6 Learning on the Neuromorphic Circuit

The Silicon Retina

In 1989 Carver Mead first developed the silicon retina, so-called because its architecture was modeled on the neural circuitry of the human retina. Machine vision in the 1980s was based on correlation; how do two images taken in sequence relate? The silicon retina, however, didn’t track, for example, the apparent change in an object over two different images that had to be individually processed. Rather, it registered the continuous variance of light levels over time, just like the peripheral vision neurons of the retina (Mahowald and Mead 1991). Put differently, the silicon retina used analog rather than digital signals to “see” because the retina for Mead was an analog computer (1990, 1632).

Even as early neuromorphic computing formally reproduced the relations between different neurons on an architectural level, it also sought to approximate the system-level functioning of organic vision through the co-constitution of memory and processing. Mead located cognition at the point of sensory input such that the senses were actively thinking and not passively sensing. The retina became as important as the brain in his model of the mind because visual noise was processed before it passed through the optic nerve: “the neural machinery that performs this first step in the chain of visual processing is located in the outer plexiform layer of the retina, just under the photoreceptors” (Mead 1990, 1632). This highly local computation was borne of his bottom-up schema of biological information processing. It started with the senses rather than the cortex, in the “vast neural iceberg beneath the cognitive tip of conscious thought” at the crest of perception (Johnson 1989, 5). In the drowned portion of this iceberg was the set of processes that faced the “‘blooming, buzzing confusion’ of sensory information that would swamp any known digital system at the time” (1989, 6). Mead repeats the phrase “blooming, buzzing confusion,” borrowed from William James, to decry the environmental noise that “system developers ignore when they build top-down AI systems: no one knows where the top is,” and too often contemporary symbolic AI systems were trapped by underestimating the amount of information processing required (2). As he wrote to control theorist Karl Astrom in 1984, “one analog circuit is worth a hundred thousand instructions,” opposing neuromorphic computers to both digital neural nets and traditional symbolic AI (Mead 1984).

From Species to Child to Chip

In a 1991 talk titled “Silicon Models of Neural Computation,” Mead further mapped out his theory of the chip as child, arguing that “many of these levels [of sensory processing] are prewired at birth, but many other sensory pathways develop from the kind of objects experienced throughout life, especially during early development” (96). This booming, buzzing confusion is the sensory input of the world onto the malleable, plastic, and nondetermined child in her early development, such that this incoherent buzzing noise of the world becomes a legible signal. The chip then reconstructs the conditions for the eye’s thinking as it sorts through this informational storm. Mead’s first chip was the silicon retina and he later developed a silicon cochlea; the initial focus on vision was due to how the eye was seen as having a privileged access to the mind and recalls the centrality of vision in contemporary experiments in artificial neural nets as well as the foundation of the connectionist approach in Rosenblatt’s Mark I Perceptron, similarly a hardware-based neural net designed to process visual signals (Lecun et al. 1989; Block 1962).

Mead’s theory of cognition was not only bottom-up in terms of the hierarchy of information that structured his computer chip; it contained a teleology as well. He argued that this “blooming, buzzing, confusion” was temporally located both as part of evolution and as occurring at the moment of birth (Johnson 1989, 5–6). In his lectures, he often paralleled how early sensory pathways learned to perceive with the development of the senses at a species level over time, starting from simpler visual and auditory processing rather than language.1 While this recalls Turing’s (1950, 460) hypothetical child-machine, in which he described a child’s brain as a largely blank notebook, here the mind of the child-machine is not a blank notebook to be filled with writing, which Turing equates to mechanism, but a self-organizing system that learns as both species and as child in which the mind is secondary to sensory processes.

Elizabeth Wilson (2010) isolates two strands of AI: the first a classical chess-driven AI with centralized decision-making, evenly ordered from perception to cognition to action; the second a child model that emerged in the 1990s, in which cognition, and presumably learning, emerges from sensory perception and not as a separately derived and encoded function. The child, whose cognition emerges from this encounter with the world, in which perception and action are co-constituted, is the figure that undergirds Mead’s earliest experiments in neuromorphic computing from the 1980s and 1990s. Yet, there is little engagement with contemporary research on how children learned in early neuromorphic computing: The child is invoked yet how the child learns remains absent. Instead, Mead focused on how the neural circuitry itself developed. This is perhaps why he focuses on the evolutionary development of the senses, an analogy operating on a radically different temporality than the infant-child for Turing. Such attempts to reproduce the evolutionary history of the human eye on a silicon substrate are an exteriorization of evolutionary memory. Bernard Stiegler termed this new relationship between organism and technology epiphylogenesis, in which evolutionary memory is exteriorized and recorded, not just at the level of the body but through technical supports (Stiegler 1998). The silicon retinal chip serves as both a prosthetic exteriorization of human visual capacity and as the inscription of an evolutionary timeline on the development of the chip itself, echoing Freud’s early claim that in childhood development ontogeny mirrors phylogeny, even as it is the very compression of ontogeny and phylogeny onto the timeline of the chip that serves as the theoretical ground for rendering evolution and learning as analogous processes of adaptive change (Freud 1915–17; Sweller 2003).

Current developments in AI and analog neural nets far exceed the limitations that Mead anticipated, but the model of cognition he isolated continues to structure the mind of neuromorphic computing, both theoretically and physically, through the architecture of neuromorphic chips themselves. The way neuromorphic chips know the world is inseparable from their materiality. This mind is distributed throughout the body, thus deprivileging the cerebral cortex, and exists at the point of sensory data collection, like the retina, and is bottom-up and infantile, produced via the sheer mass of sensory inputs rather than a centralized set of instructions. In other words, this is a theory of artificial intelligence with an embodied sensory a priori.

A short review of how exactly the silicon retina saw, and learned, meaning how its rudimentary algorithm processed visual signals, reveals how this sensory embodiment was materialized. The amplified output voltage of the chip’s photoreceptor passed through a horizontal resistive network that took a spatially weighted average of nearby photoreceptor outputs. This weighted average served as a predictive model, meaning that the output voltage of each retinal “cell” was the difference between the local light intensity and the averaged output of neighboring photoreceptors. The resistive network gave the silicon retina adaptive capabilities, allowing for automatic gain control, contrast, and edge detection (Mead 1990). Unlike a neural net run on a digital computer, the silicon retina was a non-Von Neumann computer that lacked a memory unit. Instead, it deployed an early floating gate technology allowing for rudimentary processing-in-memory, such that each neuron had its own memory similar to how each neuron in the brain has its own membrane potential (Mead 1989).

Learning here is grounded in the chip’s adaptive feedback that compared each retinal cell’s output to the overall model, meaning that its baseline was a learned relative light intensity rather than a preprogrammed and unchanging light intensity. The system became its own referent, recalling Orit Halpern’s description of midcentury cybernetic vision as “a self-generated process from within a system” (2015, 64). In Mead’s own words, even if this particular example “learns only a simple model, it illustrates a much more general principle: this kind of system is self-organizing in the most profound sense” (1989, 1634). Learning emerges from the chip’s, or organism’s, embodied sensing of the environment, which is theorized here as self-organized, self-generated, and internal to the system.

Plasticity

By foregrounding the role of environmental stimuli in childhood and chip development, Mead’s theoretical framing of the silicon retina forwarded a biology grounded in epigenetics and its constituent plasticity. Catherine Malabou (2021) argues that epigenesis is the new paradigm for artificial intelligence, one that maps onto a shift from genetics to epigenetics in the biological sciences. Epigenetics refers to processes that alter gene activity without altering DNA itself and, as a paradigm, underscores adaptability, plasticity and non-identity in the face of genetic determinism. For Malabou (2019), epigenetics is a “different name for intelligence,” whose dynamic “leads to no reification, to no substantial or essential state, but rather to what Piaget calls equilibrium, a mobile point of stability between all the intellectual, moral, and affective aspects of the individual” (11). This emergent plastic anti-essentialism theorizes intelligence and learning as adaptability and the ability to maintain a stable being, or equilibrium, in the face of outside factors.

Malabou turns to artificial intelligence and synaptic chips—one of the descendants of Mead’s experiments—to argue that as much as plasticity and neural architecture have become the framework for artificial intelligence, the shift toward plasticity and the brain in fact reveals the very incommensurability of artificial and natural brains rather than an imagined singularity (Malabou 2019). The more our brains share a formal resemblance with computers, the more it becomes apparent that computers are moving beyond simulation and producing new “forms of intelligence that no longer draw on the human,” embodying what Luciana Parisi terms AI’s alien reason (Malabou 2019, 151; Parisi 2019). At the center of Malabou’s theory are synaptic chips that “create a new generation of field-adaptable neurosynaptic computers capable of online learning,” and introduce a material plasticity to computer engineering and artificial intelligence (2019, 85). Yet these synaptic chips build on Mead’s early silicon hardware, and the silicon retina offers a different history of the current turn to analog AI, neuromorphic chip architecture, and epigenetics as a new paradigm for (machine) learning. A historically contingent chip plasticity thus emerges from the analog simulation of the human sensorium on a silicon substrate. We can see this focus on the epigenesis of artificial intelligence in Carver Mead’s theorization of the chip/child as the product of nondeterministic self-organizing learning and the product not just of genetics but an encounter with the noisy, “booming, buzzing confusion” of the world, which is first processed by the eye (Mahowald and Mead 1994). Nondeterminism thus serves as a material and theoretical link between how the chip and the child learn. At the same time, the child must be rendered as an abstract figure far from existing literature on learning in order for this linkage to hold.

In conclusion, what image of learning is found in the neuromorphic chip? In claiming to see like a person, the silicon retina collapses neurons and circuits at a physical level such that the process of human and computer sense-making are rendered one and the same. Learning is a technosocial process enacted on the disparate timelines of chip, child, and the deep time of evolution, which converge at the level of the circuit. If the eye is a two-way window we look through to discern the mind, the silicon retina, in its claim to be closer to the human mind than either symbolic AI or digital neural nets, mimics the pathways of a retinal neuron in the human eye in order to understand the mind. But does it see like a child? A species? A model? Is there a mind to be seen here, beyond one’s own reflection? No, however such investment in the liveliness of the neuromorphic chip, its constituent nondeterministic biology, and an embodied sensory a priori by virtue of how the chip architecture isomorphically reproduces the retina indexes a shift in both how AI is theorized and what exactly a computer can be.

Note

1. 1. Evolution appears multiple times throughout the Mead papers, from a video interview with collaborator Misha Mahowald in which she claims that “we need to evolve our understanding just the way evolution evolved the organism sit came up with. It didn’t sit down in advance and decide this is the blueprint for a person” (48:00) and that “we believe we can grow things, we can grow silicon chips based on a set of constraints. Silicon that can see, silicon that can use analog electronic signals” (49:05) and lecture slides that directly compare the bottom-up evolution of the senses to the bottom-up development of the silicon retina (Mahowald 1995; Mead 1994, 82.14).

References