A net acquires a history
Paper 0046 gave an idealised neural net a formal semantics. Once thresholds, inhibition and delays were fixed, McCulloch and Pitts could calculate a network’s firing proposition or build a network for a permitted proposition. Yet their calculus left the wiring fixed. A logical diagram could embody a rule, but it did not explain how experience made that rule appear. Paper 0047 supplies the next transformation. Frank Rosenblatt asks how an initially imperfect, partly random organisation can become selectively responsive through its own activity.
The shift is larger than the familiar textbook move from a logic gate to a trainable linear classifier. The 1958 perceptron is a population model. Sensory points drive association cells through excitatory and inhibitory origins; response units read broad source-sets; feedback and reinforcement alter the effectiveness of active cells. Rosenblatt does not know the exact connection pattern and does not want to know it. He seeks predictions from distributions, thresholds, cell counts and overlap probabilities. Probability theory replaces Boolean algebra because the particular wiring is accidental while the gross organisation is the proposed law.
This choice changes what an explanation looks like. An exact diagram explains one machine by tracing its paths. A statistical account explains a family of randomly wired machines by showing that, under measurable constraints, most members tend towards useful behaviour. The individual perceptron can be idiosyncratic; the theory aims to say how its expected discrimination changes with architecture and environment. Reliability is not the absence of randomness. It is an organised result obtained across randomness.
For a research programme about intelligence and consciousness, the important inheritance is methodological. Rosenblatt tries to connect physical variables to behavioural regularities without treating a remembered stimulus as a little inner photograph. That is a serious mechanistic proposal, but not evidence that the model experiences, understands or resembles every detail of cortex. Its achievement is narrower and more durable: it makes learning, generalisation and distributed damage into consequences that a specified network organisation can in principle predict.
A stored picture or a changed pathway?
Rosenblatt presents two broad answers to the storage question. The representational answer treats memory as a coded counterpart of the stimulus. If the investigator found the code or wiring diagram, the remembered image could in principle be reconstructed, as an image can be recovered from a photographic negative. Recognition then needs a further operation: compare the incoming pattern with stored contents, discover whether it has appeared before and retrieve the proper response.
The connectionist answer denies that the sensory image need ever be recorded as a reconstructable object. Retention is a change in connections, pathways or functionally equivalent transmission conditions. A later stimulus recruits the altered pathway and activates an appropriate response directly. Information is present as a preference among possible responses. “Response” is deliberately broad in the paper: it may be a muscular action, but it may also be a distinguishable internal state. The contrast therefore concerns organisation, not a crude division between behaviour and thought.
The two positions can produce the same visible answer while proposing different intermediate causes. A matcher keeps a record, retrieves it and compares. A pathway system has been changed so that current input now flows differently. Behaviour alone may not identify which organisation exists. Rosenblatt’s wager is that a plausible pathway system can be made precise enough to derive discrimination and generalisation from its physical parameters.
Causal probe 1 · Why remove the matcher?
If learning changes the route from sensory activity to response, recognition is already expressed in what the network does next. A separate search through stored replicas is unnecessary. The causal burden moves from a comparison algorithm to the plastic organisation of the pathway.
Worked example 1 · Two memories with the same answer
Suppose a striped card repeatedly produces response R₁. A template model stores the stripe image and selects R₁ after a match. A connectionist model increases the effectiveness of association cells activated by striped cards inside R₁’s source-set. Both say “R₁” on the next trial, but only the first needs a reconstructable stripe. The second needs a changed disposition shared across cells.
The exact wiring diagram is deliberately unavailable
Mid-century switching theory had shown how ideal on-off elements could realise intricate logical functions. Rosenblatt regards that success as insufficient for a biological theory. A hand-designed network can perform comparison or transformation because its connections have been chosen for that algorithm. A nervous system, however, develops with variation, imperfect elements and more local constraints. Its important learning network is not expected to have an identical detailed construction in every organism.
The perceptron therefore begins with random connections subject to gross constraints. Some connections are focal, some distributed; units have thresholds; response source-sets have sizes and overlaps; activity changes a value variable. The model’s questions are statistical. What proportion of association units responds to a stimulus? How often does a unit responding to one stimulus also respond to another? How likely is the correct source-set to acquire the greater value? A wiring diagram cannot answer for the population because no one diagram represents the population.
This is the theory of statistical separability. Two stimulus classes need not be separated by a prewired symbolic predicate. They must produce sufficiently different distributions of association activity that reinforcement can turn their difference into a response bias. The particular cells are exchangeable for much of the analysis. Counts, probabilities and expected overlaps do the explanatory work.
Randomness does not mean absence of structure. The distribution of origin points, the balance of excitation and inhibition, threshold, source-set membership and environmental classes are all constraints. Remove them and there is no reason for useful learning. Rosenblatt’s model is best read as a division of labour: accidental microscopic connectivity supplies diversity; architecture shapes the statistics of that diversity; reinforcement retains useful correlations.
Causal probe 2 · Why can an unknown network be predicted?
The investigator gives up predicting which named cell fires and predicts aggregate quantities instead. If units are sampled under known rules, the expected active fraction and coactivation probabilities can be related to physical parameters even though the individual connection list is hidden.
From a retinal pattern to a winning response
The sensory surface contains S-points. In the analysed optical models an S-point gives an all-or-none response, although Rosenblatt notes that other models could use amplitude or frequency proportional to intensity. These sensory units feed a projection area, A₁. Each projection cell receives a cluster of excitatory and inhibitory origin points whose connection probability falls with retinal distance from a centre. Focal sampling makes a cell sensitive to local contour structure instead of only to total illuminated area.
Connections from the projection area to a later association area, Aₙ, are random. An A-unit fires when the algebraic balance of its active excitatory and inhibitory origins reaches its threshold. The simplified model connects random retinal origins directly to association units. That version is analytically convenient and groups stimuli mainly by coincident illuminated areas; the fuller focal stage is expected to improve discrimination when contour location matters.
Response units receive fibres from large random sets of association cells. The A-units connected to one response form its source-set. In the simplified system, response feedback inhibits association cells outside the winner’s source-set. Responses are thus mutually exclusive: a small advantage lets one response become active and suppress its rival. This is not a soft probability vector returned by a modern classifier. It is a competitive transition in a recurrently coupled population.
Rosenblatt separates the event into two phases. During the predominant phase, the stimulus has activated association units but no response has yet dominated. During the postdominant phase, one response is active and competing activity is suppressed. Initially the winner can be random. If the active association units are reinforced, the next presentation of a related stimulus finds a changed value landscape. Learning means that the earlier winner now has a stronger tendency to recur.
Worked example 2 · Focal and scattered sampling
Imagine two patterns illuminating six of twelve retinal points. One is a compact corner; the other scatters the same six points. A cell with focal origins can respond strongly to the compact patch and weakly to the scatter even though total area is equal. A cell drawing origins uniformly across the retina is less likely to preserve that local distinction. The distribution of origins defines what counts as similarity.
Value is a disciplined placeholder
Each association cell is assigned a value, V, describing the effectiveness of its outgoing impulses. Rosenblatt leaves the physical correlate open: it might be amplitude, frequency, latency or the probability of completing transmission. Activity tends to increase value, inactivity may permit decay, and some models make cells compete for metabolic materials. V is therefore neither a modern scalar weight with a specified optimisation gradient nor a purely behavioural score. It is a proposed bridge variable constrained by what the network must do.
The α-system is uncompensated. An active cell receives an increment for each impulse and holds the gain. Repeated reinforcement raises total value indefinitely. It is simple, but unequal presentation frequencies or source-set activity can produce large biases because nothing controls the budget.
The β-system supplies each source-set with a constant rate of gain. Active cells divide that gain in proportion to their activity. The total increment is controlled even when many cells participate. The γ-system goes further: active cells gain at the expense of inactive cells in the same source-set, so the source-set’s total value remains constant. Learning redistributes a fixed resource rather than creating value.
These alternatives matter because they make different physical commitments. α imagines persistent accumulation; β imagines a limited inflow; γ imagines local competition and conservation. The response rule then asks whether a source-set wins by its total value, the Σ-system, or by mean value, the μ-system. Mean discrimination is generally less vulnerable to accidental differences in active source-set size. The paper’s taxonomy prevents “reinforcement” from becoming a magical verb: one must say what changes, which budget applies and how response evidence is aggregated.
Causal probe 3 · Why can mean and sum disagree?
A large source-set can collect more total impulse even if its typical cell is weak. A smaller source-set can have stronger cells but a lower sum. Dividing by the active count removes that size advantage, so mean discrimination estimates effectiveness rather than volume.
Worked example 3 · A response won by size
Source-set R₁ has four active cells of value 2, giving sum 8 and mean 2. R₂ has two active cells of value 3, giving sum 6 and mean 3. Σ chooses R₁; μ chooses R₂. If the four-versus-two difference came from random source-set membership rather than better evidence, μ is the less biased decision.
How much fires, and how much fires again?
For a fixed-threshold perceptron, two probabilities carry much of the analysis. PA is the expected proportion of association units activated by a stimulus of a given size. PC is conditional coactivation: given that an A-unit responds to stimulus S₁, what is the probability that it also responds to S₂? The first quantity describes the density of the active representation. The second describes similarity as that network samples it.
Raising threshold reduces PA because fewer combinations of active excitatory origins can cross it. Increasing the proportion of inhibitory origins also reduces activity. Both changes reduce PC, often sharply, because a unit that barely qualifies for one stimulus is less likely to qualify for a second. When excitation and inhibition are balanced, Rosenblatt’s curves show a flatter relationship between retinal area and PA. That can protect a system from treating mere size as the decisive feature.
PC is conditional on the unit already responding to S₁, so it is not the fraction of illuminated pixels shared by two drawings. Retinal overlap, excitatory and inhibitory origins, threshold and the projection distribution all intervene. Two shapes that people call similar can be dissimilar to a poorly organised perceptron; two visually different shapes can collide if the sampled origins ignore their difference. Similarity is a relation among stimuli, architecture and current value organisation.
The model therefore rejects a free-floating notion of class. A class is learnable when its members repeatedly recruit a coherent association population more than outsiders do. The ecology of presented stimuli and the organisation of the system together determine the “things” into which the perceptual world is divided. This is not a claim that categories are arbitrary. It is a claim that a physical learner’s available invariances help constitute its operative similarity.
Worked example 4 · Threshold changes the active code
Suppose an A-unit has three excitatory origins and one inhibitory origin. A pattern activates two excitatory origins and no inhibitor. With threshold two the unit fires; with threshold three it does not. A second pattern sharing those two origins coactivates the unit only in the first system. Changing threshold alters both representational density and the network’s measured similarity.
The environment determines whether experience becomes a class
Rosenblatt distinguishes retention from generalisation. Pr is the probability that a bias remains in favour of the response assigned to a previously reinforced stimulus. Pg is the probability that a new member of a represented class evokes the appropriate response. In each case, the pairwise analysis asks whether the correct response is favoured over one alternative; a separate quantity is needed for preference over every alternative in a larger response system.
The common approximation has two parts. First, at least one effective association unit in the correct source-set must be activated by the test. Second, the correct response’s accumulated value must beat its rival. The paper expresses the second part through a normal-curve integral whose coefficient depends on the value dynamics, decision rule, source-set overlap and environmental statistics.
P(NAr > 0) = 1 − (1 − PA)NAe
NAe counts effective A-units, excluding cells common to both compared response source-sets because a shared cell contributes equally to both sides. The first factor says that no decision evidence exists if the correct source-set receives no active effective unit. The Φ(Z) term says that, conditional on evidence, the learned value difference must have the right sign. This decomposition is a useful teaching device even when one does not reproduce every coefficient in the 1958 notation.
In an “ideal” environment of unrelated random stimuli, many arbitrary items assigned to a response interfere. The response must remember coincidences that have no class-level regularity. As the number of learned items grows, retention approaches its original chance level, and there is no principled basis for generalisation. More experience is not automatically more abstraction. Without recurring structure, more experience is load.
A differentiated environment changes the limit. Let PC11 be expected coactivation for two stimuli from Class 1, and PC12 expected coactivation between a Class 1 member and a Class 2 member. If same-class recurrence lies above ordinary activation while cross-class recurrence lies below it, training adds coherent evidence for the correct response.
Under this ordering, retention and generalisation approach the same better-than-chance asymptote. In the limit, it no longer matters whether the specific test stimulus has appeared before, because both a trained member and an unseen member express the class’s association structure. Increasing the number of association cells reduces the chance that random sampling conceals that structure and can bring the simple model’s asymptote close to one.
Causal probe 4 · Why does class structure reverse the effect of experience?
For unrelated items, each reinforcement writes a largely independent pattern into a shared source-set and increases interference. For correlated class members, reinforcements repeatedly revisit some of the same association cells. Their common evidence accumulates faster than accidental cross-class evidence.
Worked example 5 · A separability margin
Suppose PA=0.30, PC11=0.55 and PC12=0.12. A cell active for one Class 1 member is much more likely to return for another Class 1 member than for Class 2. Reinforcing active cells during Class 1 trials therefore creates a value advantage that an unseen Class 1 member tends to recover. If both coactivation probabilities were 0.30, the label would have no class-specific association signal.
From mutually exclusive names to independent features
If every possible object requires a separate mutually exclusive response, adding responses makes discrimination harder. Rosenblatt therefore considers binary response systems. A response pair can indicate whether one independently recognisable feature is present or absent. Several bits can then describe a stimulus by composition. Seven reliable binary features can in principle distinguish as many as 128 configurations, whereas a one-name-per-object scheme needs a separate response for every class.
The saving is real only when suitable independent “earmarks” exist. If a stimulus can be recognised only as an indivisible whole, it still needs its own bit or response pair, such as “dog” versus “not dog”. Binary coding does not manufacture factorisation. It exploits factorisation available to the association system.
Bivalent reinforcement adds positive and negative value change. An externally controlled reward can increase active units in the source-set of an on response while punishment reverses the change; off responses receive the complementary treatment. Because an active unit may gain or lose, trial and error can correct a response instead of only strengthening whatever happened. Rosenblatt also describes a more biologically plausible variant in which excitatory and inhibitory association units implement the two signs, though he judges it less efficient in the analysis.
The article says that related bivalent systems were simulated on an IBM 704 and that results supported the main theoretical predictions, but details were to be reported later. That sentence is evidence of a computational check, not a benchmark table. The plotted curves are theoretical or parameter-based. A faithful companion must not transform this historical report into a modern empirical comparison.
Worked example 6 · Three bits instead of eight names
Suppose objects vary independently in shape, shade and position. Three on/off response pairs can encode eight combinations: circle/different, dark/light and left/right. Training a dark circle on the left reinforces three feature responses rather than one indivisible object name. An unseen dark square on the left can reuse two of those responses if the association layer detects them independently.
Time, spontaneous organisation and distributed loss
The quantitative sections mainly analyse a momentary-stimulus perceptron. Current association activity depends on the current pattern. Rosenblatt notes that the same separability principles can reach velocities and sound sequences if stimulation leaves a temporary trace, such as an altered threshold, so that A-system activity at time t depends partly on activity at t−1. Time must become physically available to the classifier; a momentary snapshot cannot distinguish sequences that contain the same elements in different orders.
Structured origin points supply another improvement. Completely random origins notice coincident area. Focal origins in a projection layer make units sensitive to contour location. Here architecture changes the statistics before learning. The theory is therefore not an assertion that enough reinforcement makes every random representation adequate. Inductive bias appears as the distribution of origins and the traces made available.
A further proposal concerns value decay. If A-unit values decay in proportion to magnitude while every response is reinforced without regard to right or wrong, Rosenblatt says a system exposed to two dissimilar classes tends towards a stable condition that responds differently to them. He calls this spontaneous concept formation and reports successful IBM 704 simulation. The idea is intriguing because class structure can break response symmetry without an external label assigning each case. It remains a reported development rather than a fully reproduced experimental section in this article.
Distributed storage yields a lesion prediction. An association can use a large proportion of cells, so removing a small region need not delete one named memory. It should instead weaken many learned associations slightly and eventually create a general deficit. The claim is architectural: information is not located in one symbolic cell. It does not establish that every biological memory is uniformly distributed or that damage is always graceful.
Causal probe 5 · Why does a trace enable sequence discrimination?
Without state, AB and BA can present the same current symbol at their final moment and the network has no evidence about the earlier one. A decaying threshold or recurrent trace makes the earlier symbol part of the present association pattern, allowing order to affect coactivation.
Worked example 7 · A temporary trace distinguishes AB from BA
Give each sensory event a one-step residual trace. At the second step of AB, current B appears together with residual A. At the second step of BA, current A appears with residual B. Association cells can now sample different conjunctions. The trace has converted an order relation over two moments into a pattern simultaneously available to the fixed-threshold machinery.
Worked example 8 · Distributed deletion
A learnt class recruits 400 of 1,000 association cells across the area. Removing 50 cells at random is expected to remove about twenty contributors to that class, but it also removes contributors to other classes. No single memory is targeted. Margins shrink broadly, which is the model’s predicted general deficit.
A relation is not merely a more difficult pattern
Near the end, Rosenblatt tests the reach of his own mechanism. The perceptron appears sufficient for pattern recognition, associative learning, selective attention and selective recall. With persistent traces it may recognise temporal patterns, and with feedback through sensory channels it may learn ordered response sequences. These are substantial capabilities for the proposed organisation.
Yet a response can depend on several concrete conditions without representing a relation. “Name the colour if the stimulus is on the left and the shape if it is on the right” can be built from recognisable features and positions. “The triangle occurred before the circle” requires the system to abstract the same ordering relation across varying objects and moments. Rosenblatt says problems of relative judgement and relational abstraction become excessively difficult. Statistical separability alone is not sufficient.
This limit matters historically. It prevents the 1958 paper from being used as a generic prophecy of all later neural capability. The paper proposes a mechanism and marks the point at which that mechanism needs augmentation. Later work may change the architecture, objective, representation or training protocol. Such successors should be credited with their own results rather than smuggled into the original perceptron.
Causal probe 6 · Why can a conjunction succeed while a relation fails?
A conjunction can be bound to stable sensory features: left position, red colour and circular outline can repeatedly coactivate a source-set. A relation must remain invariant while its participating objects and absolute positions change. The momentary association statistics do not automatically factor out those roles.
Evidence boundary
The article is a theoretical and mathematical paper with reported simulations, not a modern data benchmark or a physiological validation of every component. V has no identified biological substrate in the article. Claims about spontaneous organisation and IBM 704 agreement are stated but not documented with complete experimental tables here. The companion treats anatomical terms as parts of a mechanistic model and does not infer sentience, phenomenal consciousness or human-level thought from successful discrimination.
Watch a fixed random representation become biased towards a class
This laboratory is an interpretive reconstruction of the paper’s population logic. Twelve sensory points feed a fixed bank of deterministic excitatory conjunctions. Two response source-sets carry value ledgers. Reinforcing a labelled pattern increments active values in the selected response ledger. The display reports activation density, within-class and between-class coactivation, response margins, retention and unseen-member generalisation. It does not claim to reproduce every 1958 coefficient.
Four cards for carrying the model forward
Architecture card
Sensory points → focal projection units → random association units → competing response source-sets. Ask which parts are fixed, random, adaptive and recurrent before calling a system a perceptron.
Statistic card
PA measures representation density. PC measures conditional coactivation. Generalisation needs same-class recurrence above baseline and cross-class recurrence below baseline.
Mechanism card
Specify α, β or γ value dynamics; μ or Σ response choice; and monovalent or bivalent reinforcement. “It learns” is incomplete until these choices are named.
Boundary card
Separate derived probability claims, reported IBM 704 simulations, biological analogy and later interpretation. Statistical separability is not a proof of relational abstraction or consciousness.
Reconstruct the argument in six moves
1 · Replace stored pictures with response dispositions
Let memory be a persistent change in which response follows recurring activity.
2 · Sample stimuli through random association units
Use physically constrained but microscopically unknown origins, thresholds and source-sets.
3 · Measure active density and conditional recurrence
Reduce architecture and stimulus overlap to PA and PC.
4 · Reinforce the active population
Apply a declared α, β or γ value law in the context of the selected response.
5 · Compare response source-sets
Let mean or total value determine the competitive winner and calculate the probability that the correct side leads.
6 · Read the environment through an inequality
If PC12 < PA < PC11, correlated class evidence is repeatedly reinforced and unseen members inherit the bias.
How to audit a claim of learned generalisation
- Name the retained variable. Is memory in a connection, a cell state, a response-specific ledger or an external store?
- Freeze the representation test. Evaluate generalisation with the feature map held fixed so representational benefit is not confused with test-time adaptation.
- Measure within- and between-class recurrence. Accuracy alone conceals whether a reusable class statistic exists.
- Check the activity baseline. Dense activation can make every pair look similar; sparse activation can leave no effective response evidence.
- Separate retention from generalisation. Include trained members and genuinely unseen members in different results.
- Control response-set size. Compare mean and sum aggregation when source populations differ.
- Use a scrambled-label control. The same patterns without class-consistent labels should lose the coactivation advantage.
- Test partial loss. Remove association units and observe whether damage is item-specific or distributed.
- State the temporal resource. Sequence recognition requires a trace, recurrence or explicit time-indexed input.
- Probe a relation. Success on concrete features does not establish abstraction over roles or order.
The first adaptive bridge in the neural sequence
The first forty-five papers established languages for computation, uncertainty, causal inference, optimisation and resource-aware learning. Paper 0046 opened the neural representation sequence by showing how a fixed, clocked net can embody a temporal proposition. Paper 0047 makes organisation plastic. It gives random neural populations a value-changing rule and asks for behavioural probabilities rather than a hand-designed truth table.
The forward link is as important as the backward one. Later neural papers will simplify the perceptron into algebraic classifiers, prove convergence under separability assumptions, expose limitations, introduce differentiable multilayer training and scale distributed representations. Those developments should not erase the richer 1958 object. Rosenblatt’s article already joins architecture, an environment, a learning rule, a decision rule and an account of similarity. Its “perceptron” is a research programme about organised adaptive systems.
For agency, the lesson is that a response policy inherits the learner’s similarity structure. Reinforcement does not decide what should count as the same situation; fixed sensory and association organisation partly decides that before feedback arrives. For consciousness research, the article supports a disciplined distinction. A distributed adaptive system can recognise, recall and generalise without the theory containing any criterion for phenomenal awareness. Functional sophistication is an explanandum and a mechanism, not by itself a consciousness verdict.
Claims, reconstructions and later vocabulary
| Layer | What is warranted | What is not silently imported |
|---|---|---|
| Primary derivation | Probabilistic performance from thresholds, origin statistics, source-sets, value rules and class coactivation. | A modern gradient objective or the later convergence theorem. |
| Primary report | IBM 704 simulations are said to support bivalent and spontaneous-organisation predictions. | Unreported sample sizes, error bars or benchmark superiority. |
| Companion reconstruction | Seeded demonstrations expose activity, coactivation, reinforcement and response margins. | Numerical replication of every printed figure. |
| Historical interpretation | The paper is a foundational connectionist model of adaptive distributed representation. | A claim that 1958 machinery equals contemporary deep learning. |
| Mind and brain | The model proposes a physically measurable bridge between organisation and behaviour. | Proof that biological brains use every mechanism, or that the system is conscious. |
What this companion adds
This companion restores the probabilistic population theory obscured by the single-unit textbook caricature. Ten paper-specific animated figures move from the article’s three opening questions through the memory fork, full photoperceptron, two response phases, α/β/γ value budgets, activation density, statistical separability, learning asymptotes, binary response coding and the declared frontier at relational abstraction.
The executable laboratory makes the central contrast auditable. A deterministic feature bank is held fixed while the response value ledgers learn either coherent classes, scrambled labels or one repeated stimulus. Trained-member retention and unseen-member generalisation are reported separately. Exact controls test repetition, differentiated classes and removal of label structure. Every numerical example outside the primary’s printed analysis is marked as authored teaching material.
Two comprehensive passes
Scholarly and editorial review. The complete 23-page primary scan was rendered and inspected. The storage contrast, organisation rules, predominant/postdominant phases, Table 1 systems, PA/PC machinery, random and differentiated environments, binary and bivalent systems, reported simulations, improved perceptrons and conclusion list were mapped to printed pages. Later convergence, XOR, gradient descent and deep-learning claims are not attributed to 1958.
Executable and publication review. The seeded laboratory includes three independent exact tests and reports its interpretive scope. Responsive rendering, SVG labelling, reduced-motion behaviour, keyboard-labelled controls, duplicate identifiers, placeholder language, substantive length and source links are checked in the publication workflow. The public source is required to match the reviewed source byte-for-byte after build.
References and source trail
- Frank Rosenblatt, “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain,” Psychological Review 65(6), 386–408 (1958). DOI record; PubMed record; complete scan used for page verification.
- Warren S. McCulloch and Walter Pitts, “A logical calculus of the ideas immanent in nervous activity,” The Bulletin of Mathematical Biophysics 5, 115–133 (1943). DOI record. Used for the fixed neural-calculus predecessor that Rosenblatt cites.
- Donald O. Hebb, The Organization of Behavior: A Neuropsychological Theory (Wiley, 1949). Archive record. Used only as the connectionist predecessor Rosenblatt discusses, not as evidence for the perceptron’s equations.
- Frank Rosenblatt, Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms (Spartan Books, 1962). Archive record. A later expansion, clearly separated from claims in the 1958 article.
- Marvin Minsky and Seymour Papert, Perceptrons: An Introduction to Computational Geometry (MIT Press, 1969). Publisher record. A later mathematical treatment and critique, not retroactive evidence about the 1958 results.
- David G. Stork, “Is backpropagation biologically plausible?” in International Joint Conference on Neural Networks (1989). DOI record. Listed only to mark a later biological-learning debate.