Review: <i>DeepAesthetics: Computational Experience in a Time of Machine Learning</i> , by Anna Munster

Across the last two decades, media studies scholar Anna Munster has consistently framed emerging new media practices and technologies in relation to philosophical questions, helping articulate the changing nature of being and knowing within a culture shifting from representation to computation. Her 2006 book, Materializing New Media: Embodiment in Information Aesthetics, for example, counters the eclipse of the body in Cartesian models of information, calling for more attention to embodiment and affect as they are imbricated in digital networks and using ideas of the baroque to help situate her argument. She writes, “The nuanced baroque idea of series, of sensations and aesthetic experiences unfolding according to differentials of degree, speed and intensity gives us a way to think about relationships between embodiment and information.”1 The concept of the differential will become central for Munster in this project as accounting for it “is to reinsert the value of those intervals of non-capture, malfunction and chance fluctuating immanent to materiality back into the series of perfect replicas.”2 For Munster, the role of the error or malfunction is central, opening up exciting new vectors for thought within what are more often instrumental or closed systems.In her subsequent book, An Aesthesia of Networks: Conjunctive Experience in Art and Technology (2013), Munster continues her interest in process philosophy and uses the work of contemporary media artists to present a form of thinking dedicated to understanding networks as relational fields. Toward the end of the book, Munster prompts us not to analyze the “logic of networks,” but instead to imagine ways to compose ourselves as “collectivities/networks.”3 She continues, asking, “What novel discoveries can we make about our/the world’s relations of betweenness, withness, toness, andness?”—with networks and our networked condition being the impetus driving the need to answer this question.4Munster’s thematic concerns—process philosophy, relational thinking, embodiment, aesthetics, malfunction, and the critical work of media artists—continue in her newest book, DeepAesthetics: Computational Experience in a Time of Machine Learning, published by Duke University Press in 2025. Here, however, Munster turns to machine learning and the language of computation, areas that have many of us in media studies scratching our heads. While the language of digital media and networks was in many ways challenging for scholars from the 2000s on, especially for film scholars exploring an expanded array of emerging forms of moving image art, the conceptual complexity of machine learning requires understanding not just a new set of terms and concepts, but completely shifting how we think about seemingly simple words such as “image” or “database.” From large language models to latent space, deep learning to GANs, embeddings to word vectors, the array of terms now necessary to consider images in their full array is daunting. However, although Munster pointedly acknowledges a new vocabulary by including pop-up definitions of key terms in boxes interspersed throughout the text, her focus in the book is on drawing out the processual qualities of machine learning, and the implications of its ongoing modulation for perception. “How can we perceptually register and even account for what occurs computationally at scales, durations, and dimensions that are nonhuman and, frequently, imperceptible?,” she asks (9). She points us toward understanding machine learning as an ensemble, one that is always both in process and in relation. Her goal overall, as with her two previous books, is to open up ways of thinking that imagine how else this technology might evolve, even as it is driven toward precise instrumental outcomes based on a narrowly construed understanding of its potentials.Munster begins her analysis by exploring errors in machine learning, specifically focusing on “spectacular category mistakes” (41). She describes a project in which she used machine learning to look back at images of its own making by building a dataset of scientific images in the open access repositor arXiv, which contains research papers across scientific domains. Noting developments in the ability of ML to recognize everything from handwritten numbers to cats, Munster was surprised to find that the image recognition algorithms were unable to classify the very visual forms it had created in arXiv. This prompts her to ask, “Could it be that the category mistakes of ML-based AI, especially pervasive in computer vision, foreground a kind of strangeness at the heart of both the operativity and sensibility of contemporary computational experience?” (41). Rather than gloat over AI’s inability to render “accurate” or “correct” responses, however, Munster instead invites us to recognize that the mistakes and failures of ML point to the “very processuality of computation” (49). Where we might want to look for identity or equivalence, Munster argues we would do better to consider a core quality of heterogeneity in ML, especially within a context that acknowledges that ML is very much the result of social factors, not simply technical infrastructures.Munster goes on to explore the strange figure of time represented by machine learning, noting that its output is connected to the past. It predicts, pointing toward the future, but only by looking back on how it was trained in the past. Munster writes, “Here we encounter the rather odd space-time of ML’s autopoiesis. We could say that while the generalized program of ML is future oriented, segmenting present data so that it aligns with definitive predictions, the becoming or heterogenesis of ML loops backward to the future, recursively iterating its outputs via what its inputs will have become” (59).Munster also explains convolutional neural networks (CNNs), generative adversarial networks (GANs), and creative adversarial networks (CANs). She reminds us that GANs strive to repeat rather than innovate. She writes, “GANs become another mechanism for promulgating visual equivalence underwritten by a model that progressively interpolates novel data instances back into ones already recognized” (71). However, creative adversarial networks, defined as “an attempt by deep learning researchers to specify a more general generative adversarial network of GAN architecture toward producing creative (image) outputs,” offer more provocative potentials (68). Because they are designed to generate images that move toward indeterminacy rather than strive for equivalence through a process that reduces difference, CANs point toward the concept of DeepAesthetics that Munster is developing. She writes, “Perhaps, then, the CAN is less a machine for producing artworks as instances of (human) creativity, as its modelers claim, and more a machine for learning to move (with) flows of images that immanently difference” (73). She continues, “We humans cannot ‘see’ this moving flow despite all the images the CAN does and might continue to produce. Instead, we register the aesthetic feeling of such moving difference both across the relationality of the CAN images (especially when seen as a grid of selections out of the past output of all the model’s image instances) and in their differencing relation to their ground—the canon of Western painting” (73–74). She goes on to celebrate the feeling we might experience of “a differenced continuity,” or a kind of modulation that moves through a tradition or canon and yet retains consistency (75). Munster notes that this experience of feeling a differenced continuity is not about attributing a specific aesthetic to processuality generally. Instead, she illuminates the differing potentials of various kinds of networks in order both to help readers understand foundational components of machine learning and to continue to sketch the potentials for novelty. She writes, “The challenge for a novel aesthesia in which ML plays a vibrant, rather than reactionary, role is to carefully probe at and think with the sociotechnicality of ML experience as it moves and multiplies its relationality via slippages and transductions of quantity and quality” (77).This search for the “vibrant role” of ML continues throughout the book as Munster explores a variety of artists’ projects that similarly examine the potentials for nonstandard uses of machine learning. In the work of Stephanie Dinkins’s chatbot Not The Only One (or N’TOO, 2018–), for example, Munster highlights the ways in which the nonconventional data used to train the chatbot helps result in a form of speech that is also nonconventional and far from perfect. Redundancies, digressions, repetitions: all of these quirks in the speech of N’TOO may be “errors” to someone seeking a more instrumental performance of speech, but to Munster, they represent the far more compelling potentials of variability and the specifically non-optimized.Munster continues to limn the concept of DeepAesthetics by looking at projects by artists such as Anna Ridler, Tega Brain, Trevor Paglen and Kate Crawford, and Monica Monin, among others. With each, she finds value in the diverse ways in which we come to understand machine learning as errant, strange, indeterminant, and mistaken, and she insistently invites us to veer away from the “pursuit of perfect alignment” and instead to celebrate “the speculative and enigmatic realms of machine learning” (178). Overall, DeepAesthetics is a challenging text for scholars unfamiliar with machine learning; however, it is also a generous text, serving a pedagogical purpose while positing a creative, productive, and even necessary perspective that is often thrilling.