{"id":"5d3ac546-fa45-4be7-8334-21285fd23c5b","arxiv_id":"2505.02248","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-signal (heterosynaptic) synaptic update rule reduces to preconditioned gradient descent at its stationary point, provided consistency scores across neurons share a sign, unifying Hebbian, anti-Hebbian and heterosynaptic plasticity.","lead":"Heterosynaptic plasticity, the modification of a synapse by signals arriving through two different pathways, is shown to be sufficient for gradient-based learning when two local conditions hold. The paper offers a unifying account of brain plasticity phenomena and points to a way of training analog AI hardware without explicitly computing gradients.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HS premise in Theorem 1 is not a generic property of the dynamics: the per-input fixed point ∆Vbar=0 cannot hold for diverse inputs, so Eq. (3) may fail exactly where learning happens.","rationale":"The reader's conditional verdict is correct in spirit, and this pass identifies the precise failure mode of the HS premise rather than a different concern. Theorem 1 is algebraically valid conditional on ∆¯V=0, but that condition is not shown to be reachable or maintainable. For a fixed ¯V and many inputs, p(x) ¯h^T(x) must equal γ ¯V for each x; this forces all p(x) to be collinear and all ¯h(x) to be collinear, which is incompatible with the high-dimensional, input-dependent signals in the paper's own simulations. Relaxing to 'small' ∆¯V does not rescue the argument because the stochastic update noise is not uniformly small per sample, and the theorem's error term is divided by a possibly small scalar ∇^T_ζℓ ζ. The paper explicitly lists HS necessity as open, and its simulations show aggregate stationarity, not per-sample residuals, so the gap is acknowledged but not closed. A direct residual measurement would settle whether Eq. (3) holds; if it does not, the paper's 'gradient computation is easy and common' conclusion is overstated, but the conditional theorem itself stands. Hence the CONDITIONAL verdict is retained.","tokens_in":24602,"tokens_out":9745,"duration_ms":136937,"concrete_test":"On the canonical SAL/random-graph setup of Figures 3 and 9, after training, compute for each held-out input x and each layer i the residual R_i(x)=||p_i(x) ¯h_i^T(x) - γ ¯V_i||_F / ||γ ¯V_i||_F. Also compute the relative error E_i(x)=||p_i(t') - φ_i H_i ∇_{p_i}ℓ|| / ||p_i(t')|| with φ_i from Eq. (9). If the median R_i or E_i over the test set is not small (say ≪0.1) while gradient alignment remains high, the per-sample HS premise fails and the observed learning must be explained by a weaker averaged property, not by Theorem 1. If instead the residuals are small, the concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Theorem 1's use of exact stationarity of the instructive weights. Eq. (12) sets p(x) ¯h^T(x) = γ ¯V for every input x. Since the left side is a rank-one matrix for each x, a single ¯V can satisfy this for many inputs only if all p(x) lie on one ray and all ¯h(x) lie on one ray. In a trained high-dimensional circuit, forward preactivations and error signals vary with x, so the per-sample fixed point is not reached; at best E[∆¯V]=0. The Remark after Theorem 3 tries to relax this to 'small' ∆¯V, but stochastic fluctuations around the expected fixed point are typically of order Var(p¯h^T)^{1/2}/γ, comparable to γ¯V, not uniformly small; and Eq. (28) divides by ∇^T_ζ ℓ ζ, so the O(ε) error is amplified when that scalar is small. Consequently Eq. (3) is not established for the actual update dynamics; the paper's aggregate 'stationarity' plots do not rule out O(1) per-sample violations. Since Eq. (3) is the only bridge from the update rule to ∇_Wℓ via the chain rule, the universality claim rests on an unverified and arguably implausible condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the 'HSDC' design principle: any circuit whose synaptic weights follow a two-signal heterosynaptic update rule of the form (2) implements matrix-learning-rate gradient descent, provided two local conditions hold: Heterosynaptic Stability (HS) and Dynamical Consistency (DC). The formal core is Theorem 1, which shows that at an exact fixed point of the instructive weights, the backward signal satisfies Eq. (7), so the forward update takes the gradient form p(t') = phi Vbar Vbar^T grad_p ell (Eq. (3)); Theorems 3-6 extend the statement to approximate stationarity, nondifferentiable activations, and architectural consistency. Simulations on CIFAR-10 with random connectivity masks, step-ReLU activations, channel rewiring, and an evolutionary growth model are presented as support. The Discussion candidly states that the necessity of HS and the mechanism of DC emergence remain open.","tokens_in":24868,"tokens_out":7508,"duration_ms":99079,"significance":"If the HS condition can be certified for actual learning dynamics, the paper would provide a genuinely unifying statement: feedback alignment, direct feedback alignment, KP, and SAL all become instances of one two-signal motif, and the PSD matrix learning rate connects the result to Adam, RMSProp, and natural gradient. The strengths of the manuscript are its transparent algebraic derivation, its explicit conditional theorem, the breadth of the simulation suite (100 random topologies, step-ReLU, rewiring, evolution), and the falsifiable predictions it generates about metaplasticity, critical periods, and the evolutionary advantage of heterosynaptic plasticity. The central limitation is that the main theorem is conditional on a stationarity assumption that is not proved to hold under the stochastic, nonstationary learning dynamics, so the 'universal' claim in the title is currently stronger than the formal content.","major_comments":[{"comment":"The HS condition is a per-input fixed-point condition. Eq. (12) requires p(x) hbar^T(x) = gamma Vbar for each input x to which the theorem is applied. For a fixed matrix Vbar, the left side is rank one, so a single Vbar cannot satisfy this equation for a diverse set of inputs whose p(x) and hbar(x) span more than one direction; at best E_x[p hbar^T] = gamma Vbar can hold. Theorem 1 as stated therefore applies at isolated inputs, not to the learning dynamics over a data set. The aggregate stationarity reported in Figure 9 is an average statement and does not control per-sample deviations, which can be of order Var(p hbar^T)^{1/2}/gamma and need not be small. Because Eq. (3) is the only bridge from the update rule to grad_W ell, the universality claim requires an additional argument showing either that per-sample stationarity is approximately maintained with high probability or that the accumulated error in the weight update is controlled.","section":"C.1, Eq. (12), Theorem 1"},{"comment":"The O(epsilon) relaxation does not resolve the difficulty. If Delta Vbar_i = O(epsilon), the proof of Eq. (28) divides by the scalar grad_{zeta_i}^T ell(F(x,t)) zeta_i(x,t), which need not be bounded away from zero; consequently the O(epsilon) residual in Eq. (33) is not uniformly small. The Remark after Theorem 3 asserts that contributions from other inputs may cancel as noise, but no averaging theorem or variance bound is supplied. Without such a bound, the claim that the actual update dynamics aligns with the gradient is not established.","section":"C.1, Theorem 3 and its Remark"},{"comment":"Assumption 1 guarantees only the existence of some c0 > 0 such that c0 Z^T grad_{h1} ell = grad_{h2} ell. The proof of Eq. (40) uses this relation as an equality with no constants, so the exact equalities phi_h = phi_p = phi_{tilde h} are not a logical consequence of the stated assumption. At most equality of signs follows. Since the DC condition (4) requires only sign agreement, this is repairable by restating Theorem 4 as a sign-consistency theorem, but as written the exact equality overreaches, and Theorem 5 inherits the same issue when it compares phi_h Vbar Vbar^T grad_p ell and phi_p W W^T grad_p ell.","section":"C.2.1, Assumption 1 and Theorem 4"}],"minor_comments":[{"comment":"The statement that the fixed points of the homosynaptic Hebbian update are 'either zero or infinity' is not correct for a fixed input vector: for Delta W = W h h^T - gamma W, nonzero fixed points of the form W = u h^T exist when gamma = ||h||^2. The qualitative point that one fixed matrix cannot serve diverse inputs remains valid, but the supporting sentence should be corrected.","section":"Section 2.2"},{"comment":"The displayed equations contain long corrupted stretches of repeated bracket symbols that are unreadable; these need to be cleaned before publication.","section":"Equation (5) and the proof of Theorem 6"},{"comment":"The notation 'Z > 0' is undefined; if it means elementwise positivity, the assumption should be restated in terms of the linear map being invertible or having a well-defined sign, since the proof uses the relation as if Z is an arbitrary matrix satisfying h1 = Z h2.","section":"Assumption 1"},{"comment":"The claim that delta_hebb and delta_hetero lower-bound the number of Hebbian and heterosynaptic updates relies on an implied causal relation ('if both i and j fired exactly once, then i caused j to fire') that is not established for the recurrent firing rule; please clarify or weaken the interpretation of these metrics.","section":"Section 2.4 and Appendix D.6"},{"comment":"There are several typographical errors, including 'Purjinke cells' (should be Purkinje cells), 'thaitn' in the Discussion, and 'cersa' in Figure 5; these should be corrected in a final pass.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central idea is appealing and the conditional theorem is sound, but the title and abstract claim universality while the proof is conditional on a per-input stationarity condition that is neither proved nor shown to be approximately maintained under the actual stochastic dynamics. The paper would be substantially strengthened by a rigorous averaged statement or an explicit bound on per-sample stationarity violations. I do not see a citation-pattern or novelty-disclosure concern, but the framing should be aligned with the sufficiency-theorem content."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat should you know: this paper claims that almost any two-signal heterosynaptic plasticity rule satisfying two local conditions implements matrix-learning-rate gradient descent. The core algebra — Theorem 1, where stationarity of the instructive weights forces the update into the form p(t') = phi H grad_p ell with H = Vbar Vbar^T PSD — is clean and correct. That result, together with the observation that deep ReLU networks are automatically dynamically consistent, is genuinely new relative to feedback alignment, DFA, KP, and SAL. The authors also deserve credit for inverting the standard view: HSP is primary, Hebbian plasticity is an emergent byproduct. That is a provocative framing worth taking seriously.\n\nWhat is actually solid: Theorem 1 is a genuine sufficient condition. The proof is short and the logic holds. The HSDC design principle is a useful organizing idea, and the paper is refreshingly explicit about its open points — it does not prove HS is necessary, and it admits the emergence of DC is not understood. The step-ReLU experiment is a nice demonstration that a non-differentiable network can be trained by a rule that behaves like gradient descent on a smoothed version of the model.\n\nWhere the soft spots are: the load-bearing premise, HS, is mostly asserted rather than established. Theorem 1 requires exact stationarity of the instructive weights for each input, and Eq. (12) is a rank-one condition per input. For diverse inputs, a single Vbar cannot exactly satisfy p(x) hbar^T(x) = gamma Vbar for all x. The paper's Remark weakens this to 'small' Delta Vbar, but the error in Eq. (28) is divided by grad^T_zeta ell zeta, which can be small, so the O(epsilon) guarantee is not robust when the gradient signal is weak. The simulations show aggregate alignment, but they do not control per-sample violations of the stationarity condition, so the step from 'conditional theorem' to 'universal gradient machine' overreaches. That said, the paper never hides the conditionality: the theorem is stated precisely. The flaw is in the interpretive leap, not in the algebra. Minor issues: Theorem 4's exact equality is not implied by Assumption 1, which contains an unspecified c0; and several appendix equations are corrupted in the text, making full verification harder.\n\nWho benefits: people working on biologically plausible learning rules and on analog or photonic hardware where explicit gradients are hard to compute. The theory gives a design rule: ensure heterosynaptic stability and sign consistency, and your circuit becomes a gradient machine. Whether this is how the brain learns is a separate question, but the paper deserves serious engagement.\n\nRecommendation: I would send it to peer review and ask for a revision that either proves a relaxed HS condition or characterizes when the error term stays small. The central idea deserves referee time.","headline":"A genuinely conditional theory of heterosynaptic gradient learning: the algebra is clean, the HSDC framing is new, but the load-bearing HS condition is demonstrated only at a level the universality claim needs.","tokens_in":25394,"tokens_out":2290,"would_cite":true,"duration_ms":31338,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92B20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that gradient learning is the generic outcome of almost any two-signal heterosynaptic plasticity rule, provided the circuit satisfies heterosynaptic stability and dynamical consistency.","keywords":["heterosynaptic plasticity","gradient learning","synaptic plasticity","Hebbian plasticity","metaplasticity","matrix learning rate","dynamical consistency"],"falsifier":"Train a heterosynaptic circuit of the form in Eq. (2) with random instructive weights $\\bar V$, and record for each step the norm $\\|\\Delta \\bar V\\|$ together with the cosine alignment between the activation update and the true negative gradient $-\\nabla_p \\ell$; if alignment is consistently high on steps where $\\|\\Delta \\bar V\\|$ is not small, the fixed-point condition is not necessary, while a tight correlation would confirm that heterosynaptic stability is the mechanism.","tokens_in":24344,"feed_emoji":"🧠","tokens_out":13005,"duration_ms":124126,"temperature":0.7,"pith_summary":"The paper tries to establish that gradient descent is not a special algorithm that a brain must implement with carefully engineered wiring; it is the generic outcome of almost any two-signal, heterosynaptic plasticity rule. The argument is summarized by the identity HS + DC = Gradient Learning: heterosynaptic stability at a synapse plus a global sign consistency among neurons makes the circuit descend a loss with a matrix learning rate. If this is right, a wide family of biologically plausible learning rules and standard machine-learning optimizers become special cases of one principle, and heterosynaptic plasticity rather than Hebbian plasticity becomes the primary learning and memory mechanism. The authors support the claim with theorems and with simulations showing that gradient-aligned updates emerge in random circuits, survive rewiring, and arise from simple evolutionary dynamics without ever computing an explicit gradient.","feed_headline":"Two-signal brain circuits are universal gradient machines","feed_subtitle":"Heterosynaptic stability plus sign consistency yields gradient descent without any explicit gradient computation.","key_machinery":"The load-bearing object is the two-signal heterosynaptic update rule, Eq. (2), together with the fixed-point condition on the instructive weight $\\bar V$. At $\\Delta \\bar V = 0$, the stationarity equation $p(t)\\bar h(t')^\\top = \\gamma \\bar V$ can be rearranged into $\\bar h(t') = \\phi \\bar V^\\top \\nabla_p \\ell$ with $\\phi = \\gamma / (\\nabla_p^\\top \\ell\\, p)$, and substituting back produces the activation update $p(t') = \\phi \\bar V \\bar V^\\top \\nabla_p \\ell$; the outer product $\\bar V \\bar V^\\top$ therefore plays the role of a plastic matrix learning rate, and the scalar $\\phi$, called the consistency score, fixes the sign of learning. Dynamical consistency, $\\phi_i \\phi_j \\ge 0$, is what turns local descent at one node into global descent for the whole circuit. The theorems show that for ReLU-type activations (called radially linear units) consistency holds automatically, and for nondifferentiable activations the same identity holds when the loss is evaluated on a pointwise differentiable approximant.","core_discovery":"On the paper's own terms, the central claim is the HSDC formula: HS + DC = Gradient Learning. For a circuit whose synapses obey the two-signal update $\\Delta \\bar V = \\eta p(t)\\bar h(t')^\\top - \\gamma \\bar V$ and $\\Delta W = \\eta p(t')\\bar h(t)^\\top - \\gamma W$, heterosynaptic stability at a node means the instructive weight is at a fixed point, $\\Delta \\bar V = 0$; a short rearrangement then gives $\\bar h(t') = \\phi \\bar V^\\top \\nabla_{p(t)}\\ell$, so the receiving neuron's signal is $p(t') = \\phi \\bar V \\bar V^\\top \\nabla_{p(t)}\\ell(p(t))$, which is gradient descent with a positive-semidefinite matrix learning rate $H = \\bar V \\bar V^\\top$ and a scalar consistency score $\\phi$. If all consistency scores satisfy $\\phi_i \\phi_j \\ge 0$ for every pair of nodes, every part of the network descends in the same direction, and the chain rule turns this activation gradient into a weight update. The paper concludes that gradient computation is a property of two-signal plasticity near heterosynaptic stability, rather than a property of any particular network architecture or learning algorithm.","pith_inferences":["A step the paper does not take is proving that the heterosynaptic fixed point is attracting under the update dynamics; if it is, the universality claim extends from stationary inputs to full training trajectories.","A testable extension for analog hardware is that any physical system with two signal pathways and weight decay that satisfies heterosynaptic stability should implement gradient descent without computing gradients, so one could build a photonic or memristive trainer and measure gradient alignment directly.","The emergence of dynamical consistency in simulations points to a developmental reading: early dense or random growth followed by pruning may be how a circuit searches over topologies for consistency, with evolution selecting heterosynaptic over Hebbian edges."],"forward_implications":["Gradient learning becomes the default of any two-signal circuit that can reach heterosynaptic stability, so biologically plausible learning does not need backpropagation-specific wiring or an explicit error-propagation pathway.","Observed Hebbian and anti-Hebbian plasticity can be reinterpreted as emergent side effects of heterosynaptic descent: after an update the activation change is proportional to the activation itself, making the rule look Hebbian even when the underlying mechanism is not.","Because the matrix learning rate $H$ is plastic, such circuits are automatic meta-learners; prior experience changes the effective learning rate of individual synapses, which is the phenomenology of metaplasticity.","Standard optimizers such as SGD, Adam, RMSProp, and natural gradient are special cases of matrix-learning-rate gradient rules, so the HSDC mechanism places them and biological circuits in one equivalence class.","The circuit can learn through nondifferentiable or low-precision components because only closeness to heterosynaptic stationarity is needed, which allows learning through spiking or quantized activations where plain SGD fails."],"supporting_citations":[{"why":"shows that SGD/backpropagation is a two-pathway circuit, making gradient descent a special case of the heterosynaptic rule.","marker":"[2]"},{"why":"supplies the SAL architecture used as the canonical example and the masking and momentum tricks in the simulations.","marker":"[29]"},{"why":"provides the feedback-alignment dynamics that the theory subsumes as a heterosynaptic special case.","marker":"[30]"},{"why":"provides the KP algorithm, another heterosynaptic special case that the HSDC theory generalizes.","marker":"[28]"},{"why":"represents the prior assumption of unitary instructive synapses that this theory relaxes through a plastic matrix learning rate.","marker":"[35]"},{"why":"documents heterosynaptic plasticity as a biological mechanism, grounding the circuit motifs in experiments.","marker":"[11]"},{"why":"provides the standard normalization-based fix to Hebbian runaway dynamics that the paper contrasts with its stable two-signal rule.","marker":"[36]"},{"why":"documents the homeostatic role of heterosynaptic plasticity and the divergent Hebbian problem that motivates the two-signal form.","marker":"[12]"}],"fun_headline_variants":["Two signals, one rule: brains are universal gradient machines","Heterosynaptic circuits: nature's universal gradient solvers","Brain's learning: same gradient math, no explicit gradient","Heterosynaptic stability turns two signals into gradient descent","Gradient computation is natural: just two signals and stability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that instructive synapses stay near their own stable fixed points while learning happens; if those synapses move substantially during training, the gradient identity in Theorem 1 does not hold.","fun_headline_variants_meta":{"raw":{"variants":["Two signals, one rule: brains are universal gradient machines","Heterosynaptic circuits: nature's universal gradient solvers","Brain's learning: same gradient math, no explicit gradient","Heterosynaptic stability turns two signals into gradient descent","Gradient computation is natural: just two signals and stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000812,"raw_usage":{"total_tokens":3604,"prompt_tokens":1029,"completion_tokens":2575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":2493}},"tokens_in":645,"tokens_out":2575,"duration_ms":18899,"temperature":1.0,"reasoning_tokens":2493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T01:00:01.187930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a heterosynaptic circuit of the form in Eq. (2) with random instructive weights $\\bar V$, and record for each step the norm $\\|\\Delta \\bar V\\|$ together with the cosine alignment between the activation update and the true negative gradient $-\\nabla_p \\ell$; if alignment is consistently high on steps where $\\|\\Delta \\bar V\\|$ is not small, the fixed-point condition is not necessary, while a tight correlation would confirm that heterosynaptic stability is the mechanism.","supporting_citations":[{"cited_title":"Deep learning without weight transport","cited_arxiv_id":null,"evidence_quote":"shows that SGD/backpropagation is a two-pathway circuit, making gradient descent a special case of the heterosynaptic rule."},{"cited_title":"Backpropagation without weight transport","cited_arxiv_id":null,"evidence_quote":"provides the KP algorithm, another heterosynaptic special case that the HSDC theory generalizes."},{"cited_title":"Direct feedback alignment provides learning in deep neural networks","cited_arxiv_id":null,"evidence_quote":"represents the prior assumption of unitary instructive synapses that this theory relaxes through a plastic matrix learning rate."},{"cited_title":"Heterosynaptic plasticity: multiple mechanisms and multiple roles","cited_arxiv_id":null,"evidence_quote":"documents heterosynaptic plasticity as a biological mechanism, grounding the circuit motifs in experiments."},{"cited_title":"Simplified neuron model as a principal component analyzer","cited_arxiv_id":null,"evidence_quote":"provides the standard normalization-based fix to Hebbian runaway dynamics that the paper contrasts with its stable two-signal rule."},{"cited_title":"Homeostatic role of heterosynaptic plasticity: models and experiments","cited_arxiv_id":null,"evidence_quote":"documents the homeostatic role of heterosynaptic plasticity and the divergent Hebbian problem that motivates the two-signal form."}],"review_version":1}