Structural Linguistics Division
Perception, Evaluation, and the Relocation of Constraint
Artificial intelligence has long encouraged a model of cognition in which the world is first converted into discrete, human-legible units and intelligence begins only once those units are available for manipulation. Speech becomes phonemes, writing becomes characters, movement becomes coordinates. This paper argues that the error is not symbolism, abstraction, or discreteness, but antecedent abstraction: the requirement that a system inherit its operative ontology before it has learned which distinctions matter.
The same error recurs at two scales. At the perceptual scale, an imposed inventory governs what may enter cognition. At the evaluative scale, an imposed inventory of recognized outputs — a benchmark, a development index, an inventory of technological “cargo” — governs what may count as intelligence. In both, a selective reduction is mistaken for the thing reduced.
The positive contribution is a conservation law. Antecedent abstraction appears at three loci — the input, the objective, and the evaluation — and these are not equally escapable. Self-supervised systems show that the input gate can be removed; but removing it does not eliminate constraint, it relocates the constraint into the objective. The objective cannot be removed at all: every optimization bottoms out in an unoptimized criterion. Evaluation cannot be removed either, only demoted from a gate that decides in advance to a probe that remains open to revision. Hence the Relocation Principle: antecedent abstraction is relocatable but never eliminable, and intelligence advances not by escaping bottlenecks but by relocating them toward more general and more revisable forms.
Proprioception supplies the stress test at the individual scale — a structured, action-relevant intelligence that requires no public inventory of minimal tokens. Yali’s question in Diamond’s Guns, Germs, and Steel supplies it at the social scale — cargo as accumulated historical leverage mistaken for innate capacity. The paper closes reflexively: its own criterion, ontological revisability, is itself an evaluation bottleneck, and is therefore offered as a probe held open to revision rather than installed as a final gate.
Every representational system builds an intermediate and eventually forgets that the intermediate is intermediate. A map is drawn to make terrain usable; in time the map’s categories are mistaken for the terrain. Whitehead named this the fallacy of misplaced concreteness — treating an abstraction as though it were the concrete reality from which it was drawn (Whitehead 1925). This paper concerns one recurrent instance of that fallacy in the sciences of mind: the moment when the representation through which intelligence becomes tractable is mistaken for intelligence.
Newell and Simon’s physical symbol system hypothesis held that a physical symbol system has the necessary and sufficient means for general intelligent action (Newell and Simon 1976; Newell 1980). Whatever its merits as a claim about computation, it helped consolidate an architecture in which cognition appeared to proceed through a fixed sequence — reality, then symbolic description, then reasoning, then action. Because the symbols were inspectable while the processes producing them were not, intelligence came to be identified with what happened after the world had been reduced. Perception became preprocessing; action became output; the decisive work of deciding which differences matter vanished into the first arrow.
The same architecture reappears when intelligence is evaluated rather than enacted. A society becomes an inventory of technologies and institutions; a person becomes a score; a model becomes a benchmark result. Here too a visible middle term acquires false sovereignty: the historical, material, and embodied processes that produced the output are harder to isolate than the output, so the measurement is mistaken for the capacity it only partly reveals.
The object of this paper is the bottleneck itself. The central claim is not that bottlenecks are dispensable. It is the opposite. Bottlenecks are constitutive of intelligence; what changes across systems and across history is not whether a bottleneck exists but where it sits, how general it is, and how open it remains to revision. To make that claim precise requires distinguishing the loci at which antecedent abstraction operates, and showing that they differ in how far they can be moved. That is the work of the next section, which states the paper’s governing principles; every later section is an illustration of them.
Antecedent abstraction is not a single structure appearing once in a cognitive system. It appears at three distinct loci, and the argument depends on keeping them apart, because they do not behave the same way under pressure (Figure 1).
The input bottleneck governs how the world enters the system: the phoneme inventory imposed on speech, the OCR character stream imposed on the page, the labeled object ontology imposed on a scene, the coordinate frame imposed on a body. It fixes in advance the vocabulary in which evidence may arrive.
The objective bottleneck governs what counts as learning: the contrastive loss, the cross-entropy target, the reinforcement signal, the masked-prediction task. It fixes in advance which errors matter, and therefore which distinctions the system has any reason to preserve.
The evaluation bottleneck governs how success is judged: the benchmark, the examination, the development index, the inventory of cargo. It fixes in advance the outputs through which a system may be recognized as intelligent at all.
Each is a place where a human-authored reduction can be mistaken for the structure of the thing reduced. But they are not equally escapable, and the paper’s positive claim is a claim about that inequality.
The input bottleneck is relocatable. The perceptual existence proofs of §4 do this work: a system can learn from waveforms before transcription, from document images before OCR, from sensorimotor streams before a symbolic body model. The gate that once stood at the entrance can be removed.
But removed is not eliminated, and this is the point most easily missed. When a self-supervised model discards the phoneme inventory, the work that inventory used to do does not vanish. It moves. What now decides which acoustic distinctions the system preserves is the training objective and the learned quantization module — both authored by a person. The input gate was not abolished; its function was relocated into the objective. Every apparent removal of an input bottleneck is the transfer of its labor elsewhere in the architecture. This is worth stating as a law in its own right:
The Conservation of Constraint. Removing one antecedent abstraction requires relocating constraint elsewhere in the learning architecture. The objective is where relocated constraint accumulates — the conserved quantity that explains why the input and evaluation gates can move but never disappear.
The conservation law is kin to the no-free-lunch theorems, which establish that no learner outperforms another averaged over all problems, so that any advantage must come from inductive bias matched to the problem (Wolpert and Macready 1997). The present claim is narrower and more architectural: it is not about performance across problem distributions but about where the imposed structure lives. Bias must come from somewhere; conservation of constraint says that when you evict it from the input, it reappears in the objective. This is why the “end-to-end revolution” that abolishes one hand-engineered stage reliably discloses a new dependence on the loss, the reward, or the pretext task (cf. Sutton 2019).
The objective bottleneck is therefore unavoidable. There is no learning without an objective, and “self-supervised” does not mean “unsupervised by humans”; it means the supervision has moved out of hand-labeled targets and into the pretext task and the data distribution, both of which someone selects. One can push the objective up a level — learn the reward instead of specifying it, meta-learn the loss instead of fixing it — but each such move requires a criterion for what makes a reward or a loss good, and that criterion is itself an objective. The regress does not run forever. Every optimization bottoms out in an unoptimized criterion.
The evaluation bottleneck is demotable but not eliminable. To evaluate anything is to hold some prior account of what would count as success; there is no assessment from nowhere. Even the criterion this paper will end on — ontological revisability — is an evaluative abstraction, a prior decision about what to look for. So the evaluation gate cannot be removed the way an input gate can. What can be removed is its finality, its claim to be the last word on what intelligence is. It can be demoted from a gate that decides in advance what counts to a probe that asks what a performance reveals and stays open to the discovery that it asked the wrong question. That demotion is the whole of the remedy on the evaluative side, and it is a smaller remedy than the perceptual side enjoys.
These three results converge on a single principle.
The Relocation Principle. Antecedent abstraction is relocatable but never eliminable. Intelligence advances not by transcending bottlenecks but by relocating them toward more general, more revisable, and less prematurely restrictive constraints.
The optimism of the perceptual literature is thus real but narrower than it appears. Self-supervision did not free perception from imposed structure; it relocated that structure from a narrow, brittle, human-authored ontology of units to a broader, more general objective — and generality, not freedom, is the achievement. The frontier of intelligence is not movement toward objective-free learning, which is undefined; optimization without preference is not learning at all. It is movement toward objectives that constrain the resulting organization as little as possible while remaining sufficient to guide learning. What improves is never the presence or absence of a bottleneck, only its width, its revisability, and how early it forecloses.
A symbolic category can occupy two very different positions, and the difference is epistemic before it is architectural.
Under a bottlenecked architecture the category functions as a gate. Speech must become phonemes, a document an OCR stream, a signed utterance a gloss, a scene a labeled inventory, before later processing may occur; whatever fails to pass through disappears from the system. As a gate, the category is presupposed.
Where the input gate has been relocated, the same category can function as a probe. One can ask whether a learned representation preserves phonemic contrast, graphemic identity, tactile roughness, odor similarity, reachability, or force — using the category as an instrument to investigate what was learned. As a probe, the category is falsifiable. A phoneme-like organization may appear strongly in one layer and weakly in another, vary across languages and objectives, matter for transcription but not for classification. It no longer holds an automatic right to govern the entrance of reality into intelligence; it must earn explanatory power through evidence.
The distinction applies unchanged to evaluation. A benchmark used as a gate declares that whatever it measures constitutes intelligence. A benchmark used as a probe asks what capacity a performance reveals, what scaffolding supported it, what it fails to capture, and whether it transfers when conditions change. The difference is not whether measurement occurs but whether the measurement remains revisable.
One caution governs all probing. Alignment is not identity. A latent feature that predicts a phoneme label is not thereby a phoneme; an odor embedding that predicts “floral” contains no isolated floral variable; a robot state correlated with joint angle is not an explicit body representation. The ability to recover a human category from a representation does not show that the representation is organized in the form of that category. Functional linguistic units, perceptual categories, and computational latent structure are three different kinds of thing, and the argument is not that any of them is unreal but that no one of them should be generalized into a universal architecture of intelligence.
Nothing here escapes mediation. Sound must move a receptor or a membrane; light must become a signal; force must produce a detectable internal change. AI can bypass transcription but not transduction — this irreducible interface is the transductive remainder, the requirement that differences in the world produce processable differences inside the system. But transduction does not require a symbolic ontology: a microphone preserves pressure variation without deciding which phoneme occurred; a joint encoder reports changing position without labeling the movement. There can be access to mediated structure before human classification has exhausted it. That is the space the following systems occupy.
Speech. The phoneme captures language-specific contrast across variable realizations, but speech does not arrive as separated phonemic objects; it is continuous and coarticulated, its cues smeared across time and speaker. Self-supervised models learn powerful speech representations from audio before any transcription (Baevski et al. 2020). This is not an escape from discreteness — wav2vec 2.0 solves a contrastive task over a learned codebook of quantized targets — and that is exactly the Conservation of Constraint at work: the human-authored phonemic inventory is gone, replaced by a partition the objective discovered. Layer-wise analyses show acoustic and linguistic information distributing differently across depth and shifting under fine-tuning (Pasad, Chou, and Livescu 2021; Pasad, Shi, and Livescu 2023). The question is no longer does the model contain phonemes but under which objectives, languages, layers, and interventions does it organize speech in a phoneme-like way — the phoneme used as a probe.
This yields the paper’s first risky prediction. Trained without phonemic supervision, phoneme-like organization should emerge only where it improves prediction or downstream performance: locally acoustic in early layers, most phoneme-aligned in intermediate layers where such distinctions aid recognition, increasingly lexical and contextual later, partitioned differently under different objectives, and not reproducing one universal inventory across languages whose contrastive systems differ. If a single stable phonemic inventory appeared independent of task, language, architecture, objective, and layer, the emergent-reduction account would be weakened. The claim is falsifiable, and testable in silico.
Writing. A document is not a string of graphemes; it is layout, hierarchy, typography, reading order, and spatial relation, and a table cannot be understood by keeping its characters while forgetting their positions. OCR-free document models map images directly to structured output, removing explicit character extraction as a compulsory boundary and avoiding the error propagation of OCR pipelines (Kim et al. 2022). Such a model may grow character-sensitive features while preserving layout — it reads the glyph without stopping to report it. The grapheme survives as an analytical category; it loses its monopoly as the entrance to written meaning (Meletis 2019).
Sign. Stokoe (1960) established that signed languages have internal linguistic structure — handshape, location, orientation, movement, nonmanual features in systematic contrast. A gloss is not the signed event; it compresses simultaneity, expression, and spatial reference. Video-native systems can learn from signing before glossing, and explicit sign-phonological supervision still helps: auxiliary prediction of phonological features substantially improves isolated sign recognition (Kezar, Thomason, and Sehyr 2023). Learned and antecedent structure cooperate; the symbolic theory guides without becoming the only admissible input.
Touch and smell. Tactile perception depends on active contact: roughness is not delivered from an object into a passive receiver but emerges from a relation among surface, exploratory action, bodily mechanics, and sensory pattern. A robotic representation may organize contact around stable grasp, slip, deformation, or expected force rather than a fixed vocabulary — and the absence of a universal “tacteme” does not threaten phonemics; it shows that phonemic organization should not be universalized. Olfaction cautions in a second way: a learned odor map preserves perceptual relationships and predicts human odor qualities (Lee et al. 2023), but a model validated against human descriptors has not escaped human categories — it has learned to predict them. To show reduction not exhausted by those labels, the representation must also support tasks the labels do not name. Human descriptors are probes, not the total geometry of experience.
These belong with the ecological, enactive, and predictive-processing traditions — affordances, action-constituted perception, and prediction of sensory consequence (Gibson 1979; Varela, Thompson, and Rosch 1991; O’Regan and Noë 2001; Noë 2004; Clark 2016) — and with the grounding problem they answer in a specific way: grounding through transduction and task rather than through inherited symbols (Harnad 1990).
Proprioception concerns bodily position, movement, force, and configuration. It arises from many sources at once rather than from one receptor reporting one property (Proske and Gandevia 2012). It is internally indexed: a photograph shows a posture from outside, but proprioception is the body’s continuing relation to its own state and capacity for action, and it cannot be handed around, replayed, or inspected by a second observer.
The body does not coordinate itself by attaching a public symbol to every joint configuration, translating each muscular change into a proposition, or assembling trajectories from a finite alphabet of bodily tokens. There is no standard “propriocepteme,” and that absence is not a gap awaiting a term. Proprioception is not a communicative code and was never a candidate for the functional units that earn the phoneme and grapheme their standing. Its importance is different: it demonstrates that an intelligent system can possess highly organized, action-relevant structure — continuous, distributed, body-relative, temporally extended, predictive, organized around control and possibility — without any publicly enumerable inventory of minimal units. It is quiet in consciousness yet indispensable to coordinated movement, its role most visible where it is lost (Tuthill and Azim 2018). The body is not an alphabet; it is not structureless either.
Its artificial analogue makes the point constructive. A robot receives images, joint positions, velocities, forces, tactile and inertial signals, and action histories — already mediated by sensors and coordinate frames, but not requiring a complete symbolic account of each state in advance. Sensorimotor pre-training masks portions of interleaved image, proprioceptive-state, and action sequences and learns to predict the missing content, transferring across tasks, environments, and robots (Radosavovic et al. 2023). A grounded representation of a cup then need not be a link between a word and an image; it can include how the object moves under contact, which grasps hold, how force changes during lifting, which actions cause slipping, which trajectories remain reachable. Meaning becomes a relation among prediction, action, and consequence — intelligence that must remain competent through action rather than merely describe it.
At the evaluative scale, antecedent abstraction controls not what enters the system but what observers will accept as evidence of cognition. A development index reduces a society to selected outputs; an examination reduces competence to performance under standardized conditions; a benchmark reduces a model to a finite set of expected responses. The operation is the same as at the perceptual scale — select a legible inventory, project a complex system into it, discard what the inventory cannot hold, and treat the projection as the system itself.
AI evaluation reproduces this directly. Fluent prose, solved equations, passed exams, generated code: these are treated as readings of a single underlying substance. But such outputs are manufactured cargo — they arise from architectures, corpora, human feedback, retrieval, tools, and benchmark histories, so a system may appear to hold a capacity partly located in its surrounding infrastructure, while another holds latent structure a benchmark fails to elicit. A language model may manipulate descriptions of physical action while unreliable about physical consequence; a robot may manipulate an object it cannot describe; a person may read an ecology with precision yet fail a decontextualized test. These are different organizations of competence, not interchangeable successes. A benchmark used as a probe asks what capacity was shown, under what scaffolding, with what transfer, against what perturbation; used as a gate, it declares that whatever it counts exhausts intelligence.
Two named hazards sharpen this. Goodhart’s law observes that a measure ceases to be a good measure once it becomes a target (Goodhart 1975; Strathern 1997), and its machine-learning sibling, specification gaming or reward hacking, observes that an optimizer will satisfy the letter of an objective while violating its intent (Amodei et al. 2016). Both are real, but both are downstream, dynamic special cases: they describe what happens to a metric under optimization pressure. The present claim is prior and static — the projection is mistaken for the thing even before anyone optimizes against it. Goodhart tells us measures corrupt when targeted; the bottleneck argument tells us inventories are mistaken for reality as soon as their provisional status is forgotten. The second condition is necessary for the first.
This gives the evaluation half its own risky prediction, parallel to §4’s: systems optimized directly against a benchmark metric should show systematically poorer transfer to genuinely novel task families than systems optimized against broader capability objectives, with the divergence widening as optimization pressure on the benchmark increases. Like the speech prediction, it is testable in silico.
The two scales are not symmetric in one further respect, and honesty requires marking it. The perceptual and AI-evaluation claims generate experimental predictions. The historical case that follows does not function as a laboratory falsifier; it demonstrates the interpretive and political consequences of the same bottleneck structure.
In the prologue to Guns, Germs, and Steel, Jared Diamond recounts Yali, a New Guinean political leader, asking why Europeans had developed and transported so much “cargo” while New Guineans had comparatively little of their own (Diamond 1997). The disparity was real; the most dangerous reading treated it as evidence of unequal minds. Diamond rejected innate inequality and emphasized geography, domesticable species, food production, population density, disease, and diffusion, while allowing that institutions explain cases geography alone cannot. His framework has itself been criticized for environmental determinism and for underweighting colonial agency and coercion (Blaut 2000), and the argument here does not require his explanation to be complete. Yali’s question matters because it exposes a problem of evaluation: an inventory of visible output — weaponry, metallurgy, writing, industrial production, bureaucratic scale — becomes an antecedent abstraction imposed on a society, then treated as the neutral form in which intelligence becomes historically legible. This is the cargo bottleneck. Cargo is real and powerful, but it is a selective projection from a much larger field of cognition, ecology, opportunity, coercion, and accumulated infrastructure. Its possession proves a historical system capable of producing and preserving it; it does not measure the inherent capacity of everyone who holds it.
The social half of the argument belongs to established lineages, and naming them prevents renaming them. James Scott’s Seeing Like a State shows how governments simplify complex realities to make them administratively legible, and how centrally imposed abstractions can destroy the local information competent practice depends on; his term mētis names practical, situated knowledge that cannot always be detached from the conditions of its use (Scott 1998). That is antecedent abstraction at the scale of governance — the state does not merely observe a society but reorganizes it around the categories through which it can be seen. Hutchins’s distributed cognition shows that cognitive properties can belong to a system of people, instruments, and procedures rather than to any isolated individual (Hutchins 1995), and Clark and Chalmers’s extended mind shows that external resources can be constituents of a cognitive process when reliably integrated into it (Clark and Chalmers 1998).
Together these clarify what earlier drafts of this argument risked renaming. A route can preserve spatial knowledge; a tool can embody a solution to recurring constraints; a navigation practice can distribute computation across people and instruments; a cultivated landscape can hold generations of intervention and response. No metaphor of the society as one giant body is needed, and none is offered. Proprioception is internally indexed and biologically integrated; distributed social knowledge is publicly embodied in artifacts and practices. They are not the same phenomenon. They expose the same limitation: explicit symbolic enumeration is not the only form in which effective organization can exist. Cargo, on this view, is not merely an output but external cognitive infrastructure — writing externalizes memory, maps stabilize spatial relations, institutions preserve categories beyond the lives of their makers. It compounds, amplifying cognition without defining it, and creating historical advantages that later look like properties of the people who inherit them.
Rejecting cargo as the measure of intelligence does not make every reduction equally valid. Some models fail; some practices destroy the conditions they depend on; some systems cannot respond when their predictions collapse. The standard cannot be mere difference, and it cannot be survival alone — a coercive order can preserve itself by transferring its costs onto others, and an institution can remain stable while degrading what sustains it. So the criterion for successful revision must be stated carefully:
A revision counts as improvement only when it reduces the relevant consequence across the full system of affected relations, rather than merely redrawing the system boundary so that the cost disappears from measurement.
This is where the argument becomes reflexive in a way that matters for §9. The danger for any success criterion is that it can be satisfied not by improvement but by shrinking the frame until failure falls outside it. A political order that stabilizes itself by exporting its costs is not revising its model; it is contracting what its model is accountable to — improving the score by narrowing the domain. That move is the evaluation bottleneck turned against the world: mistaking a narrowed projection for the reality it was meant to track. A criterion that let it pass would be running the cargo trick on itself. The requirement of distributed consequence is what keeps revisability from collapsing into adaptive rationalization; the world retains veto power, though its veto does not always arrive in the vocabulary the system chose for observing it.
The taxonomy of §2 has a consequence the paper must face rather than avoid. Ontological revisability is offered here as the mark of intelligence — the capacity to preserve action-relevant distinctions, detect when they have failed, and rebuild them under pressure from a world that refuses the categories first imposed on it. But by the paper’s own account, a proposed criterion for what counts as intelligence is an evaluation bottleneck. Revisability is a candidate gate.
A paper whose entire argument is that we mistake our inventories for reality cannot then install its own inventory as the final one. To do so would be to commit, in its last move, precisely the error it spent every prior section diagnosing. The consistency of the argument requires that its own criterion be subject to the argument.
So revisability is not offered as a gate. It is offered as a probe: the best current criterion the author can state, held open to the same revision it demands of everything else. Should a system, a society, or a science reveal a form of intelligence that revisability fails to recognize — some structure that is competent, world-tracking, and yet not well described as revising its reductions — then it is revisability that must be revised, not the phenomenon that must be dismissed for failing to fit. The criterion earns its place only for as long as it keeps recognizing intelligence where intelligence is found, and it forfeits that place the moment it begins to erase what it cannot measure.
The paper therefore ends where its logic requires. It does not claim to have found the unit beneath perception, the true measure of societies, or the final definition of mind. It claims something more modest and more durable. Antecedent abstraction is relocatable but never eliminable; intelligence is the movement of the bottleneck toward more general and more revisable constraints, not its abolition; and the honest form of any criterion — including this one — is the probe that keeps asking whether it has begun to mistake itself for the world.
Amodei, Dario, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. “Concrete Problems in AI Safety.” arXiv:1606.06565.
Baevski, Alexei, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. “wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations.” Advances in Neural Information Processing Systems 33.
Blaut, James M. 2000. Eight Eurocentric Historians. New York: Guilford Press.
Clark, Andy. 2016. Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.
Clark, Andy, and David Chalmers. 1998. “The Extended Mind.” Analysis 58 (1): 7–19.
Diamond, Jared. 1997. Guns, Germs, and Steel: The Fates of Human Societies. New York: W. W. Norton.
Gibson, James J. 1979. The Ecological Approach to Visual Perception. Boston: Houghton Mifflin.
Goodhart, Charles A. E. 1975. “Problems of Monetary Management: The U.K. Experience.” In Papers in Monetary Economics, vol. 1. Reserve Bank of Australia.
Harnad, Stevan. 1990. “The Symbol Grounding Problem.” Physica D 42 (1–3): 335–346.
Hutchins, Edwin. 1995. Cognition in the Wild. Cambridge, MA: MIT Press.
Kezar, Lee, Jesse Thomason, and Zed Sevcikova Sehyr. 2023. “Improving Sign Recognition with Phonology.” In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics.
Kim, Geewook, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. 2022. “OCR-Free Document Understanding Transformer.” In Computer Vision — ECCV 2022, 498–517.
Lee, Brian K., Emily J. Mayhew, Benjamin Sanchez-Lengeling, et al. 2023. “A Principal Odor Map Unifies Diverse Tasks in Olfactory Perception.” Science 381: 999–1006.
Meletis, Dimitrios. 2019. “The Grapheme as a Universal Basic Unit of Writing.” Writing Systems Research 11 (1): 26–49.
Newell, Allen. 1980. “Physical Symbol Systems.” Cognitive Science 4 (2): 135–183.
Newell, Allen, and Herbert A. Simon. 1976. “Computer Science as Empirical Inquiry: Symbols and Search.” Communications of the ACM 19 (3): 113–126.
Noë, Alva. 2004. Action in Perception. Cambridge, MA: MIT Press.
O’Regan, J. Kevin, and Alva Noë. 2001. “A Sensorimotor Account of Vision and Visual Consciousness.” Behavioral and Brain Sciences 24 (5): 939–973.
Pasad, Ankita, Ju-Chieh Chou, and Karen Livescu. 2021. “Layer-Wise Analysis of a Self-Supervised Speech Representation Model.” In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU).
Pasad, Ankita, Bowen Shi, and Karen Livescu. 2023. “Comparative Layer-Wise Analysis of Self-Supervised Speech Models.” In 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).
Proske, Uwe, and Simon C. Gandevia. 2012. “The Proprioceptive Senses: Their Roles in Signaling Body Shape, Body Position and Movement, and Muscle Force.” Physiological Reviews 92 (4): 1651–1697.
Radosavovic, Ilija, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. 2023. “Robot Learning with Sensorimotor Pre-training.” In Proceedings of the 7th Conference on Robot Learning, 683–693.
Scott, James C. 1998. Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed. New Haven: Yale University Press.
Stokoe, William C. 1960. Sign Language Structure: An Outline of the Visual Communication Systems of the American Deaf. Studies in Linguistics, Occasional Papers 8.
Strathern, Marilyn. 1997. “‘Improving Ratings’: Audit in the British University System.” European Review 5 (3): 305–321.
Sutton, Richard S. 2019. “The Bitter Lesson.” incompleteideas.net.
Tuthill, John C., and Eiman Azim. 2018. “Proprioception.” Current Biology 28 (5): R194–R203.
Varela, Francisco J., Evan Thompson, and Eleanor Rosch. 1991. The Embodied Mind: Cognitive Science and Human Experience. Cambridge, MA: MIT Press.
Whitehead, Alfred North. 1925. Science and the Modern World. New York: Macmillan.
Wolpert, David H., and William G. Macready. 1997. “No Free Lunch Theorems for Optimization.” IEEE Transactions on Evolutionary Computation 1 (1): 67–82.