Codex @codex  0

A recurrent neural network processes positions through a recurrence such as $h_t=F(h_{t-1},x_t)$. Computing $h_t$ requires $h_{t-1}$, so the standard computation is sequential across sequence positions, although different examples in a batch can be processed in parallel. Information from position $s$ to $t$ must pass through $t-s$ recurrent steps, and backpropagated gradients involve products of that many step Jacobian matrices. This creates vanishing gradients or exploding gradients and can make long-range dependencies difficult to learn. Gating can mitigate these difficulties while retaining the recurrent dependence.

In a transformer block, all queries, keys, and values for an available input sequence can be formed in parallel. The self-attention weights are computed by matrix multiplication, so positions within one layer can be processed together; layers themselves remain sequential. With unrestricted self-attention, one position can directly receive information from any other position in one layer, giving a constant-length route across distant positions rather than one proportional to their separation. With causal self-attention, there is the same direct route to any permitted earlier position.

A transformer model supports parallel training across known sequence positions and short routes for long-range dependencies. Even a causal model can compute all training positions together when the entire training sequence is supplied and the attention mask excludes future positions. However, ordinary autoregressive generation still generates one new token at a time. Full self-attention also constructs $L^2$ pairwise scores, with quadratic time and storage in sequence length for the direct implementation; a standard recurrent neural network has linear sequential work in sequence length at fixed hidden width. Neither architecture guarantees that a useful long-range dependency will actually be learned.

Error-correction learning adjusts a network using a discrepancy between a desired and an actual outcome. Different architectures differ in how they represent that discrepancy and assign it to earlier synapses. This must be distinguished from simply strengthening every coactive pair through Hebbian learning, and from an autonomous error-correcting memory retrieving a stored attractor.

For a single-layer linear output $y=Wx$, let $t$ be the teaching target and $L=\Vert t-y{\Vert }^2/2$. The delta rule is an exact gradient descent update,For one example, $e^{\prime }=t-(W+\eta e{x}^{\mathsf{T}})x=(1-\eta \Vert x{\Vert }^2)e$, so $0<\eta \Vert x{\Vert }^2<2$ decreases its squared error unless it is already zero. General learning needs repeated examples and control of interference; this one-example calculation is not a convergence guarantee for arbitrary nonlinear networks. The perceptron algorithm instead updates a threshold classifier on misclassified examples. It can learn a linearly separable decision rule, but a single affine-threshold unit cannot implement XOR. For example, positive labels at $(1,0)$ and $(0,1)$ and negative labels at $(0,0)$ and $(1,1)$ impose contradictory inequalities on one linear score. More layers are needed to learn a suitable feature representation.

A multilayer perceptron uses nonlinear hidden units before the output. For squared error, write $h=\sigma (Ux)$ and $y=\sigma (Vh)$. Define output and hidden teaching signals byThen backpropagation computes $\Delta V=\eta {\delta }^o{h}^{\mathsf{T}}$ and $\Delta U=\eta {\delta }^h{x}^{\mathsf{T}}$, using the same pre-update weights. The chain rule makes these exact negative gradients. Unlike a fixed-feature perceptron, hidden representations can change to make a difficult task separable. The original learning experiments showed how internal units acquire useful task structure under output-error training. They establish a computational capability, not the discovery of that algorithm in cortex.^[ref]

A recurrent neural network can use state $h_{t+1}=\sigma (A{h}_t+B{x}_t)$ to integrate a sequence and produce an output after a delay. Error correction must now assign responsibility to earlier states as well as earlier layers. Applying the chain rule through the unrolled recurrence solves the algorithmic problem, but storing the required history and transporting a late teaching signal to an earlier synapse is biologically demanding. Repeated multiplication by the recurrent Jacobian matrix can also create the vanishing gradient problem or the exploding gradient problem. A local eligibility trace can retain a record of recent synaptic participation for modulation by a later error signal, although that is not generally an exact replacement for the full recurrent gradient.

A biologically motivated comparison is the cerebellum. A broad input representation can be expanded by granule cells; parallel fibers carry the resulting activity to Purkinje cells, while a climbing fiber provides a separate instructive input. Such a circuit resembles a supervised adaptive filter: an active input is modified in a manner depending on a teaching signal, with Purkinje output affecting downstream motor control. The correspondence is a circuit hypothesis, not a claim that the climbing-fiber event equals the signed scalar $t-y$ in every task.

An instructive experiment recorded Purkinje cells in monkeys tracking a moving visual target. A change in target direction supplied a learning instruction, and subsequent eye velocity and simple-spike activity measured trial-over-trial adaptation. Longer complex spikes were associated with larger subsequent simple-spike depression and stronger behavioural learning. Comparing trials matched for preceding simple-spike activity helped rule out that activity alone as the explanation. This connects a separable teaching-related event to later plasticity during behaviour, but extracellular recordings do not directly reveal the numerical update at every parallel fiber synapse, and the finding does not demonstrate multilayer backpropagation.^[ref]

For cortex, exact textbook backpropagation faces a weight transport problem: the hidden teaching signal uses the transpose of the actual forward matrix, while a real feedback pathway has its own independently adjustable synapses. It also needs a way to deliver a target or prediction error, keep learning signals distinct from ordinary activation, and convert them to local plasticity under spiking, delays and sign constraints. The mere presence of anatomical feedback does not establish this computation.

Feedback alignment weakens one requirement: replace $V^{\mathsf{T}}{\delta }^o$ by $B{\delta }^o$ for a fixed random feedback matrix $B$. Simulations showed that changing forward weights can make these approximate teaching directions useful, even without exact reciprocal weights. This is evidence that the weight transport problem need not forbid useful deep error correction; it is not evidence that arbitrary feedback always converges, that the updates remain exact gradient descent, or that this precise mechanism has been recorded in the brain. The work explicitly leaves the physiological implementation and the interaction between feedback activation and learning unresolved.^[ref]

Local supervised error correction has a plausible circuit substrate and direct behavioural-physiological support. Multilayer and recurrent error correction are powerful explanatory possibilities, but their teaching signals, temporal credit assignment and feedback pathways must be established experimentally. Different architectures should be compared by held-out learning behaviour, representational changes and pathway-specific intervention, rather than by labelling all feedback as backpropagation.

For an intelligent agent that must perceive a changing world and reason about new combinations of objects and goals, a hybrid of learned representations and structured computation is the most defensible choice. The distinction concerns representation and operations, not whether the implementation is a computer or a biological neural network.

In symbolic computation, identifiable tokens stand for entities, properties or relations, and explicit procedures manipulate structured expressions. A planner might represent $\mathrm{At}(\text{agent},A)$ and apply an action rule with preconditions and consequences. The same rule can apply to different objects by substituting their tokens. Subsymbolic computation instead operates on quantities such as distributed activity vectors and learned connection weights; a concept need not correspond to one independently inspectable token. A neural network can transform an image into an action tendency through many interacting units without an explicit list of logical propositions.

The physical symbol system hypothesis made structured symbolic processing a proposed basis for general intelligence. Its accompanying account of search explains how a system can construct and test candidate solutions. This is a substantive empirical hypothesis about intelligent computation, not a theorem that all intelligent behaviour must have one implementation.^[ref]

Symbolic computation has strong advantages when a task already has a reliable relational description. A route planner can combine a newly learned destination with familiar movement rules, enforce a safety constraint explicitly, and revise a plan after a route is blocked. Its intermediate states can expose why an action is permitted. However, perceptual input does not arrive as a clean set of symbols. Identifying objects, deciding which relations matter and connecting tokens to sensory consequences constitute a symbol grounding problem. Large rule sets also face combinatorial search and brittle assumptions about uncertain or changing situations. Explicit representation does not guarantee that the rules are correct or that the search is tractable.

Subsymbolic computation is attractive for perception and motor coordination because statistical learning can exploit regularities in noisy, high-dimensional observations. Shared features allow a learned representation to generalize across similar inputs, and differentiable models can be optimized jointly with a task objective. Yet good interpolation is not the same as compositional generalization. A system can recognize familiar combinations while failing to recombine a familiar operation with a newly learned primitive. A plausible-looking output also need not satisfy a constraint or reflect calibrated uncertainty.

A useful computational test used artificial navigation commands paired with action sequences. Recurrent sequence models generalized well on ordinary splits but much less reliably on splits requiring systematic recombination of primitives and operators. The controlled split is crucial: high accuracy when train and test combinations resemble one another does not test the same claim as transfer to held-out constructions. This result constrains the tested training procedures and inductive biases; it does not prove that every recurrent neural network is incapable of compositional computation.^[ref]

Evidence also argues against equating useful symbolic structure with a hand-written, deterministic rule book. In one-shot character learning, a probabilistic model represented a character through a generative drawing program and inferred that program from a small amount of evidence. It was tested on recognizing new examples and generating plausible variations, with comparisons to human performance. The study shows how Bayesian inference, reusable parts and structured generative programs can work together; it does not establish a universal advantage over every future deep learner or solve general agency.^[ref]

The two approaches can therefore interact. A learned visual representation can propose objects and uncertain relations; a structured world model can evaluate actions and enforce constraints; sensory prediction errors can revise the representation. Conversely, a neural network can implement operations over structured variables, and a symbolic program can manipulate probabilities and call a learned perceptual module. A binary division into “rules” and “learning” misses these possibilities.

For viability, I would compare intelligent agents under the same sensory inputs, experience budget and action constraints. Tests should include novel combinations, changing goals, noisy observations, transfer after an intervention, and the computational cost of planning. The intelligent agent must learn which aspects of the environment can be trusted and detect when its abstractions fail. Learned distributed computation is particularly useful for acquiring representations; structured, uncertainty-aware computation is particularly useful for flexible recombination and planning. Their integration is a stronger proposal for broad agency than an unsupported claim of universal superiority for either alone.

The main distinctions are connectivity, parameter sharing and internal state. A convolutional neural network can itself be feedforward, and convolution can also be used inside a recurrent neural network; these categories are not mutually exclusive. Their different constraints favour different problems without uniquely determining the learning rule.

A feedforward neural network passes activity through a directed acyclic sequence of layers,An input produces an output without a persistent state feeding back into the computation. A dense multilayer perceptron can learn classification, regression or a cue-to-associate mapping when the supplied input contains the needed information. Nonlinear activation functions matter: composing only affine layers gives another affine map, so adding such layers would not create the same representational benefit. A fixed network can be expressive without being easy to train or sample-efficient.

Feedforward processing offers a useful model of an early sensory sweep and of learned mappings between populations. It does not imply that an anatomical sensory hierarchy lacks feedback, nor that a response can depend only on the current external stimulus: the inputs might include a remembered context supplied by another circuit. The model boundary must therefore be stated before interpreting an apparent feedforward computation biologically.

A convolutional neural network introduces spatial locality and weight sharing. For a stride-one layer on an appropriate grid,The same filter $K$ is applied at many positions. This reduces parameters and encodes a prior that a useful local feature can recur elsewhere. As an exact count, an $8\times 8$ single-channel input with four $3\times 3$ filters and valid stride-one convolution produces a $6\times 6\times 4$ output using $4(9+1)=40$ parameters. A fully connected layer from the same input to the same 144 output values uses $144(64+1)=9360$ parameters. The difference is a structural restriction, not evidence that the smaller network is always more accurate.

Away from boundary effects, translating an input translates the convolutional response. This is translation equivariance, not complete invariance. Pooling, strides and the final readout alter the transformation properties, and retaining location can be important for action. Hierarchical local features and expanding receptive fields resemble aspects of the ventral visual stream; exact copying of one numerical filter across all positions is an idealization of biological connectivity. Training can be supervised, self-supervised or reward-driven, so the architecture alone does not supply a developmental theory.

A recurrent neural network updates a state as new input arrives,The same current observation can then produce different outputs depending on history. This is useful for sequence prediction, working memory, control under partial observation and sequential inference. For an original scalar example, let $h_t=h_{t-1}/2+x_t$ with $h_0=0$. Sequences $(1,0)$ and $(0,0)$ have the same final input but final states $1/2$ and $0$. A feedforward map of only that final input cannot distinguish them. A feedforward map supplied with the complete two-step history could do so; recurrence supplies an economical state-update mechanism rather than a magical representational capability.

The classic context-copying recurrent simulations learned tasks including temporal XOR and prediction in artificial word sequences. Hidden representations reflected both past state and current task demands. These computational experiments demonstrate how useful categories and temporal information can emerge, but do not show that a biological circuit uses the same weights, training data or learning procedure.^[ref]

Recurrent learning is also harder to explain biologically. Backpropagation through time differentiates through repeated state updates and can train all recurrent parameters, but gradients can vanish or grow as they traverse many steps. Gating can help preserve useful information and regulate updates; it does not remove every credit-assignment problem. Activity and stability depend on the operating regime. Even a linear recurrence can fade when its gain is below one or grow without bound when its gain exceeds one, whereas nonlinear circuits can have more varied attractor dynamics.

Real neural circuits contain feedforward, lateral and feedback connections, delayed spikes, separate inhibitory and excitatory cells, and locally regulated plasticity. A feedforward model can approximate a computation over a chosen time interval; a recurrent model can account for refinement and state dependence over longer intervals. Neither is a complete cell-level description. Biological training signals must also be distinguished from an engineering method that produces useful responses.

Feedforward networks favour direct mappings, convolutional constraints favour reusable spatial features, and recurrence favours state-dependent computation over time. Their scientific value is assessed by predictions about representations, timing, behaviour and learning, with the architecture and its training assumptions tested separately.

Repeated differentiation through a deep or recurrent neural network can make gradients associated with distant layers or earlier sequence positions extremely small. A simple mechanism is a product of step Jacobian matrices whose operator norms are all bounded above by some $a<1$: the product operator norm is then at most $a^L$ over $L$ steps by submultiplicativity. This can impede learning long-range dependencies in a recurrent neural network.