences were found between NCL and MVL [χ²(1) = 4.15, adjustedp-value = 0.08] or, NCL and Wulst (Fisher′sExacttest:adjustedp-value = 0.11).

Although only a small percentage of neurons were considered pigeon selective, we were interested in assessing which moments in the video drove the most amount of modulation. We fitted the subpopulations of pigeon-responsive neurons, grouped by region, with model m2 (see Methods). For every video, the 200 ms bins in which the populational activity significantly differed from the populational baseline were identified. Subsequently, in each of these bins, the activity recorded during the presentation of a pigeon video and the respective greeble control was compared.

In Wulst and MVL, the populations diverged from their respective baselines, for at least a period of 200 ms, in most videos (MVL: 11 out of 16; Wulst: 12 out of 16). In the videos depicting courtship or eating behaviors, every time bin that significantly differed from baseline corresponded to frames where the animal was performing a bowing motion or walking (Figure 6). During flying videos, the significant moments depicted wing flapping and, in one instance for both MVL and Wulst, small body movements while the animal was stationary. In walking videos, all relevant bins corresponded to the animal moving forward (relative to its original trajectory).

In MVL, 2 out of 11 videos had bins where there was a significant difference of activity between the actor pigeon, and the actor greeble. Both instances depicted the animal walking. In Wulst, 9 out of 12 videos contained bins where activity significantly differed from baseline and differed between the pigeon and the greeble presentations of the video.

In NCL, only 5 out of 16 videos had bins where the population activity was significantly modulated from the populational baseline. Furthermore, the identified bins did not differ between the pigeon and the greeble presentations.

Lastly, we analyzed the populational dynamics of the representations, generated by each stimulus, with a population state-space analysis. The aim of this analysis was to identify patterns of activity that differ between the representation of the pigeons and greebles (see Methods for details). The trajectories (Figure 7a) describe instantaneous activation patterns, learned from the data (Pang et al., 2016): they represent the progression of the populational response, during 600 ms of baseline and the entire video presentation, for each condition (video x actor). The trajectories were projections of the first three principal components on the data. The first three principal components contain 17% of the populational variance in MVL, 13% in Wulst and 11% in NCL (Supplementary Figure 2).

In MVL and Wulst, it is possible to observe a destabilization of the population baseline activity, for every behavioral category (Figure 7). Within each category, the videos differ regarding how much separation there is between the trajectories of the actor pigeon and the actor greeble. We quantified the differences in trajectories by measuring the Euclidean distance between them, at every timepoint (Figure 7b). The moments in which the neuronal populations distinguish between the pigeons and the greebles occur in brief periods, not in a sustained way: the distances between trajectories return to values like the ones observed during the baseline after a visible peak.

For each video, we identified the two time points with the largest Euclidian distance between trajectories. During courtship videos, the biggest differences between activity related to pigeons and greebles were found while the pigeon performed bowing motions and walking. During eating videos, it coincided with bowing motions, pecking, mandibulation and walking. In the flying videos, the differences corresponded to two distinct types of behavior. On one hand, moments where the pigeon was flapping its wings and, therefore, creating a lot of movement on screen, but also in moments where the animal was stationary and only performing slight body shifts, such as small head movements. Lastly, in the walking videos, the differences between pigeons and greebles corresponded to the coordinated motion of stepping and head-bobbing. All behaviors were represented in the modulation of the population of every region.