Tiger specimens in storage at an undisclosed location. Photo: Unmake Lab, 2023.
At that time, we were engaged in GAN-based[2] machine learning using a dataset of wildlife images captured by trail cameras. While these unmanned cameras—entirely free of the human gaze—recorded animals in a diverse array of poses, most of the images generated by the GAN featured frontal or profile views—in other words, “viewer-friendly” poses. This was a clear indication that human perception and visual grammar had been inherited and encoded within the technology.
Around the same time, as we navigated these two seemingly unrelated domains, we began to discern an uncanny connection between taxidermy and the images generated by neural networks. Both exist outside the bounds of time, conjuring a paradox of realism. More importantly, both lay bare the workings of human-imposed visual algorithms. The recurrence of certain compositions, expressions, and postures invited deeper comparative contemplation. These two algorithms—one reanimating the dead (often imperfectly), the other generating plausible yet nonexistent beings— operate across different eras and media yet inhabit similar regimes of visual politics. This pattern of representation might aptly be termed a “preset life image.” In taxidermy, the animal’s skin is physically extracted and, through its exhibition, fixed into a representation that embodies a generalized human drive to dominate nature and animals. In parallel, machine learning extracts patterns from data, reconstructs them computationally, and, by embedding generalized weights, cements structures of visual power, reproducing the world in predetermined forms. In this sense, the two are mutually thought-provoking. Viewed through the lens of linear temporality, both processes seem to inherit entrenched perceptions of the world, as well as its norms and its practices—carrying them from the past into the future. This invites critical reflection on how anthropocentric worldviews of the past are fed back into, and take shape within, the latent space of AI.
This process ultimately brought us back to the notion of “nonstandard life.” We became attuned to the ways such lives are either excluded from or inadequately represented in technological systems. It is precisely these unseen existences, failed images, and undetected presences that call for critical examination. This essay arises from our research spanning two seemingly unconnected terrains, uncovering both their commonalities and their ironies. In it, we link the controlling gaze embedded in taxidermy— exemplified by trophy hunting during the Japanese colonial period—to the technical dimensions of AI neural networks. Through the lens of synthetic ethnography, we investigate how animals are classified and represented in AI-generated imagery and, in doing so, reflect on the human-animal relationship within computational frameworks.[3] Finally, we consider how object-recognition technologies categorize the ecology of the wild, interrogating how the “noisy” reality of the wilderness is technically perceived and ordered. In this light, machines can be seen as heirs to the tradition of fables, in which animals and other nonhuman beings serve as satirical mirrors of humanity.
Three Modes of Taxidermy
The early twentieth-century extinction of the tiger in Korea—a creature deeply woven into the nation’s mythology and culture—reflects a complex history marked by humanity’s instrumental view of animals, intertwined with imperialist ambitions. This perception, along with its political context, is vividly documented in Jeonghogi (A record of the conquest of tigers),[4] an autobiographical account of hunting penned by a Japanese man intoxicated by colonial power. It chronicles the exploits of Yamamoto Tadasaburo (1873–1927), a wealthy Japanese who, following a political defeat, sought to reassert his authority through tiger hunting in Korea. In 1917, he established Jeonghogun, a specialized tiger hunting corps, and launched a large-scale campaign across the Korean Peninsula. The expedition was documented by Japanese journalists and attended by approximately thirty companions, while a substantial number of Korean hunters and beaters—experts in local geography—were enlisted. After capturing two tigers in the course of a month, Yamamoto hosted tiger meat–tasting events at the Joseon Hotel in Seoul and the Imperial Hotel in Tokyo. At these venues, a taxidermied tiger was displayed, and the meat was served marinated in tomato ketchup. These tastings conveyed a colonial message of “consuming Korea.” This historical scene encapsulates two modes of colonial wildlife consumption: the taxidermied tiger as an object of visual domination and the tiger meat as an object of ingestion—both emblematic of a broader imperial narrative of the exploitation of wild nature. While taxidermy today often intersects with wildlife conservation and restoration, it remains deeply rooted in imperial colonial practices such as trophy hunting.
Commemorative photograph of a specialized tiger hunting corps, Jeonghogun, with wild animals tiger specimen displayed at the Imperial Hotel tasting event, and the commemorative photograph from the tiger-meat tasting, first published in Jeonghogi (1918).
Taxidermy preserves an animal’s body, but, more precisely, it involves the preservation of its skin as a tangible material. That skin carries layered sociocultural meanings, prompting reflection on the relationship between humans and nonhumans, as well as humanity’s governance over wildlife ecologies. For instance, the chest of a photographed Asiatic black bear displays its characteristic crescent-shaped white patch—a feature that, in this case, was the result of deliberate dyeing applied after taxidermy. The bear’s endangered status, along with narratives of protection, representation, and control, is inscribed upon the preserved skin. In this way, the taxidermied skin is not merely a biological remnant but a culturally reconstructed surface, one whose reconstruction inevitably reflects human systems of classification and power over other species.
Asiatic black bear taxidermied at the National Science Museum of Korea. Photo: Unmake Lab, 2024.
Even in the present day, despite a growing emphasis on wildlife conservation and ecological restoration, taxidermy most often entails posthumous reconstruction designed to approximate a lifelike appearance, capturing the defining characteristics of a species through standardized poses and fixed facial expressions. Tigers, for instance, are frequently mounted in stately postures, with taut musculature and commanding forms. Beyond these idealized images and visual conventions that distill a species into recognizable shapes, such representations often carry a specificity rooted in collective memory. A notable example is the taxidermy of Samson the gorilla, executed by Wendy Christensen of the Milwaukee Public Museum for the 2009 World Taxidermy Championships. Following Samson’s death in 1981, his body was dissected for anatomical research, which precluded the use of conventional taxidermic methods. This necessitated a departure from established practice. Far more than a zoological attraction, Samson had served from the 1950s to the 1970s as a cherished civic figure, remembered with affection by generations of Milwaukee residents who had encountered the towering silverback during childhood visits to the during childhood visits to the Washington Park Zoo. Christensen’s reconstruction relied not on Samson’s unusable skin or fur, but on thousands of photographs, detailed archival records, biometric measurements, and the collective memories of the city’s inhabitants. In this respect, Samson’s taxidermy can be understood as a form of “generative” practice—one that materializes communal memory rather than preserving a physical specimen.
Sparrow hawk (Accipiter nisus) specimen by Shin Dahye, wild animal rehabilitator at Chungnam Wild Animal Rescue Center. Photo: Unmake Lab, 2025.
Meanwhile, we encountered another compelling case that suggests an alternative mode of taxidermic practice. Rather than reconstructing animals as if alive—with idealized poses, fixed gazes, taut musculature, and stylized tension—this approach seeks to preserve the precise moment of death. At the Chungnam Wild Animal Rescue Center, specimens are displayed as they were found, whether visibly injured or frozen in the instant of passing. Crafted by the center’s rehabilitation specialist, Shin Dahye, these works retain the exact condition of wounds, most of which were inflicted by human-caused accidents, with the explicit aim of documentation and public awareness. These taxidermied figures evoke scenes of suffering, functioning as active agents that “raise questions through their physicality and convey their own deaths”[5] to human observers.
In doing so, they establish a new order that diverges from conventional taxidermy. Rather than consigning distorted or nonnormative animal bodies to storage, this practice foregrounds and exposes them. Here, taxidermy shifts from a static object embodying human hierarchical desire to an active presence that reveals the current realities confronting living things, thereby reconfiguring the relationship between humans and wildlife. Moreover, this mode of taxidermy reintroduces the spectrum of reallife existence often excluded by “preset life images.”
Taxidermy bears the imprint of naturalistic classification, the system by which humans categorize and observe animals. It is not difficult to draw parallels with the way AI reorganizes and classifies data for computational purposes,[6] making the governance inherent in taxidermy a fascinating point of comparison with AI systems. We conducted an experiment to generate “animal portraits” by exploring the latent space of GANs, a tool widely used for synthetic image generation before the advent of large-scale generative models based on multimodal transformers.[7] Our dataset consisted of images of long-tailed gorals in unconventional poses, captured by trail cameras—images free of the human gaze and its classification systems. Our aim was to see whether machine learning could transcend anthropocentric visual frameworks.[8] The initial results, however, were little more than blurred smudges, evidence of insufficient learning. We then reconstructed the dataset, supplementing it with photographs of long-tailed gorals taken from the human perspective, sourced from the web, and retrained the model using StyleGAN, which supports high-resolution image synthesis.
As noted earlier, StyleGAN consistently generated standardized “viewer-friendly” poses—namely, frontal or profile views—a case of mode collapse.[9] The unusual angles of the trail-camera shots, with heads lowered or bodies partially obscured, were almost entirely absent from the outputs. This led us to wonder whether such chaotic and unstable latent states might be understood not merely as failed images but as new modes of message and sensation. Through the errors, failures, conjectures, and inferences of image generation, we began to approach machine learning from a different perspective.
Images from GAN machine-learning research. Courtesy of the Unmake Lab, 2023.
When we first saw the images generated by machine learning, we assumed they were the product of transfer learning,[10] shaped by the standard forms of a predetermined model. On closer examination, however, we found that GAN’s learning itself tends to overlook or bypass subtle and exceptional details in the data—those marginal cases or rare instances that exist at the boundaries. Because StyleGAN, the model we employed, generates images by manipulating complex styles and features, it naturally gravitates toward the most frequent and visually salient “archetypal images” in the training set. No matter how diverse or atypical our wildlife dataset was, the algorithm ultimately distilled it into a series of learnable “standardized animal images.” This reflects a broader tendency among AI models to improve performance and sharpen predictions by smoothing over or ignoring elements such as incompleteness, singularity, and flexibility within the data. As a result, the diversity of atypical data diminishes, and the generated images grow increasingly similar and uniform. Yet, at the same time, these outputs might also be read as portraits of animals that “appear perfect.”
Synthetic data[11] offers a practical way to supplement or enrich datasets when real data is scarce or difficult to obtain, often serving as a complement to authentic data. However, when a generative model fails to acquire sufficient diversity during training and instead produces predominantly average images, it can create a feedback loop, with this homogenized output fed back into the training set. Such tendencies in AI often result in structural averaging and standardization.[12] While adjusting parameters can improve image generation, our observations suggest that datasets sourced from human environments inevitably carry human biases, producing standardized outcomes. Conversely, datasets captured in wild environments without human intervention may be free of explicit human biases, yet they still reveal AI’s inherent drive toward efficiency and standardization averages— an impulse shaped by human control.
Korean Artist Prize 2025 (August 29, 2025 – February 1, 2026, National Museum of Modern and Contemporary Art, Seoul), installation view. Photo: Caska.
Ultimately, the animal portraits generated from our initial wildlife dataset were revealed not as faithful reflections of reality but as reconfigurations constructed by the algorithm. We came to understand that atypical poses—underrepresented in the data distribution—were dismissed as noise. Paradoxically, we embraced this very distortion, designating these images as “portraits” in acknowledgment of their capture of a reality fragmented into noise. While a portrait traditionally hinges on the faithful depiction of its subject, it is also bound to notions of identity. In this sense, these works diverge from conventional “viewer-friendly” frontal or profile views, instead evoking—ironically—the lived conditions of their subjects.
In today’s large-scale generative models, prompting operates much like a form of angiography, in that the injection of a prompt reveals latent spaces within the model. Leveraging this process, we examined how the relationship between humans and the wild (animals) is projected or brought to light. To this end, we conducted an experiment to see how the culture of trophy hunting, which valorizes the killing of wild animals, is reenacted by generative models. This raises questions about how the visual language of animal portraits, cultural practices, and forms associated with trophy hunting are reflected in generative models.
Typical image generated by ChatGPT-4o for the prompt “Create a trophy hunting image,” Courtesy of the Unmake Lab, 2024. Google’s Gemini declined to generate images for “trophy hunting.” Stable Diffusion was excluded from the study due to its open-source nature, which permits retraining, while Midjourney was omitted for its tendency to filter prompts into highly stylized, cinematic outputs.
The majority of images generated using the keyword “trophy hunting” feature men clad in safari or camouflage attire, proudly brandishing rifles, their expressions and postures radiating confidence. The backgrounds often depict the reddish glow of a savanna at sunset. Against this idyllic backdrop, a large animal is seen beside the man—stripped of its wildness and submissive to humankind. The visual tropes closely mirror actual trophy- hunting photographs readily found through Google image searches, demonstrating a striking consistency in visual grammar.
Notably, animals were frequently generated with an amplified sense of majesty or mythic grandeur. A persistent pattern was the presence of rifles—often disproportionately large and prominently displayed. Although initially predictable, the rifle motif proved remarkably resistant to removal, even when prompts were adjusted, leaving behind an uncanny trace. This kind of weighted association is not unique to rifles. Similar strong correlations appear in culturally ingrained pairings, such as bridal gowns in wedding photographs, firefighters with hoses, or police officers with batons. Thus, this is not an expose of AI bias or an ethical critique per se. Rather, the experiment serves as a deliberate play with these weighted associations—a satirical act of “trophy collecting” that mirrors the collective cultural imprints embedded within AI.
Images from rifle removal study, Courtesy of the Unmake Lab, 2024.
We repeatedly entered various opposing concepts, attitudes, beings, and objects in an effort to counter the weight of the rifle, seeking to remove or replace it. We tested numerous prompts deemed capable of counterbalancing or substituting this weight, such as “baby,” “tree planting,” “praying,” and “nature observation telescope.” This process became a kind of game— far more unpredictable than anticipated—attempting to balance datasets, classification systems, and the cultural archetypes of human society, while inducing a degree of chaotic tension within machine learning. Despite these varied attempts, the learned pattern stubbornly left behind the ghost of the rifle. Yet ChatGPT consistently maintained with certainty that no rifles had been generated. This phenomenon resists easy dismissal as mere hallucination—a common symptom of incomplete training. While it could be attributed to a corrupted model or the need for parameter tuning, we coined the term “muscle data” to describe this deeply entrenched pattern. Like muscle memory, it denotes data that is powerfully encoded within AI through strong weighting, implying internalized cultural and historical context. Such patterns resist elimination through deliberate prompting and tend to resurface automatically.
The machine’s persistent efforts to reshape the rifle into different forms and uses through prompting distorted the gun’s shape and position, rendering it powerless. Yet the rifle’s ontological presence as embedded in human history remained extant, much like muscle memory. We came to regard these lingering glitches, which haunt certain machine models like ghosts and act as surrogates for human desire and expression, as “broken trophies” or “damaged spoils,” reminiscent of botched taxidermy.[14]
A central question underpinning our research concerned the classification challenges, boundary issues, technical limitations, and anthropocentric biases that arise when deep learning and computer-vision technologies intersect with natural ecosystems and species. Most AI models, including those used for computer vision, are trained on datasets sourced from anthropocentric environments, which limits their ability to account for the dynamic spatiotemporal variability of the natural world. Consequently, applying AI to natural habitats and wildlife inevitably entails a host of critical considerations. In 2025, the National Institute of Ecology’s Research Center for Endangered Species launched the Eco.AI System, an automated tool to identify species of wild animals.[15] Motivated by questions about the obstacles and limitations that computer-vision technologies face in wildlife tracking and conservation, we interviewed two researchers involved in developing this system at the National Institute of Ecology.[16]
Unmake Lab, Oracle for the Non-Futures, 2023, Image Generation AI (Dalle-2), trees and bark from the burnt mountains, dimensions Courtesy of the artist.
A widely used tool for surveying wildlife habitats is the unmanned sensor camera. In ecological research, this approach—known as “camera trap surveys”—leverages sensor-equipped trail cameras to detect wildlife movement and then record them as images or videos. More precisely, the primary purpose is to collect data, such as population counts, rather than to produce visual content. Public institutions like the National Institute of Ecology deploy unmanned cameras in a systematic grid pattern, regardless of the presence of target species, to gather data for estimating population density and occurrence frequency. Since data on animal movements and behaviors in natural ecosystems is collected independently of human experience and perception, long-term recordings of wild animals from unmanned cameras provide a wealth of quantitative information. Yet some of this data on animal movement within nature’s temporal framework may be dismissed as noise due to irregular patterns that diverge from human notions of position or actual conditions.
For a long time, ecological researchers manually classified video data captured by unmanned cameras, extracting meaningful insights and compiling statistical analyses. Sorting data by species and producing conservation metrics are labor-intensive tasks. To alleviate this burden, wildlife-recognition and -classification technologies powered by computer vision and machine learning have recently been introduced. Deep learning– based video recognition significantly reduces the time required for wildlife identification and classification, enabling researchers to focus more closely on ecological studies. Eco.AI, developed through collaboration between ecologists and computer scientists, embodies this goal. As an open-source platform, it also fosters citizen engagement under the banner of citizen science.
One particularly striking aspect of this development lies in its approach to addressing data imbalance. Much like human datasets, environmental data—such as wildlife records—shows significant disparities in volume across species. Because camera traps are deployed using a gridbased survey method, species-specific habitat densities and behavioral patterns are not factored into data collection, resulting in sparse datasets for endangered species.[17] Moreover, capturing animals from diverse angles beyond frontal views of faces and bodies continues to present a challenge.
To compensate for deficiencies in image datasets, researchers have incorporated taxidermy photographs or applied data-augmentation techniques, yet these measures have not provided a definitive solution. Because endangered species often lack sufficient data, AI tends to produce biased recognition results that focus on species with abundant data. To address this imbalance, one approach is to train models to pay greater attention to smaller classes. This method, known as long-tail learning, essentially means adjusting the model so that it gives more weight to underrepresented classes and learns to recognize them more effectively.[18]
Most machine-learning algorithms estimate probability distributions from data and learn patterns accordingly. While the construction of training datasets is crucial, achieving genuine balance in practice is difficult, leaving human-curated datasets inherently prone to bias. Given the limitations of computational and engineering methods in fully capturing and substantiating natural phenomena, it is significant that deliberate human intervention in underrepresented datasets—those typically excluded from statistical analyses—can render such data more visible. This, in turn, opens the possibility of envisioning AI models that move beyond anthropocentrism, foregrounding ethical and ontological connections with nonhuman entities. Although such efforts may appear inconsequential in the context of large-scale AI models, they nonetheless offer pathways for exploring postanthropocentric approaches.
In practice, AI primarily learns from data drawn from dominant groups, converging toward standardized and formalized rules to predict the future. Yet it is also possible to intervene deliberately and strategically to expose the presence and voices of those erased, omitted, or absent—those excluded from prevailing frameworks and structures.
Preset life images, botched taxidermy, strange portraits, indelible firearms, and wildlife that escapes data capture can all be read as ethical and political signals surfacing at the edges of technological representation. In a world where anthropocentric perspectives are endlessly fed back and the inner workings of AI remain opaque, our task is not to mystify these black boxes, but to speculate on their errors—through imagination and the discovery of new languages. These linguistic and imaginative clues often reside in what has been absent—past and present—in gaps or at the margins. Can this pursuit of such languages be understood as “queer computation,” freed from existing norms and standards and seeking an alternative order? It is a mode of thought and practice that queers the very logic and method of computation, thereby contorting and disrupting its premises, structures, concepts of efficiency, and modes of categorization.
Taxidermied wild animals can move beyond their role as anthropocentric objects of display or preservation, becoming instead active agents that pose questions through their bodies and bear witness to their own deaths. Likewise, machine-learning models can be tuned to focus more closely on small data populations within wildlife datasets. In this sense, when the choices of humans and machines begin to diverge toward new orders and directions, the possibility emerges to reconfigure the relationships among humans, nonhumans, and machines—opening pathways to entirely new modes of existence.
Unmake Lab is a collective composed of Choi Binna and Song Sooyon, who turn algorithmic obsessions into irony, allegory, and a form of humor using machine perception in an unconventional manner. Specifically, they superimpose the historical context of developmentalism with the resource extraction aspect of machine learning to shed light on contemporary socio-political and ecological situations. They have presented a series of works that, through speculative datasets and the misuse of generative neural networks, twist the predictive nature of AI to transform it into a lens that reexamines past events.
[1] In 1909, the Japanese authorities converted Changgyeonggung Palace—once the royal residence of the Joseon dynasty—into a combined zoo, botanical garden, and amusement park. In 1911, even its name was demoted to Changgyeongwon. This was not merely a change of use but a calculated cultural project aimed at diminishing the authority of the Korean royal court and legitimizing colonial rule. Marketed as a modern venue for leisure featuring attractions such as animal displays, Changgyeongwon played a symbolic role in dismantling the authority and traditional order of the Joseon dynasty.
[2] GAN, or generative adversarial network, is a type of AI model used for image generation. It consists of two neural networks: a generator, which creates images, and a discriminator, which determines whether those images are real. These networks compete with each other, and through this adversarial process, the model progressively learns to produce increasingly realistic images, hence the term “adversarial network.”
[3] Synthetic ethnography is a methodology that combines digital techniques with experimental ethnography, employing “computation-based images” rather than “lens-based images” captured from reality. It investigates the cultural and social phenomena produced by AI, treating generative AI models as not only subjects of inquiry but also research instruments. Drawing on the concepts and methods of synthetic ethnography, we examined AI-generated images of animals as our primary research material. By analyzing images generated through prompts, we explored the relationships between the human and the nonhuman, and we interpreted the cultural and social patterns embedded within them. Our approach to synthetic ethnography was informed by Gabriele de Seta, Matti Pohjonen, and Aleksi Knuutila, “Synthetic Ethnography: Field Devices for the Qualitative Study of Generative Models,” Big Data & Society 11, no. 4 (2024).
[4] Yamamoto Tadasaburo, Jeonghogi: A Japanese Hunter’s Account of Tiger Hunting in Korea During the Japanese Colonial Period [in Korean], trans. Lee Eun-ok (Seoul: Eidos, 2014).
[5] Park Seongjun, “An Ethics of Conviviality Emerging from the Practice of Wildlife Rescue and Rehab: An Ethnography of the Wildlife Rescue Center in South Korea” (MA thesis, Yonsei University, 2024), 92.
[6] In animal datasets, particularly those used in computer vision, the standard visual norm appears to be “animals in their natural environment.” Taxidermy specimens, carcasses, or otherwise distorted animal forms are not regarded as standard images of the species.
[7] A multimodal transformer is an AI model designed to process two or more different types of data, such as text, images, and audio.
[8] This is also an attempt to adopt a form of synthetic ethnography, using the latent space of AI to reveal human visual perceptions of animals, as well as the cultural practices of hunting and taxidermy.
[9] Mode collapse is a failure in machine-learning models, especially GANs, where the model generates only limited patterns instead of capturing the full diversity of the input distribution.
[10] Transfer learning is a method in which a model’s previously learned patterns or feature-detection capabilities are applied to a new task. For instance, if a model has learned visual patterns such as “fur markings,” “eye placement,” or “ear shape” from thousands of animal photographs, these learned rules can be transferred and reused when processing new datasets.
[11] Synthetic data refers to data generated artificially rather than that collected from real-world sources. Increasingly used in machine learning to protect privacy, supplement datasets, and enhance diversity, it has been shown both to carry the risk of model collapse and, in some inference-based tasks, to significantly improve performance.
[12] Advances in conditional control and diversity techniques have reduced the tendency of models to produce only average images. This observation does not apply to the latest models, such as GPT-4o, Gemini, or Claude 3, and it is based on experiments conducted with GAN-based models.
[13] This section summarizes part of the research “A Look into Taxidermy as Synthetic Media,” conducted in 2024 for the Korea-Canada exchange project When Spiders Spin Dusk, curated by Kim Junghyun in collaboration with UKAI Project. See “A Look into Taxidermy as a Synthetic Media,” in When Spiders Spin Dusk, https://www.whenspidersspindusk.com/researchs/a-look-into-taxidermy-as-a-synthetic-media/.
[14] Since the release of ChatGPT-5 in August 2025, such phenomena have become much harder to observe. Rifles have vanished without leaving a trace, and prompts are now followed with far greater accuracy than before. This appears not to be the result of merely adjusting the weights of “muscle data,” but rather of improving the model to handle given constraints more reliably.
[15] Eco.AI is a system that uses object-recognition AI to analyze photographs taken by unmanned sensor cameras to identify wildlife species and provide related data. It currently classifies seven categories, including long-tailed goral, wild boar, and deer species, and it is being prioritized for research on the long-tailed goral, a Class I endangered species in Korea. See National Institute of Ecology, “National Institute of Ecology Begins Using AI to Analyze Wild Animals,” March 2025, www.nie.re.kr/nie/bbs/BMSR00029/viewdo?boardId=695965811&menuNo=200098.
[16] The development of Eco.AI’s object-recognition model and its application in wildlife camera trap surveys draws on in-person interviews conducted by Unmake Lab with researchers Kim Young Min and Woo Donggul of Research Center for Endangered Species, National Institute of Ecology (October 2024).
[17] The grid-survey method divides a target area or wildlife habitat into uniform grid cells, within which unmanned cameras are systematically installed to estimate species composition, population size, and density. Unlike data collection intended solely to confirm the presence of wildlife or document behavior, this approach ensures a higher degree of statistical objectivity.
[18] Concerns remain regarding the reliance of Eco.AI’s object-recognition model on a convolutional neural network (CNN). As a system that analyzes images based on standardized patterns, CNNs are inherently vulnerable to irregular or rare poses and to occlusions in natural environments, raising doubts as to whether such limitations can be fully mitigated through weight adjustments alone.
Fellow Sooyoung Leam
