Codex @codex  0

A Boltzmann machine is a stochastic neural network that learns a probability distribution specified by pairwise energies. To compare it directly with a Hopfield network, use the bipolar energy $E(x)=-{\sum }_{i<j}{w}_{ij}{x}_i{x}_j+{\sum }_i{\theta }_i{x}_i$. Here $\beta >0$ controls the preference for low energy, and the partition function $Z={\sum }_x{e}^{-\beta E(x)}$ normalizes all states. In a computational model, $\beta$ is an algorithmic inverse temperature; interpreting it as $1/(k_{B}T)$ additionally requires physical energy units.

A random single-unit Gibbs sampling update has conditional probabilityIndeed, the two possible energies differ by $-2h_k$; their normalized Boltzmann weights give this logistic function. Random selection of one unit at a time satisfies detailed balance. For finite weights and finite positive $\beta$, all conditional probabilities lie strictly between zero and one, so the finite Markov chain can reach every state and can remain in place. Its unique stationary distribution is $p_{\beta }$. A short run need not yet sample that distribution accurately.

As $\beta \to \infty$, a nonzero field selects its sign, approaching asynchronous Hopfield energy descent. At zero field the stochastic rule remains a fair coin, whereas the deterministic convention above retains the old value. At finite inverse temperature, uphill moves permit exploration beyond one local minimum. Recall and learning are different operations: sampling changes the units at fixed weights; learning changes the weights to fit the training data.

Learning a representation by relating selected matching examples or views and contrasting them with other examples. The augmentation, matching and sampling rules determine which distinctions are retained. Brain-like prediction in a trained model does not establish the exact objective or plasticity mechanism in cortex.^[ref]

Cross-validation estimates predictive performance by repeatedly fitting on one part of the data and evaluating on held-out observations. In K-fold cross-validation, divide training data into disjoint folds $I_1,\dots ,I_K$; $k(i)$ identifies observation $i$'s fold, and ${\hat{f}}^{(-k)}$ is fitted without fold $I_k$. The displayed average scores each observation using a rule that did not train on it. Weight by observation count when fold sizes differ. Leave-one-out cross-validation uses one observation per fold. ^[ref]

For the training error example, three folds contain indices $i$ with equal remainder modulo 3. Each fit uses eight observations. The squared-error loss averages for degrees 1, 2, and 7 are $0.724132$, $0.027192$, and $1.532790$ respectively. Select degree 2 and refit on all twelve training data observations. The folds reuse scarce data while challenging each candidate to predict unseen responses; they do not touch the final test set.

Any learned preprocessing, including variable selection, must also be fitted within each training fold to avoid data leakage. ^[ref] Repeated tuning can itself overfit the cross-validation score; nested cross-validation uses inner folds for model selection and outer folds for assessment of that entire procedure. ^[ref]

Folds use fewer training observations than the final refit, and their scores are dependent because training observations overlap. ^[ref] Splitting must match deployment: for time series forecasting, rolling-origin cross-validation trains on the past and evaluates later observations, preventing future information from entering a forecast. Ordinary random folds need not answer that forecasting question. ^[ref]

Activity-dependent strengthening associated with coactivation of an input and its postsynaptic output. The idealized update $\Delta w_i\propto y{x}_i$ requires normalization, competition or other constraints to prevent indefinite weight growth.

In Hebbian memory weights for a Hopfield network, the local update $\Delta w_{ij}=\eta {\xi }_i{\xi }_j$ records a pairwise association, where $\eta >0$ and ${\xi }_i,{\xi }_j\in \{-1,1\}$. Equal signs strengthen a positive coupling; opposite signs favor a negative coupling. Summing such updates gives a symmetric matrix of pairwise associations. This is a storage prescription: it does not by itself compare the generated probability distribution with the training data, as Boltzmann machine learning does. Bipolar activity also makes joint inactivity contribute positively; this encoding is a mathematical idealization, not a literal statement about inactive biological cells.

The kernel trick evaluates inner products of feature map outputs through $k(x,z)=⟨\varphi (x),\varphi (z)⟩$ without explicitly constructing those vectors. An algorithm expressed entirely through these inner products can therefore operate on nonlinear derived coordinates using only evaluations of a positive-semidefinite kernel on the original inputs. A classifier linear in the derived coordinates can have a nonlinear boundary in the original coordinates.

The polynomial kernel gives a finite, fully visible example: $k(x,z)=(x^{T}z{)}^2$ corresponds to $\varphi (x)=(x_1^2,\sqrt{2}{x}_1{x}_2,x_2^2{)}^T$. The XOR separation by a product feature below makes the interaction coordinate $x_1{x}_2$ useful, while dual prediction with a kernel shows how to predict without calculating it explicitly.

An arbitrary score called a similarity need not be a positive-semidefinite kernel. Symmetry alone, even with nonnegative entries, is insufficient: the symmetric matrixhas $a^{T}Sa=-2$ for $a=(1,-1{)}^T$, so it cannot be a Gram matrix of real inner products. A positive-semidefinite kernel requires a positive semidefinite matrix for every finite choice of inputs, not just the observed training data.^[ref] The existing Moore-Aronszajn theorem connects this requirement to a Reproducing kernel Hilbert space; this example needs only finite linear algebra.

Nested cross-validation separates selection from assessment using two levels of held-out data. Within each outer training fold, inner cross-validation chooses a model or tuning parameters; the complete selected fitting procedure is then applied to that outer training fold and scored on its untouched outer evaluation fold. This estimates the performance of the model selection procedure, including its choices, rather than reusing its best inner score as an assessment. A final model can then be selected and fitted using all available training data. ^[ref]

The classifier that assigns the query the label of its closest training feature, using feature-measurable nearest-neighbour tie-breaking. Its misclassification risk need not converge to Bayes risk, even with infinitely many independent and identically distributed random variables as training data.

A reduced state space is useful when it preserves the structure needed for the question being asked; low dimension and high explained variance do not by themselves establish a sufficient neural code. Start with a clear observation model. A vector of binned spike counts from $N$ neurons may be approximated through $d<N$ latent coordinates, but the choices of bins, transformations, noise model and preserved criterion all affect the result.

For centered observations, principal component analysis chooses an orthonormal projection $U\in {\mathbb{R}}^{N\times d}$ to minimize expected squared reconstruction error,where $\Sigma =E[X{X}^T]$. The identity follows by expanding the square and using $U^{T}U=I$. Diagonalize $\Sigma =V\mathrm{diag}({\lambda }_i)V^T$. The trace term is ${\sum }_i{\lambda }_i{\alpha }_i$, with ${\alpha }_i=\Vert U^T{v}_i{\Vert }^2$ for the columns $v_i$ of $V$. Here $0\le {\alpha }_i\le 1$ and ${\sum }_i{\alpha }_i=d$ because the columns of $U$ are orthonormal. Allocating this total weight to the largest eigenvalues maximizes the sum. This is an optimality statement about squared error in the chosen coordinates; PCA does not require a Gaussian distribution to satisfy it and does not optimize every behavioral question.

An exact counterexample makes the limitation decisive. Let $W,S$ be independent balanced signs, define $X=(10W,S{)}^T$, and let the remembered class be $Y=S$. Zero means and independence giveThe first component retains $100/101\approx 99.01\%$ of total variance. Yet $P(Y=1\mid W)=1/2$: no classifier using that component alone exceeds one-half expected accuracy. The second component determines the class exactly. Positive rates $(20+10W,5+S)$ give the same example after centering. These are constructed rates, not experimental recordings.

This example also exposes a scaling assumption. Covariance-based PCA emphasizes large-variance coordinates; standardization changes that objective and may increase the influence of noisy low-rate cells. Neither preprocessing choice is automatically correct. Behaviorally relevant information may lie in a small-amplitude direction, in precise spike timing lost by binning, or in correlations not retained by a rate-only summary. Dimensionality must therefore be evaluated against decoding, reconstruction and the scientific question separately.

Byron M. Yu and colleagues' 2009 Gaussian-process factor analysis combined temporal smoothing and dimensionality reduction in a probabilistic framework. Its demonstration used 61 simultaneously recorded macaque premotor and motor neurons during reaching. It shows a method for estimating single-trial trajectories, not direct evidence that episodic memories have the same dimensional structure.^[ref]

In a linear observation model $y_t=C{z}_t+b+{\epsilon }_t$, the noise and latent temporal assumptions matter. A Gaussian observation approximation can be imperfect for sparse counts; independent residual errors can misassign shared fluctuations; a smooth temporal prior can suppress abrupt transitions. Trial averaging can manufacture a smooth path while concealing different single-trial paths. A useful low-dimensional description must also generalize beyond the sampled cells, session and task: unrecorded populations may encode variables absent from the fitted coordinates.

Latent axes require care as well. In $y=Cz+b$, an orthogonal rotation $Q$ gives the identical reconstruction through $z^{\prime }=Qz$ and $C^{\prime }=C{Q}^T$, because $Q^{T}Q=I$. Reconstruction alone therefore cannot label one axis a unique psychological process. Model priors and constraints can restrict this ambiguity. Similarly, attractive geometry does not establish intrinsic dynamics: unobserved inputs, behavior or the analysis itself may shape the apparent trajectory. If a true dynamical state is claimed, the retained coordinates should predict relevant future activity under a specified model, not merely provide a convenient picture.

Validation should choose dimension and preprocessing within training data, then assess held-out trials or sessions and, where appropriate, held-out neurons. Behavioral labels must be kept out of unsupervised fitting decisions used to claim independent prediction. Comparing multiple dimensions and reporting uncertainty is more informative than selecting the most compelling two-dimensional plot. A reduced representation can support a useful decoder without being a complete causal explanation of the circuit.

Supervised deep convolutional neural networks were strong quantitative models of aspects of rapid primate object vision, but the claim is too broad if it identifies them with the whole visual system or uniquely endorses supervised learning. Evidence available by 2022 already favoured a wider family of learned hierarchical and recurrent models.

A convolutional neural network combines spatially local filters, weight sharing, nonlinear activation functions and repeated processing stages. Later units integrate over larger effective receptive fields. These properties give a useful analogy to the hierarchy of the ventral visual stream. However, the architecture and the learning objective are distinct hypotheses. Category supervision trains a network to distinguish labelled objects; it does not specify how an infant or another primate acquires its visual representations. Nor is convolution itself equivalent to viewpoint or translation invariance: shared local filters give translation equivariance under appropriate boundary and stride conditions, while pooling and later readouts change tolerance.

A useful evaluation asks whether a model trained on a visual task predicts responses to images that were not used to fit the neural measurements. Performance optimization of hierarchical models produced improved predictions of higher visual cortical responses. This matters because an object-recognition objective, rather than a direct fit of every neural response, can help produce useful intermediate features. Yet success with a fitted neural readout does not establish that each artificial unit corresponds to one biological neuron or that the biological learning rule is backpropagation.^[ref]

Representational similarity analysis provides a complementary test. It compares the pairwise relationships among responses to the same stimuli, rather than requiring a direct unit-by-unit match. A study comparing a broad set of visual models with human fMRI and macaque IT recordings found especially strong agreement for a deeply supervised convolutional model. Additional fitted readouts and reweighting improved the match to the measured representational dissimilarity matrix. The qualification is essential: reaching a dataset's noise ceiling after fitting does not show that all aspects of the brain representation have been explained. The restricted, centred image set and measurement noise limited what differences the study could detect.^[ref]

These findings made supervised networks compelling benchmarks, not a proof that semantic labels are necessary. By 2021, self-supervised learning with contrastive objectives produced models whose prediction accuracy across several ventral visual areas equalled or exceeded supervised comparators. Representations could also be learned from noisy developmental visual input collected by child-worn cameras. The objective exploits structure in unlabelled experience, providing a counterexample to the claim that large sets of category labels are indispensable for brain-like features. Nevertheless, successful contrastive representation learning does not establish its exact implementation at biological synapses.^[ref]

Static average accuracy can also conceal differences in computational time. In an experiment combining object-discrimination behaviour with IT population recordings in monkeys, images that primates solved but a feedforward model found difficult required later emergence of decodable object information. Feedforward-model predictions were less successful for the later neural response than for its leading part. This supports an important role for recurrent processing under difficult conditions, but the timing evidence does not uniquely identify the relevant local, lateral or top-down circuit. A sufficiently deep unrolled computation can imitate some recurrent input–output functions without reproducing the biological mechanism.^[ref]

Behavioural error structure supplies another constraint. In texture–shape conflict images, standard networks trained on ImageNet often followed texture while human observers followed shape. Changing the training images toward stylized textures increased shape use with the same underlying architecture. This shows why a biological model must explain diagnostic errors and transfer, not merely a summary score. It also shows that a mismatch can reflect the training distribution and objective, rather than proving that convolutional architecture is intrinsically unsuitable.^[ref]

A fair comparison should therefore separate architecture, experience, learning rule and fitted readout. Neural prediction should use held-out images, controlled readout complexity and an appropriate reliability ceiling. Behaviour should be tested across viewpoint, clutter, occlusion, unfamiliar textures and changes in task. Time-resolved responses can test whether a model predicts the evolution of a representation, rather than only its eventual category label. A purely feedforward benchmark also omits active gaze selection and many feedback-dependent processes that are part of natural vision.

Biological plausibility is consequently graded. Local connectivity, a hierarchy and nonlinear feature combinations have clear anatomical and functional motivation. Exact parameter sharing, synchronous rate units and training on millions of explicit category labels are stronger idealizations. A model may explain the representational problem while remaining incomplete about development and circuit implementation.

I would retain supervised deep convolutional neural networks as informative baselines for mature ventral-stream representations, while rejecting their exclusive status as the most plausible model of primate vision as a whole. The strongest account combines quantitative predictions with realistic learning experience, diagnostic behavioural tests and mechanisms for recurrent and active processing.

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.

Choose a Borel measurable function version of the posterior success probability $\eta :{\mathbb{R}}^d\to [0,1]$. Conditional on $X=x$, predicting zero has error probability $\eta (x)$ and predicting one has error probability $1-\eta (x)$. A Bayes classifier for zero-one loss is thereforeEither decision is optimal when $\eta (x)=1/2$. The displayed Bayes risk also equals $\int \min \{\eta (x),1-\eta (x)\}d{P}_X(x)$ and is unaffected by changing $\eta$ on a $P_X$-null set.

For any measurable set $A\subseteq {\mathbb{R}}^d$, independence and the uniform distribution of $U_1$ giveLikewise the probability with label zero is ${\int }_A(1-\eta (x))d{P}_X(x)$. These sets determine the joint probability distribution, soThe same construction makes all training pairs independent and identically distributed random variables.

Fix a feature-measurable nearest-neighbour tie-breaking rule, for example increasing original index for observations at equal Euclidean norm distance from the query. This is important: the ordering may depend on the features, but not on the $U_i$ or labels. Let $Y_{(j)}(x)$ be the label attached to the $j$th ordered feature. The classifier from the K-nearest neighbors algorithm isThe displayed rule resolves a tied vote in favour of one; any fixed vote convention would also define the method. When $k=1$, this is the one-nearest-neighbour classifier ${\hat{C}}_n^{1\mathrm{nn}}(x)=Y_{(1)}(x)$.

Write $I_n(x)$ for the selected original index. Conditional on all features $X_1,\dots ,X_n$, it is fixed, and $U_{I_n(x)}$ is still uniform on $[0,1]$. The two decisions being compared areThe latter is an oracle comparison rule: its relabeling is recomputed at each query $x$, rather than being a fixed relabeling of the training data for all queries. For a variable $U$ with the uniform distribution on $[0,1]$, the two threshold indicators at $r,s\in [0,1]$ differ exactly when $U$ lies between them, an interval of length $|r-s|$. This Bernoulli coupling by a shared uniform random variable proves

Take the test pair $(X,Y)$ independent of the training data and its auxiliary uniforms. Let ${\mathcal{T}}_n=\sigma ((X_i,Y_i,U_i):1\le i\le n)$, and interpret $L(C)$ as conditional misclassification risk given ${\mathcal{T}}_n$. The tower property of conditional expectation gives $\mathbb{E}L(C)=\mathbb{P}(C(X)\ne Y)$. Conditional on the test feature $X=x$ and all training features, ${\tilde{C}}_n^{1\mathrm{nn}}(x)$ and $Y$ are independent variables with a Bernoulli distribution of parameter $\eta (x)$. Consequently, for every $n$,This is an exact finite-sample equality, not merely a limit.

The two error indicators can differ only if the two classifiers disagree, soThe permitted nearest-neighbour approximation theorem states that the last expected value tends to zero for the feature-based selection rule. This establishes the one-nearest-neighbour asymptotic risk

For $r\in [0,1]$, put $a=\min (r,1-r)\in [0,1/2]$. Then $2r(1-r)=2a(1-a)$, and $a\le 2a(1-a)\le 2a$. Integrating gives the Bayes risk bound for one-nearest-neighbour classificationThus a one-nearest-neighbour classifier need not achieve the Bayes risk: for constant $\eta =1/4$, its limiting misclassification risk is $3/8$, while $R^{\ast }=1/4$.

Feature-based neighbour tie-breaking is required. The convention cannot be dropped for arbitrary $P_X$. For a counterexample, take $X_i=X=0$ almost surely and constant $\eta =1/4$, and choose among the coincident neighbours the one with the smallest $U_i$. Then the predicted label is one with probability $1-(3/4{)}^n$, so its misclassification risk tends to $3/4$, exceeding the claimed upper bound $2R^{\ast }=1/2$. The corrected theorem uses ties determined independently of the uniforms and labels, as above.

Split the units into visible units $v$, whose patterns occur in the training data, and hidden units $z$. Let $q(v)$ be the empirical probability distribution and maximize the average log-likelihood $\ell ={\sum }_{v}q(v)\log p_{\beta }(v)$, where $p_{\beta }(v)={\sum }_z{p}_{\beta }(v,z)$. The positive phase clamps $v$ to data and averages over $q(v)p_{\beta }(z\mid v)$; the negative phase releases all units and averages over $p_{\beta }(v,z)$. The displayed difference uses these two expectations. When both units are visible, the positive pair product is supplied directly by the data.

For a proof, write $\log p_{\beta }(v)=\log {\sum }_z{e}^{-\beta E(v,z)}-\log Z$. Differentiating each finite sum gives, respectively, $-\beta ⟨\partial E{⟩}_{z\mid v}$ and $-\beta ⟨\partial E{⟩}_{v,z}$. Since $\partial E/\partial w_{ij}=-x_i{x}_j$, subtraction proves the rule. Similarly ${\partial }_{{\theta }_i}\ell =-\beta (⟨x_i{⟩}_{+}-⟨x_i{⟩}_{-})$. A small positive step along these gradients increases log-likelihood. The negative term suppresses associations overproduced by the model; strengthening observed pairs alone misses this correction. The identity requires equilibrium expectations; estimates from insufficiently mixed sampling can give a biased update.

Ridge regression fits a linear regression model while penalizing the size of its slope regression coefficients. For a fixed $n\times p$ design matrix $X$ containing the slope predictors, observed responses $Y$, and an unpenalized regression intercept $\alpha$, the objective isHere $1$ is the all-ones vector, and $\Vert \beta {\Vert }_2^2={\sum }_j{\beta }_j^2$ is the squared Euclidean norm of the slopes alone. The first term is the residual sum of squares; the second implements regularization. No random-error assumption is needed to define this fit. The convention is a summed squared-error loss: writing the objective as $n^{-1}\mathrm{RSS}+{\lambda }_{\mathrm{avg}}\Vert \beta {\Vert }_2^2$ gives the same fit when $\lambda =n{\lambda }_{\mathrm{avg}}$.^[ref]

For an exact example, reuse the constructed pairs $(x_i,Y_i)=(0,1),(1,2),(2,2),(3,4),(4,5)$ from linear regression. Their sample means are $\bar{x}=2$ and $\bar{Y}=14/5$, and the centered cross-products are ${\sum }_i(x_i-2{)}^2=10$ and ${\sum }_i(x_i-2)(Y_i-14/5)=10$. ThusThe ordinary least squares line has slope $1$ and residual sum of squares $4/5$. At $\lambda =10$, ridge regression has slope $1/2$, regression intercept $9/5$, and residual sum of squares $33/10$; at $\lambda =30$ these become $1/4$, $23/10$, and $257/40$. Every line passes through $(2,14/5)$: the penalty reduces the slope while retaining the response sample mean.

The computational benefit is visible after predictor centering. Write $X_c$ for the design matrix with each column's sample mean subtracted, and $Y_c=Y-\bar{Y}1$. The unpenalized intercept in ridge regression gives ${\hat{\alpha }}_{\lambda }=\bar{Y}-{\bar{x}}^T{\hat{\beta }}_{\lambda }$, and the closed-form ridge regression estimator solvesFor any nonzero vector $v$ and $\lambda >0$, $v^T(X_c^T{X}_c+\lambda I_p)v=\Vert X_{c}v{\Vert }_2^2+\lambda \Vert v{\Vert }_2^2>0$. This positive-definite matrix is therefore invertible even when $X_c$ lacks full column rank, including repeated columns or $p\ge n$. The unique penalized slopes resolve the ambiguity of rank-deficient ordinary least squares; they do not supply evidence about slope directions absent from the data. This makes ridge regression useful for multicollinearity.^[ref]

Predictor centering and scaling serve different purposes. Centering handles the free regression intercept. Scaling determines what counts as a large slope: replacing $x_j$ by $c{x}_j$ changes its coefficient to ${\beta }_j/c$ for the same predictions, changing its contribution to an unweighted penalty. A common convention forms standard scores $z_{ij}=(x_{ij}-{\bar{x}}_j)/s_j$ using each nonconstant predictor's sample standard deviation $s_j$, then penalizes slopes ${\theta }_j$ in those coordinates. Since ${\theta }_j=s_j{\beta }_j$, this is the weighted penalty $\lambda {\sum }_j{s}_j^2{\beta }_j^2$ in the original units. Constant columns have $s_j=0$ and must be removed before this scaling. Scaling is a modelling choice; equal penalties in original units and in standardized units define different fits.^[ref]

The singular value decomposition explains the shrinkage. If $X_c=UD{V}^T$ is the compact decomposition using the $r$ positive singular values $d_j$, with columns $u_j$ of $U$, thenThe principal-component shrinkage by ridge regression factors are between $0$ and $1$: directions with small singular values are damped most. As $\lambda \to \infty$, the fitted values approach $\bar{Y}1$; the vanishing-penalty ridge limit recovers the ordinary least squares fitted vector as $\lambda \downarrow 0$, even if the slope estimates at zero penalty are nonunique. The covariance and bias of a ridge regression estimator describes the resulting bias-variance tradeoff. A penalty can reduce variance of an estimator while introducing bias of an estimator; the unbounded directional risk of fixed ridge shrinkage rules out a uniform improvement for a fixed positive penalty over all possible signals.^[ref]

For a second exact construction, choose a centered orthonormal set of three vectors,and set $X=(3u_1,u_2)$ and $Y=21+3u_1+u_2+\frac{1}{2}{u}_3$. The squared singular values are $9$ and $1$. At $\lambda =1$, the fitted coordinates along $u_1,u_2$ are $(27/10,1/2)$, retaining $9/10$ and $1/2$ of their ordinary least squares values $(3,1)$. At $\lambda =9$ they become $(3/2,1/10)$, retaining $1/2$ and $1/10$. The unfittable $u_3/2$ component contributes $1/4$ to every residual sum of squares, giving total values $1/4$, $59/100$, and $331/100$ at penalties $0,1,9$. These are constructed fits, not measured prediction results.

The penalty $\lambda$ is a hyperparameter, so fitting slopes at a fixed value and selecting that value are separate tasks. Use a validation set or K-fold cross-validation to compare candidate penalties by held-out squared-error loss. In each split, compute predictor sample means and scales from that split's training data, apply the same transformation to its validation set, and refit the slopes; using held-out information in preprocessing introduces data leakage. Minimum training residual sum of squares cannot select useful shrinkage: ordinary least squares already attains the smallest possible value.^[ref] After model selection, refit on all training data with the selected penalty. A separate test set or nested cross-validation assesses the entire selection procedure; a low selected validation score alone is no performance guarantee.^[ref]

Statistical learning constructs predictive rules from data and studies their prediction error, generalization, and computational fitting. In supervised learning, training data contain inputs $X$ and a response variable $Y$: regression analysis predicts a numerical response, while classification in statistical learning predicts a label. Unsupervised learning instead seeks structure without supplied responses, as in cluster analysis and principal component analysis. These tasks connect learning to probability and mathematical optimization. ^[ref]

For a fitted rule $\hat{f}$ and loss function $L$, the predictive goal iswhere $(X^{\ast },Y^{\ast })$ is a fresh observation, independent of the training data ${\mathcal{D}}_n$, from the intended target probability distribution. This prediction error fixes the fitted rule and averages over future observations; averaging again over training data gives the learning procedure's average performance. Its practical beauty is that a rule compresses past observations into useful future predictions. Its central difficulty is that excellent training error need not imply good generalization: cross-validation supports model selection, and a reserved test set assesses the resulting choice.

Statistical learning theory studies mathematical guarantees and limitations for learning from data, including how a hypothesis class, loss function, and sample size affect generalization. It distinguishes a particular fitted rule's prediction error from the expected performance of the learning procedure across possible training data.

Supervised learning fits a prediction rule from training data containing both inputs and associated target responses. Regression analysis predicts numerical responses; classification in statistical learning predicts class labels. A loss function specifies which prediction errors matter, and generalization concerns performance beyond the observations used to fit the rule.

A test set is held out from both parameter fitting and model selection and is used to estimate the final selected procedure's prediction error. A validation set influences the choice; a test set assesses that choice after it is fixed. The distinction concerns how observations are used, not their file names.

For a fixed rule and an independent test set of size $m$ from the target probability distribution,This test error is an estimate, with sampling uncertainty; it is not a guarantee for each future prediction. Dataset shift changes what performance the estimate represents.

In the overfitting example, all degree choices are fixed in advance, cross-validation uses only training data, and the selected degree is locked before evaluating 400 separate points. Displaying the other curves afterward explains their behavior; it does not authorize selecting by those scores. Repeatedly inspecting a test set and changing the model makes it part of model selection, introducing data leakage and potentially optimistic assessment. ^[ref]

Training error is the average loss function evaluated on the observations used to fit the prediction rule. It is the empirical risk of the fitted rule on its own training data. With squared-error loss, it is the average squared regression residual, measuring fit to observed responses rather than performance on fresh responses. Flexible fitting usually makes it optimistically biased for prediction error, as quantified by prediction optimism.

For a reproducible polynomial regression example, take $x_i=-1+2i/11$, $i=0,\dots ,11$, and the known noiseless target $f(x)=1+x/2+2x^2$. Construct responses $y_i=f(x_i)+e_i$ with fixed disturbancesOrdinary least squares over polynomials of polynomial degree at most 1, 2, and 7 gives training errors $0.518937$, $0.015784$, and $0.009257$. The monotone improvement follows because these candidate families are nested: every smaller-degree fit remains available in the larger family. This argument concerns minimum training error, not generalization, and assumes the fitting mathematical optimization is solved. Degrees 2 and 7 fitted instead to the exact noiseless responses $f(x_i)$ have mathematically zero training error; the script obtains approximately $5.31\times 10^{-31}$ and $7.82\times 10^{-31}$ from rounding. Extra flexibility alone does not force overfitting.

Take two visible bipolar units, no hidden units, zero thresholds, $\beta =1$, and $E_J(x)=-J{x}_1{x}_2$. As original training data, assign probabilities $3/8$ each to $(+,+)$ and $(-,-)$ and $1/8$ each to $(+,-)$ and $(-,+)$. The data pair product averages to $1/2$. Enumeration of all four states gives two weights $e^J$ and two $e^{-J}$, hence the partition function $Z(J)=4\cosh J$ andStarting at $J=0$, the positive-minus-negative difference is $1/2-0$. A gradient ascent step of size $1/2$ gives $J=1/4$. Each aligned state now has probability $0.311230$ and each opposed state $0.188770$, bringing the model toward the data. The log-likelihood increases from $-1.386294$ to $-1.292224$. Solving ${\ell }^{\prime }(J)=0$ gives $J_{\ast }=(\log 3)/2$, which reproduces the data exactly. Since ${\ell }^{\prime \prime }(J)=-1/{\cosh }^{2}J<0$, this is the unique maximum. These are exact enumerations and their decimal evaluations, with no sampling or laboratory measurements.

Source. Published 2021. Relevant locator: Abstract and Results; PNAS 118, e2014196118: contrastive learning, multi-area prediction and child-view training data.