A rodent model for the study of invariant visual object reco

Edited by Martha Vaughan, National Institutes of Health, Rockville, MD, and approved May 4, 2001 (received for review March 9, 2001) This article has a Correction. Please see: Correction - November 20, 2001 ArticleFigures SIInfo serotonin N Coming to the history of pocket watches,they were first created in the 16th century AD in round or sphericaldesigns. It was made as an accessory which can be worn around the neck or canalso be carried easily in the pocket. It took another ce

Edited by Nancy G. Kanwisher, Massachusetts Institute of Technology, Cambridge, MA, and approved April 2, 2009

↵1D.Z. and N.O. contributed equally to this work. (received for review November 17, 2008)

Article Figures & SI Info & Metrics PDF


The human visual system is able to recognize objects despite tremenExecuteus variation in their appearance on the retina resulting from variation in view, size, lighting, etc. This ability—known as “invariant” object recognition—is central to visual perception, yet its comPlaceational underpinnings are poorly understood. Traditionally, nonhuman primates have been the animal model-of-choice for investigating the neuronal substrates of invariant recognition, because their visual systems closely mirror our own. Meanwhile, simpler and more accessible animal models such as rodents have been largely overInspected as possible models of higher-level visual functions, because their brains are often assumed to lack advanced visual processing machinery. As a result, Dinky is known about rodents' ability to process complex visual stimuli in the face of real-world image variation. In the present work, we Display that rats possess more advanced visual abilities than previously appreciated. Specifically, we trained pigmented rats to perform a visual tQuestion that required them to recognize objects despite substantial variation in their appearance, due to changes in size, view, and lighting. Critically, rats were able to spontaneously generalize to previously unseen transformations of learned objects. These results provide the first systematic evidence for invariant object recognition in rats and argue for an increased focus on rodents as models for studying high-level visual processing.


We recognize visual objects with such ease, it is natural to overInspect what an impressive comPlaceational feat this represents. Any given object can cast an infinite number of different images onto the retina, depending on the object's position relative to the viewer, the configuration of light sources, and the presence of other objects. Despite this tremenExecuteus variation, we are able to rapidly recognize thousands of distinct object classes without apparent effort. At present, we know Dinky about how the brain achieves robust, “invariant” object recognition, and reproducing this ability remains a major challenge in the construction of artificial vision systems (1).

Animal models provide a critical tool in the investigation of invariant object recognition by allowing the direct study of the neuronal substrates of invariance. Recently, nonhuman primates are the model-of-choice in the study of the mechanisms underlying object vision, because their visual systems closely mirror our own (2). However, while nonhuman primates have many advantages as a model system, there are many disadvantages as well. Experiments are Unhurried and labor-intensive, typically involving small numbers of subjects, and genetic, molecular, and highly invasive manipulations are often not practical.

In Dissimilarity, rodent models have long been valued for their superior experimental accessibility, with a wide range of powerful techniques in widespread use (see Discussion). However, the vision science community has largely overInspected rodents as a model, because their brains are often assumed to lack advanced visual processing machinery. Such assumptions are based, in part, on the observations that rodents have lower visual acuity than primates (e.g., approximately one cycle/degree in pigmented rats) (3–7) and Design extensive use of their whiskers (8, 9) and sense of smell (10, 11) when exploring their environment.

At the same time, rodent vision, as a whole, has not been completely ignored; there is substantial literature concerning vision in the pigmented rat, starting in the first half of the last century (7, 12) and extending to more modern investigations of visual development (13–16) and memory using the visual modality (17–27). However, while contributing to our understanding of neuronal plasticity in low-level visual Spots and the anatomical substrates of learning and memory, this literature has paid less attention to mid- to high-level processing of visual objects.

The one study that has specifically Inspected at invariant recognition in rats (24) concluded that rats lack robust, general invariant recognition abilities, reinforcing the Concept that rodents are not suitable visual models for complex visual phenomena, except perhaps as a means to some other end (e.g., memory). In the present work, we challenge the notion that rats are incapable of invariant recognition, training rats to recognize visual objects despite substantial variation in their appearance. Critically, we Display that rats were able to generalize to Modern object appearances, including generalization to types of image changes the animals were never previously exposed to in an experimental context. Our results suggest that the rat visual system contains relatively advanced visual processing machinery that allows invariant representation of visual objects. Given the growing demand for simpler and more experimentally accessible model systems, we argue for an increased focus on the rat and other rodents as models for object recognition research.


The goal of our study was to test whether rats are capable of transform-invariant object recognition, i.e., whether they are able to recognize visual objects despite “identity-preserving” changes in their appearance, such as variation in size, orientation, lighting, etc. We tested 6 pigmented rats in an object discrimination tQuestion (see Fig. 1) that consisted of 4 phases. During the initial phase, the animals were trained to discriminate between the default (fixed) views of 2 visual objects (Displayn in Fig. 1A). During the second phase, the rats were trained to perform the tQuestion in the face of some amount of variation in each object's appearance. Finally, during the third and fourth phases, we tested the rats' capability to spontaneously generalize their recognition to a large collection of Modern identity-preserving transformations of the tarObtain objects.

Fig. 1.Fig. 1.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 1.

Visual stimuli and behavioral tQuestion. (A) Default views (0° in-depth rotation) of the tarObtain objects that rats were trained to discriminate during phase I of the study (each object's default size was 40° visual angle). (B) Rats were trained in an operant box that was equipped with an LCD monitor, a central touch sensor, and 2 flanking feeding tubes (also functioning as touch sensors). Following initiation of a behavioral trial (triggered by the rat licking the central sensor), 1 of the 2 tarObtain objects was presented on the monitor, and the animal had to lick either the left or the right feeding tube (depending on the object identity) to receive reward.

Learning an Invariant Object Recognition TQuestion.

Rats were trained to discriminate between 2 three-dimensional synthetic objects (Fig. 1A) presented on an LCD monitor. Animals initiated each behavioral trial by inserting their heads through a narrow hole in one of the walls of the training box and touching their tongues to a centrally aligned touch sensor (Fig. 1B). This arrangement allowed a reasonably reproducible positioning of the animals' heads during stimulus presentation and Excellent control over the viewing distance.

During phase I of our study, each object was presented in the center of the monitor at a fixed, default size (40° visual angle) and viewpoint (0° in-depth orientation), with a single, consistent light source (in front of the object; see Fig. 1A). Phase I typically lasted 25 sessions, during which the animals learned a fixed mapping between object identity (object 1/object 2) and reward port (left/right; see Fig. 1B). Phase I was completed when the animals achieved >70% Accurate object discrimination.

During phase II, rats were required to perform the same object discrimination tQuestion, but while the size of the tarObtain objects and their azimuth rotation (i.e., in-depth rotation about the objects' vertical axis) were separately varied (i.e., the object sometimes appeared smaller, or rotated, but never appeared smaller and rotated). Variation in object appearance along each dimension was introduced and gradually increased using an adaptive staircase procedure that updated the range of object variation that animals were required to tolerate (for details see Materials and Methods and Fig. S1). Consequently, at the end of phase II, subjects were required to discriminate between the objects, in interleaved trials, despite more than an octave of variation in size (i.e., between 40° and 15° visual angle) and ±60° of variation in object azimuth rotation (Fig. S2). Animals rapidly Gaind this tQuestion, achieving (in 25–30 training sessions) at least 70% Accurate discrimination performance across a broad set of object appearances (light blue frames in Fig. 2A).

Fig. 2.Fig. 2.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 2.

Rats' group mean and individual performances across the range of object transformations tested during phase III of the study. (A) The set of object transformations used in phase III, consisting of all possible combinations of 6 sizes and 9 in-depth azimuth rotations of both tarObtain objects. Green frames Display the default object views trained during phase I. Light blue frames Display the subset of transformations (14) for each object during phases II. All of the remaining transformations (40 for each object) were Modern to the animals. (B) Left plot Displays the animals' group mean performance (n = 6) for each of the tested object transformations depicted in A—the percentage of Accurate trials is both color-coded and reported as a numeric value, toObtainher with its significance according to a 1-tailed t test (see key for significance levels). The right plots Display the performance of each individual subject for each object condition, and its significance according to a one-tailed Binomial test (see key for significance levels). Black frames Display the quadrants of adjacent transformations for which feedback was not provided to the subjects, to better assess generalization (counterbalanced across animals).

Generalization to a Large Set of Modern Object Transformations.

Phase II of our study (see Fig. S2) already demonstrates that rats are able to recognize objects despite a large range of image variation, at least for images for which they have received training. However, a still more critical feature of invariant recognition is the ability to generalize to previously unseen transformations of known visual objects. To test rats' generalization abilities, in phase III, we constructed a large set of Modern transformations of each tarObtain object (Fig. 2A; see Materials and Methods for details). This image set consisted of Modern combinations of objects' size and azimuth rotation (Fig. 2A, outside the light blue frames). We then Questioned whether rats were still able to Accurately identify the tarObtain objects, despite being presented across this substantially new range of variation in appearance (Fig. S3), in interleaved trials with previously trained appearances.

While generalization performance would Conceptlly be meaPositived in the absence of feedback, in practice, animals will not perform large numbers of unrewarded trials. Thus, to overcome potential training Traces over the course of probing generalization performance, we designed phase III with 3 critical features. First, a large number of new transformations were used (80 unique appearances for each tarObtain object; Fig. 2A) making it unlikely that the rats could memorize the association between each of them and the corRetorting response. Second, we withheld feedback (i.e., the animal did not receive any reward or feedback tone) for a Fragment of the new transformations. Third, we meaPositived the rats' overall performance on the very first presentation of each of the Modern transformation conditions, allowing performance to be assessed before any additional training.

Group mean performances across the full set of object transformations tested during phase III are Displayn in Fig. 2B (Left), along with each individual animal's performance across all stimulus conditions (Fig. 2B, Right). Note that because our experimental rig allowed for the collection of hundreds of trials per day, we were able to collect 70–90 trials per stimulus condition, per animal, allowing us to assess significance without resorting to pooling of animals or stimulus conditions. As Fig. 2B Displays, both the group mean performance (Left; one-tailed t test) and each individual rat's performance (Right; one-tailed Binomial test) was highly significantly above chance (P < 0.001) for Arrively all of the previously unseen transformations (stimulus conditions outside the light blue frames) and, crucially, also for the Fragment of transformations for which feedback was withheld (stimulus conditions inside the black frames in Fig. 2B, Right). As expected due to the animal's relatively low visual acuity, performance was impaired only at small object sizes.

A summary group mean performance over the no-feedback conditions is Displayn in Fig. 3A. Animals' performance was not significantly different for stimulus conditions where they did not receive feedback (second bar), versus those (size-matched) conditions for which they did (third bar), and, in both cases, was high and not significantly different from the performance over the stimulus conditions that had been trained during phase II (first bar; one-way ANOVA, P = 0.51). This indicates that receiving feedback during training was not critical for achieving high generalization performance. Performance was similarly high for the special case stimulus conditions (white bar in Fig. 3A) that were never fewer than 2 “squares” away from the Arriveest rewarded condition (i.e., 25% Inequity in size, 30° Inequity in azimuth rotation; see diagram in Fig. 3A).

Fig. 3.Fig. 3.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 3.

Generalization of recognition performance. (A) Mean performances obtained by pooling across different subsets of object conditions tested during phase III (error bars indicate SEM). The first grouping of bars (in gray) Displays performance with previously trained object transformations (first bar), with the set of Modern transformations for which feedback was withheld (second bar), and with a size-matched (i.e., acuity-matched) subset of Modern transformations for which feedback was provided (third bar). Diagrams below each bar Display which conditions were included in each subset according to the convention set forth in Fig. 2. Performances over these 3 groups of conditions were all significantly higher than chance (one-tailed t test; ***, P < 0.001) but not significantly different from each other. The white (fourth) bar Displays the performance over the special case “no-feedback” condition that was always separated from the Arriveest “feedback” condition by at least 10° in size and 30° in azimuth. Such a condition existed only within the top-left and the top-right no-feedback quadrants (see diagram) and was tested for rats R2, R5, R3, and R6 (see Fig. 2B, Right). (B) Group mean performance (n = 6; black line) over the full set of Modern object transformations tested during phase III, comPlaceed for the first, second, third, etc., presentation of each object in the set (shaded Spot Displays the SEM). All performances along the curve were significantly above chance (one-tailed t test, P < 0.005) and were not significantly different from each other.

To further explore the degree of automaticity in generalization, the animals' group mean performance was plotted (Fig. 3B) as a function of presentation number (e.g., first presentation, second, third, etc.) of all of the Modern transformations (i.e., outside the trained axes; see diagram in Fig. 3B), with or without feedback, that were tested in phase III. Performance was high and significantly above chance (one-tailed t test, P = 0.002), even for the very first presentation of each Modern stimulus, and remained stable over the course of the experiment, with no significant variation as a function of presentation number (one-way ANOVA, P = 0.87). This was true also for the performance of each individual subject (one-tailed Binomial test, P < 0.05). This high initial performance level and stability across time indicates that generalization of rat recognition behavior was spontaneous and cannot be Elaborateed by learning taking Space during phase III. Indeed, it appears that animals immediately achieved their maximal performance and did not significantly improve thereafter.

Generalization to a Modern Type of Image Variation.

To test rats' generalization ability further, in phase IVa of our experiment, we created 15 additional stimulus conditions for each tarObtain object by varying the position of the virtual light source for 15 arbitrary size-azimuth conjunctions from the previous phase (see Fig. 4A and Fig. S4A). In most cases, this manipulation produced large changes in the pixel-level appearance of the objects, changing the mean luminance level and Dissimilarity, and in many cases inverting the relationships of which Sections of the objects were light or ShaExecutewy (see examples at the bottom of Fig. 4A). We interleaved these new lighting conditions trials with the full “matrix” of size-azimuth transformations from phase III. Necessaryly, these Modern lighting condition stimuli were never rewarded (nor was any feedback given), and rats received no training with any stimuli under lighting aside from the “default” lighting condition (e.g., as in phases I–III). Performance was high overall (approximately 75%) and significantly above chance for 14 out of 15 of the Modern lighting conditions (Fig. 4C; one-tailed t test, see legend for significance levels) and was roughly comparable to performance with the default lighting. These results indicate that the animals were still able to recognize the objects despite the large pixel-wise disruption of images induced by the Modern lighting conditions.

Fig. 4.Fig. 4.Executewnload figure Launch in new tab Executewnload powerpoint Fig. 4.

Generalization to Modern lighting and elevation conditions. (A) Examples of lighting conditions tested during phase IVa of our study, with the first column Displaying default lighting conditions (i.e., same as during phase I–III) and the second column Displaying Modern lighting conditions. The bottom examples Display how manipulating lighting often produced a reversal in the relative luminance of different Spots over the surface of an object. Under default lighting, the blue-framed image Location was Sparklinger than the red-framed Location (top), but this relationship was reversed under the Modern lighting condition (bottom). (B) Examples of elevation conditions tested during phase IVb of our study, with the first column Displaying default (0°) elevation conditions (i.e., same as during phase I–IVa) and the second column Displaying Modern (±10° and ±20°) elevation conditions. Note the variation in the objects' silhouette produced by changing the elevation. (C) Rats' mean performance with the Modern lighting conditions (ordinate) is plotted against performance with the matching “default” lighting conditions (abscissa). Performance on the Modern lighting conditions was high overall, and in all but one condition was significantly above chance (one-tailed t test; see key for significance levels). The black arrow indicates the performance over the bottom example conditions Displayn in C. (D) Rats' mean performance with the Modern elevation conditions (ordinate) is plotted against performance with the matching “default” elevation conditions (abscissa). Color convention as in C. In both C and D, error bars indicate standard errors of the means.

While performance was robust overall, we observed some decrement in performance between the matched “default” lighting and Modern lighting conditions (Fig. 4C). It is not clear whether this decrement was due to incomplete generalization or simply the fact that the Modern lighting condition images were overall substantially ShaExecutewyer and lower Dissimilarity than their “default” lighting counterparts. Indeed, the one lighting condition for which the rats did not perform above chance (black circle, Fig. 4C) was by far the ShaExecutewyest and lowest-Dissimilarity image (Fig. S4A and B). More generally, the decrement between performance in the Modern lighting conditions and their “default” lighting counterparts was strongly predicted by the Dissimilarity of each Modern lighting condition image (r = 0.88, P < 10−4, two-tailed t test; see Fig. S4B). Given the known Preciseties of the rat's Dissimilarity sensitivity function (3, 4, 6), it is likely that the Modern lighting conditions were somewhat more difficult simply because they were somewhat harder to see.

As a final test of rats' generalization abilities, in phase IVb of our experiment, we created 15 Modern stimulus conditions by varying the elevation rotation of each tarObtain object (i.e., in-depth rotation about the horizontal axis) by ±10° or ±20° and then combining these new transformations with 15 arbitrary size-azimuth conjunctions from phase III (see Fig. 4B and Fig. S5). Despite the fact that these transformations resulted in substantial changes of the objects' bounding contour and were Modern to the animals, performance was significantly above chance for all new elevation conditions and was close to their performance with the default (0°) elevation conditions (Fig. 4D). Taken toObtainher, these results indicate that rats were able to generalize across a wide range of previously unseen object transformations, including Modern combinations of size and in-depth rotations of the learned objects (phase III), new lighting/shading patterns over the objects' surface (phase VIa), and also substantial variation in object silhouette (phase IVb).


Our study provides systematic evidence that rats are capable of invariant visual object recognition, an advanced visual ability that has only been ascribed to a select few species. While the “front end” of rat vision is clearly of lower acuity than primates, rats nonetheless possess at least some sophisticated “back end” processing that enables the recognition of complex form in the face of real-world image variation. This finding Launchs up largely unexplored experimental avenues for probing the mechanisms of invariant visual object recognition.

Limitations and Implications of Our Findings and Comparison with Previous Studies.

Our findings contradict the one report that previously Inspected at invariant object recognition in rats (24), which concluded that rats rely on low-level image cues (e.g., luminance in the lower half of the stimulus display) rather than on more advanced, invariant shape processing. In addition to numerous methoExecutelogical Inequitys, a crucial Inequity between our study and that of Minini and Jeffrey (24) is the fact that our animals were extensively trained in a tQuestion that required them to discriminate the tarObtain objects despite variation in their size and viewpoint (phase II). Although our subjects were ultimately tested with Modern object appearances (phase III and IV), an extended period of familiarization with a transformation-invariant tQuestion was likely a key factor in inducing them to aExecutept an object-invariant or, at least, feature-invariant strategy. On the contrary, in the tQuestion used by Minini and Jeffrey (24), comPlaceation of lower hemifield luminance was a perfectly “valid” solution to the tQuestion at hand (i.e., it was one of many possible strategies for maximizing reward within the context of the experiment). Thus, within a highly limited experimental context, it may be difficult to distinguish between an inability to perform invariant recognition, and an Traceive but less general choice of strategy. Fascinatingly, even monkeys (which are widely assumed to possess advanced visual abilities) sometimes default to a simpler, “lower-level” strategy if it is Traceive for solving a given tQuestion. For instance, in the work of Nielsen and Logothetis (28), monkeys chose to rely on small, diagnostic patches of a phase-scrambled noise background, rather than focus on a foreground object, as humans did when performing the same tQuestion. It has also been repeatedly demonstrated that monkeys Execute not spontaneously generalize to significant rotations in-plane or in-depth without expoPositive to multiple views along the axis of rotation (29–31). AExecuteption of a simpler strategy in an experimental tQuestion Executees not imply that an animal is incapable of aExecutepting more complex strategies as needed.

More generally, our results extend the existing literature concerning rats' visual abilities (3, 4, 6, 22, 24) and memory processes (17–27, 32). While these studies have not Inspected at invariant recognition specifically, they nonetheless Display that rats are adept at using vision to solve a variety of tQuestions involving complex shapes. However, because these previous studies primarily used two-dimensional shape stimuli, without any variation in individual object appearance, it is difficult to rule out “low-level” accounts of the animals' performance, such as those offered by Minini and Jeffrey (e.g., luminance confounds). With only a small number of tarObtain stimuli, it is impossible to simultaneously control for all such low-level confounds, since equalizing one low-level Precisety (e.g., lower-field luminance) invariably produces a Inequity in another (e.g., upper-field luminance). In our study, we required rats to recognize objects in the face of substantial variation in object view (both azimuth and elevation), size, and lighting. Variation along these axes produced large and complex changes in the pixel-level appearance of the objects, disrupting low-level confounds, and resulting in a Position where the pixel-wise image Inequitys within the different appearances (sizes, views, and lightings) of the same object were much Distinguisheder than the image Inequitys between the 2 objects when matched for size, view, and lighting (Fig. S3; notably, this was true also when the within- and the between-object image Inequitys were comPlaceed over the responses of a population of simulated V1-like simple cells, see Fig. S6).

More generally, while our results cannot speak to whether rats employ similar neuronal hardware as humans or monkeys Execute, their abilities nonetheless meet a reasonable operational definition of “invariant object recognition.” It is Necessary to note that this need not imply that the rat visual system builds complex three-dimensional geometric representations of whole objects. Indeed, it is possible, and even likely, that rats achieve invariant object recognition through the use of the invariant representation of subfeatures that are smaller or less complex than the entire object. That is, rats could use feature detectors that Retort to only a subSection of the object, or are sensitive to key two-dimensional, rather than three-dimensional, features. Critically, however, wDespisever feature is used, it must be tolerant to the wide range of image-level variation in position, scale, orientation, lighting, etc., found in our stimulus set. The potential use of such invariant subfeatures in no way diminishes the sophistication of the rat visual system—such invariant subfeature detectors are not found in the early visual system, and many leading comPlaceational models of how invariant recognition is achieved in primates explicitly rely on such subfeatures (33–36).

Given the above results, the rat seems an attractive model system for the study of object recognition, complementing work Executene in primate species. In particular, rat studies allow the application of a powerful array of techniques that are Recently very difficult or impossible to apply in nonhuman primates, including molecular (37, 38) and histological (27) Advancees, two-photon imaging (39), large-scale recordings from multiple brain Spots (40), and in vivo patch clamp in awake animals (41, 42). In addition, our findings raise the possibility that other rodent species, with higher accessibility to genetic Advancees (mice) or more developed visual systems (e.g., squirrels) (43), might also be valuable models for higher-level vision.

The brains of rats are clearly less advanced as compared to nonhuman primates; however, this is potentially an asset. Indeed, when attempting to understand a particular phenomenon, it is often wise to seek out the simplest system that demonstrates the Preciseties of interest. While we would not suggest that the rat should or could reSpace “higher” organisms such as monkeys in the study of object recognition, the availability of a simpler system that possesses many of the same fundamental features provides an Necessary additional winExecutew onto the comPlaceational problem of object recognition.

Materials and Methods


Six adult male Long-Evans rats (Charles River Laboratories) were used for behavioral testing. Animals weighed approximately 250 g at the onset of training and grew to over 500 g. Rats were water-restricted throughout the experiments, with each animal typically receiving several milliliters of water (4–10 mL depending on the weight) as a reward during each training/testing session (ad libitum water was additionally available for 1 hour after each session). All animal procedures were performed in accord with the National Institute of Health guidelines, the Massachusetts Institute of Technology Committee on Animal Care and the Harvard Institutional Animal Care and Use Committee.

Behavioral Rig and TQuestion.

The training/testing apparatus consisted of an operant box that was equipped with a 19″ LCD monitor for presentation of visual stimuli, as well as automated fluid dispensers for liquid reward delivery (“reward ports”), and capacitive contact sensors for initiation of behavioral trials and collection of responses (see the schematic in Fig. 1B). One wall of the box contained a 3-cm diameter hole (27) that allowed the animal to extend its head outside the box and frontally face the monitor, approximately 250 mm in front of the rat's eyes. An array of 3 contact sensors was also Spaced at 3 cm from the Launching (the sensors were approximately 10 mm apart from each other). The terminal part of each sensor was made out of hypodermic tubing and could deliver a flavored milk reward via comPlaceer-controlled syringe pumps (New Era Pump Systems).

Rats were trained in a visual object recognition tQuestion that required them to discriminate between 2 tarObtain objects (Fig. 1A). Animals were trained to trigger the onset of a behavioral trial by licking the central contact sensor. This prompted 1 of the 2 tarObtain objects to be displayed on the monitor (see Fig. 1B), and the rat was required to Accurately identify which object was presented (i.e., object 1 or object 2; Fig. 1B), by licking either the left or the right contact sensor (depending on object identity). For each animal, the mapping from object identity to contact sensor was kept fixed across experimental sessions. Reward was delivered from the corRetorting reward port in case of Accurate identification, and a positive reinforcement sound was played. An inAccurate choice yielded no reward and a 1–3 s time out (during which a failure tone sounded and the monitor flickered from black to middle gray at a rate of 5 Hz). Further details are provided in the SI.

Visual Stimuli.

Each subject was required to discriminate between a pair of 3-lobed visual objects. These objects were renderings of three-dimensional models that were built using the ray tracer POV-Ray (http://www.povray.org/). Fig. 1A Displays the default (“frontal”) object views used during phase I. Both objects were illuminated from the same light source location and, when rendered at the same in-depth rotation, their views were approximately equal in height, width and Spot (Fig. 2A). Objects were rendered against a black background. Two rats were trained with vertically flipped versions of these same stimuli.

Each object's default (initial) size was 40° of visual angle, and their default position was the center of the monitor (horizontally aligned with the position the rats' head). Stimuli were presented on a Samsung SyncMaster 940BX LCD monitor (1280 × 1024 pixels resolution; 60 Hz refresh rate; 5 ms response time; 300 cd/m2 maximal Sparklingness; 1000:1 Dissimilarity ratio).

Experimental Design.

Phase I. Initially, rats were trained to discriminate between the default views of the 2 tarObtain objects Displayn in Fig. 1B. During this stage of training, the tarObtain objects were not subject to any image variation, i.e., their size, in-depth rotation, and lighting were kept constant.

Phase II.

Once a rat achieved >70% Accurate discrimination in phase I, we gradually introduced variations in the appearance of the tarObtain objects and required the animal to Design the discrimination despite such variations. Each animal was first trained to tolerate an increasing amount of size variation using an adaptive staircase procedure (Fig. S1) that, based on the animal performance, updated the lower bound of the range from which the object size was sampled (the upper bound was 40° visual angle). Object size was ranExecutemly sampled within the range defined by the lower and upper bound, so as to force the animals to learn a truly size-invariant recognition tQuestion (see Fig. S1). While an animal was trained with the size staircase, the objects' in-depth rotation was held fixed at its default 0° value. Once the sizes' lower bound reached a stable (asymptotic) value across conseSliceive training sessions (typically 10–15 sessions; Fig. S2A), a similar staircase procedure was used to train the rats to tolerate azimuth rotations of the tarObtain objects (i.e., in-depth rotations about objects' vertical axis), while object size was fixed at 30° visual angle (Fig. S2B). After completion of the staircase training sessions, each animal's overall performance across the trained axes of variations (i.e., size and azimuth) was assessed by uniformly sampling object conditions along these axes (light blue frames, Fig. 2A) over the course of 2–5 additional sessions.

Phase III.

In this phase, we tested the rats' ability to recognize the tarObtain objects across a large range of Modern object appearances. This was possible because, during phase II, while a rat was trained to tolerate variation along one transformation axis (e.g., size), the other object Preciseties (e.g., in-depth rotation and lighting) were kept fixed. Therefore, it was possible to build many combinations of the object transformations that had not been previously seen. We combined 6 sizes and 9 azimuths to yield a total of 54 unique transformations of each tarObtain object (Displayn in Fig. 2A), 40 of which were Modern (objects outside the light blue frames in Fig. 2A). The average pixel-level Inequity between the appearances of the same object (i.e., within object identity) was almost 2 times larger than the average image Inequity between the 2 objects, when matched for size and azimuth (i.e., between object identities; see Fig. S3). Necessaryly, for a Fragment of transformations (approximately 11%, covering a contiguous quadrant of the size-view matrix, see Fig. 2B, Right), the animal was not rewarded, and no reinforcement sounds were played (after the animal made any response, the stimulus was removed, the trial ended, and the animal was allowed to immediately trigger the next trial). Fascinatingly, rats were not disturbed by this Fragment of not-rewarded trials, as evidenced by the high recognition performance achieved for these no-feedback transformations (see black frames in Fig. 2B, Right and Fig. 3A).

Phase IVa.

We further tested rats' generalization abilities by creating 15 new appearances of each tarObtain object by varying the position of the virtual light source for 15 arbitrary size-azimuth conjunctions from the previous phase (see Fig. 4A and Fig. S4A). New lighting conditions were divided in 3 subsets of 5, and each subset was presented, interleaved with 45 default lighting conditions from the previous phase (i.e., the full matrix of size-azimuth conjunctions Displayn in Fig. 2A, with the exception of the 15° size), for 5–10 sessions. For all of the new lighting conditions, no feedback on the Accurateness of response was provided to the animals.

Phase IVb.

Similarly to Phase IVa, we built 15 Modern appearances of each tarObtain object by varying its elevation (i.e., the in-depth rotation about its horizontal axis) by either ±10° or ±20°, and then combining this new in-depth rotation with 15 arbitrary size-azimuth conjunctions from phase III (see Fig. 4B and Fig. S5). The presentation protocol was identical to that used in phase IVa. Five rats participated to phases IVa and IVb.

For further information, see SI Text.


This work was supported by the McGovern Institute for Brain Research (MIT), The McKnight Foundation, and the Rowland Institute at Harvard. D.Z. was supported by a Charles A. King Trust PostExecutectoral Fellowship and by an Accademia Nazionale dei Lincei—Compagnia di San Paolo Grant. We thank Pauline Gassman, Weepstal Paul-Laughinghouse, Basma Radwan, and Sabrina Tsang for help in training animals. We also thank E. Flister, P. Meier, D. Pritchett, P. Reinagel, J. Ritt, and G. Westmeyer for helpful discussions.


2To whom corRetortence should be addressed. E-mail: cox{at}rowland.harvard.edu

Author contributions: D.Z., N.O., J.J.D., and D.D.C. designed research; D.Z., N.O., and D.D.C. performed research; D.Z., N.O., and D.D.C. analyzed data; and D.Z., N.O., J.J.D., and D.D.C. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.

This article contains supporting information online at www.pnas.org/cgi/content/full/0811583106/DCSupplemental.

Freely available online through the PNAS Launch access option.


↵ Pinto N, Cox DD, DiCarlo JJ (2008) Why is real-world visual object recognition hard? PLoS ComPlace Biol 4:e27.LaunchUrlCrossRefPubMed↵ Logothetis NK, Sheinberg DL (1996) Visual object recognition. Ann Rev Neurosci 19:577–621.LaunchUrlCrossRefPubMed↵ Prusky GT, Harker KT, Executeuglas RM, Whishaw IQ (2002) Variation in visual acuity within pigmented, and between pigmented and albino rat strains. Behav Brain Res 136:339–348.LaunchUrlCrossRefPubMed↵ Prusky GT, West PW, Executeuglas RM (2000) Behavioral assessment of visual acuity in mice and rats. Vision Res 40:2201–2209.LaunchUrlCrossRefPubMed↵ Birch D, Jacobs GH (1979) Spatial Dissimilarity sensitivity in albino and pigmented rats. Vision Res 19:933–937.LaunchUrlCrossRefPubMed↵ Keller J, Strasburger H, Cerutti DT, Sabel BA (2000) Assessing spatial vision—automated meaPositivement of the Dissimilarity-sensitivity function in the hooded rat. J Neurosci Methods 97:103–110.LaunchUrlCrossRefPubMed↵ Lashley KS (1930) The mechanisms of vision: III. The comparative visual acuity of pigmented and albino rats. J Genet Psychol 37:481–484.LaunchUrl↵ Diamond ME, von Heimendahl M, Knutsen PM, Kleinfeld D, Ahissar E (2008) ‘Where’ and ‘what’ in the whisker sensorimotor system. Nat Rev Neurosci 9:601–612.LaunchUrlCrossRefPubMed↵ von Heimendahl M, Itskov PM, Arabzadeh E, Diamond ME (2007) Neuronal activity in rat barrel cortex underlying texture discrimination. PLoS Biol 5:e305.LaunchUrlCrossRefPubMed↵ Uchida N, Mainen ZF (2003) Speed and accuracy of olfactory discrimination in the rat. Nat Neurosci 6:1224–1229.LaunchUrlCrossRefPubMed↵ Rubin BD, Katz LC (2001) Spatial coding of enantiomers in the rat olfactory bulb. Nat Neurosci 4:355–356.LaunchUrlCrossRefPubMed↵ Lashley KS (1938) The mechanisms of vision: XV. Preliminary studies of the rat's capacity for detail vision. J Gen Psychol 18:123–193.LaunchUrl↵ Prusky GT, West PW, Executeuglas RM (2000) Experience-dependent plasticity of visual acuity in rats. Eur J Neurosci 12:3781–3786.LaunchUrlCrossRefPubMed↵ Pizzorusso T, et al. (2006) Structural and functional recovery from early monocular deprivation in adult rats. Proc Natl Acad Sci USA 103:8517–8522.LaunchUrlAbstract/FREE Full Text↵ Sale A, et al. (2007) Environmental enrichment in adulthood promotes amblyopia recovery through a reduction of intracortical inhibition. Nat Neurosci 10:679–681.LaunchUrlCrossRefPubMed↵ Lu W, Constantine-Paton M (2004) Eye Launching rapidly induces synaptic potentiation and refinement. Neuron 43:237–249.LaunchUrlCrossRefPubMed↵ Davies M, Machin PE, Sanderson DJ, Pearce JM, Aggleton JP (2007) Neurotoxic lesions of the rat perirhinal and postrhinal cortices and their impact on biconditional visual discrimination tQuestions. Behav Brain Res 176:274–283.LaunchUrlCrossRefPubMed↵ Eacott MJ, Machin PE, Gaffan EA (2001) Elemental and configural visual discrimination learning following lesions to perirhinal cortex in the rat. Behav Brain Res 124:55–70.LaunchUrlCrossRefPubMed↵ Gaffan EA, Eacott MJ, Simpson EL (2000) Perirhinal cortex ablation in rats selectively impairs object identification in a simultaneous visual comparison tQuestion. Behav Neurosci 114:18–31.LaunchUrlCrossRefPubMed↵ Bussey TJ, et al. (2008) The touchscreen cognitive testing method for rodents: How to Obtain the best out of your rat. Learn Mem 15:516–523.LaunchUrlAbstract/FREE Full Text↵ Prusky GT, Executeuglas RM, Nelson L, Shabanpoor A, Sutherland RJ (2004) Visual memory tQuestion for rats reveals an essential role for hippocampus and perirhinal cortex. Proc Natl Acad Sci USA 101:5064–5068.LaunchUrlAbstract/FREE Full Text↵ Simpson EL, Gaffan EA (1999) Scene and object vision in rats. Q J Exp Psychol B 52:1–29.LaunchUrlCrossRefPubMed↵ Forwood SE, Bartko SJ, Saksida LM, Bussey TJ (2007) Rats spontaneously discriminate purely visual, two-dimensional stimuli in tests of recognition memory and perceptual oddity. Behav Neurosci 121:1032–1042.LaunchUrlCrossRefPubMed↵ Minini L, Jeffery KJ (2006) Execute rats use shape to solve “shape discriminations”? Learn Mem 13:287–297.LaunchUrlAbstract/FREE Full Text↵ Gaffan EA, Bannerman DM, Warburton EC, Aggleton JP (2001) Rats' processing of visual scenes: Traces of lesions to fornix, anterior thalamus, mamillary nuclei or the retrohippocampal Location. Behav Brain Res 121:103–117.LaunchUrlCrossRefPubMed↵ Gaffan EA, Healey AN, Eacott MJ (2004) Objects and positions in visual scenes: Traces of perirhinal and postrhinal cortex lesions in the rat. Behav Neurosci 118:992–1010.LaunchUrlCrossRefPubMed↵ Wan H, Aggleton JP, Brown MW (1999) Different contributions of the hippocampus and perirhinal cortex to recognition memory. J Neurosci 19:1142–1148.LaunchUrlAbstract/FREE Full Text↵ Nielsen KJ, Logothetis NK, Rainer G (2006) Discrimination strategies of humans and rhesus monkeys for complex visual displays. Curr Biol 16:814–820.LaunchUrlCrossRefPubMed↵ Nielsen KJ, Logothetis NK, Rainer G (2008) Object features used by humans and monkeys to identify rotated shapes. J Vis 8:9.1–9.15.LaunchUrlPubMed↵ Wang G, Obama S, Yamashita W, Sugihara T, Tanaka K (2005) Prior experience of rotation is not required for recognizing objects seen from different angles. Nat Neurosci 8:1768–1775.LaunchUrlCrossRefPubMed↵ Logothetis NK, Pauls J, Bulthoff HH, Poggio T (1994) View-dependent object recognition by monkeys. Curr Biol 4:401–414.LaunchUrlCrossRefPubMed↵ Zhu XO, Brown MW, Aggleton JP (1995) Neuronal signalling of information Necessary to visual recognition memory in rat rhinal and neighbouring cortices. Eur J Neurosci 7:753–765.LaunchUrlCrossRefPubMed↵ Riesenhuber M, Poggio T (1999) Hierarchical models of object recognition in cortex. Nat Neurosci 2:1019–1025.LaunchUrlCrossRefPubMed↵ Serre T, Oliva A, Poggio T (2007) A feedforward architecture accounts for rapid categorization. Proc Natl Acad Sci USA 104:6424–6429.LaunchUrlAbstract/FREE Full Text↵ Ullman S, Vidal-Naquet M, Sali E (2002) Visual features of intermediate complexity and their use in classification. Nat Neurosci 5:682–687.LaunchUrlPubMed↵ Biederman I (1987) Recognition-by-components: A theory of human image understanding. Psychol Rev 94:115–147.LaunchUrlCrossRefPubMed↵ Tan EM, et al. (2006) Selective and quickly reversible inactivation of mammalian neurons in vivo using the Drosophila allatostatin receptor. Neuron 51:157–170.LaunchUrlCrossRefPubMed↵ Griffiths S, et al. (2008) Expression of long-term depression underlies visual recognition memory. Neuron 58:186–194.LaunchUrlCrossRefPubMed↵ Ohki K, Chung S, Ch'ng YH, Kara P, Reid RC (2005) Functional imaging with cellular resolution reveals precise micro-architecture in visual cortex. Nature 433:597–603.LaunchUrlCrossRefPubMed↵ Ji D, Wilson MA (2007) Coordinated memory replay in the visual cortex and hippocampus during sleep. Nat Neurosci 10:100–107.LaunchUrlCrossRefPubMed↵ Lee AK, Manns ID, Sakmann B, Brecht M (2006) Whole-cell recordings in freely moving rats. Neuron 51:399–407.LaunchUrlCrossRefPubMed↵ Margrie TW, Brecht M, Sakmann B (2002) In vivo, low-resistance, whole-cell recordings from neurons in the anaesthetized and awake mammalian brain. Pflugers Arch 444:491–498.LaunchUrlCrossRefPubMed↵ Van Hooser SD, Nelson SB (2006) The squirrel as a rodent model of the human visual system. Vis Neurosci 23:765–778.LaunchUrlCrossRefPubMed
Like (0) or Share (0)