{"id":"3d90ef14-d761-44d1-bb40-a9c0cf0b2a4c","arxiv_id":"2506.21554","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper defines similarity-finding and surprise-response algorithms on Numenta's neocortex model, with a matching non-Bayesian probability update.","lead":"This preprint proposes small modifications to Numenta's thousand-brains neocortex model: algorithms that search for objects similar to a given object, and an algorithm that responds to unexpected sensory input by resetting prior expectations. The algorithms are described in detail and tested on toy grid objects, but no code or brain data is provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The similarity-search claim is conditional on an externally supplied distance function D; the paper gives no account of how the brain acquires D, and the experiments do not validate that active objects are semantically similar.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the similarity-search algorithms presuppose a fixed distance function D and assume the brain knows which features are similar. This is indeed the most important gap because it sits at the input level of the central claim. If D is not learned or calibrated, the algorithms do not add an association capability; they only propagate an externally injected measure. The paper's own experiments in Table 4.1 do not validate the semantic meaning of 'similar': they test whether a predetermined O' remains active, not whether the algorithm finds objects that humans or a ground-truth model would judge similar. This concern is real, but it is not fatal to the paper's proposal: the paper explicitly frames the setting as restricted and states the assumption rather than hiding it. The appropriate verdict remains CONDITIONAL, as the reader concluded, because the central claims are plausible and could be substantially strengthened by specifying or learning D, adding ground-truth-based evaluation, releasing code, and providing convergence guarantees. No change to the reader's verdict is needed.","tokens_in":21528,"tokens_out":9021,"duration_ms":105691,"concrete_test":"Build a controlled object set with known ground-truth similarity (e.g., features drawn from a continuous color space and quantized into SDRs by a fixed code, with human-rated or generative-model object similarity). Run Algorithm 4.2 with three distance functions: (i) the true similarity-derived D, (ii) a random permutation of SDR labels, and (iii) a D learned from co-occurrence statistics of features across objects. Measure precision/recall of the active objects against ground truth. If performance collapses for (ii) and (iii), the algorithm depends on an unmodeled oracle; if (iii) matches (i), the missing learning step can be filled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4, Algorithms 4.1 and 4.2 replace W_in^t with N_d(W_in^t), where N_d is defined via a fixed distance function D on feature SDRs (Eq. 4.1). The paper explicitly assumes 'the brain knows that these are similar,' i.e., that D(W_in(blue), W_in(purple)) is small. This makes the central claim 'find objects similar to O' conditional on D already encoding the true feature-similarity relation. If D is arbitrary or poorly calibrated, the algorithms merely propagate that external measure, and the objects they return are 'similar' only in the circular sense that D declares their features close. The numerical experiments in Table 4.1 do not break this circularity: they only measure whether a pre-chosen object O' remains active under random feature placements, not whether the active objects match a ground-truth similarity judgment. No experiment compares algorithm output against an independent similarity ranking, and no learning rule for D is proposed. Without such an account, the association capability is not implemented in the model; it is assumed at the input level.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two extensions of the Numenta neocortex model. Algorithms 4.1 and 4.2 search a learned object set Ω for objects similar to a given object O by replacing the sensory SDR W_t^in with a neighborhood N_d(W_t^in) defined through an external distance function D; Algorithm 5.1 responds to prediction failures (surprise) by resetting past output activity and setting the movement vector to zero. The paper further claims that Algorithm 5.1 can be interpreted as a Bayesian or non-Bayesian updating process for the latent location L_t(O), and formalizes this as Definition 5.1 together with Assumption 5.1.","tokens_in":21839,"tokens_out":5086,"duration_ms":51488,"significance":"If the claims held, the paper would add two capabilities—similarity-based association and surprise-driven active inference—to a biologically motivated model, with clear pseudo-code that aids reproducibility. The strength of the manuscript is its explicitness: algorithms are stated step-by-step, thresholds are listed, and the toy experiments are honest in reporting low success rates. However, the central claims are only conditional at present: the similarity algorithms rely on an externally supplied distance function whose origin is unexplained, and the non-Bayesian interpretation is constructed to match the algorithm rather than derived from it. These issues directly affect the main assertions of the paper.","major_comments":[{"comment":"The central claim that Algorithms 4.1 and 4.2 'find objects in Ω similar to O' is conditional on an external distance function D, with the assertion that 'the brain knows that these are similar' (text following Eq. (4.1)). The algorithms propagate this externally injected similarity measure; no learning rule for D is proposed, and no experiment compares the output against an independent ground-truth similarity ranking. Table 4.1 only tests whether a randomly chosen O' remains active under random feature placements, not whether the active objects correspond to semantically similar objects. As written, the association capability is assumed at the input level rather than implemented in the model.","section":"§4, Eq. (4.1)"},{"comment":"The non-Bayesian updating process is reverse-engineered to match Algorithm 5.1: γ_t is defined in (5.5) exactly by the algorithm's condition (5.2), and Assumption 5.1 forces the first term in (5.6) to vanish whenever γ_t=1, leaving only the second term. The claimed correspondence between algorithm steps and probabilities (Table 3.1) is therefore asserted by construction rather than derived or independently validated. The paper should either weaken the claim to a definitional correspondence or provide a derivation that starts from the algorithm's actual mechanisms.","section":"§5, Definition 5.1 and Assumption 5.1 (Eqs. (5.5)–(5.6))"},{"comment":"The proposed state-space model is formally incomplete: the sums in (5.6) are over the sets Ω_{t,L}∩Ω_{t-1,+} and Ω_{t,L}∩Ω_{t-1,0}, which depend on the informal threshold 'P>0' introduced before (5.5), and no explicit transition distribution for L_t(O)→L_{t+1}(O) is specified beyond the phrase 'by the movement vector' in (5.7). Without these probabilistic primitives, the claim that Algorithm 5.1 'can be interpreted as Bayesian or non-Bayesian updating' is not precise enough to be checked.","section":"§5, Eq. (5.6)–(5.7)"},{"comment":"The experiments show that with Γ=1, which the paper notes is exactly Algorithm 4.1, the probability that O' is active at the end is only 7.0%, 0.4%, and 0.1% for T=3, 4, 5; for Γ=2 the probabilities also decay quickly (59.6%, 10.2%, 1.6%). These are high failure rates for the central similarity-search claim, yet the paper reports them without discussing why they are acceptable or providing a performance criterion. Moreover, the experimental setup excludes the (NL) case, making the results optimistic. The simulations therefore do not currently support the claim that the algorithms reliably find similar objects.","section":"§4, Table 4.1"}],"minor_comments":[{"comment":"The two thresholds in Definition 3.1 are both rendered as θ_out^o in the text (with and without an overbar); the notation should be distinguished explicitly, e.g., θ^out_o and θ̄^out_o, to avoid confusion.","section":"§3.4, Definition 3.1"},{"comment":"The sentence 'Real systems almost always encounter errors, and in the following, the equations contain few of such errors, unless otherwise noted' is unclear and should be rephrased or removed.","section":"§1"},{"comment":"The assumption that the brain 'knows that these are similar' should be listed as a formal assumption with the other parameters (d, Γ, θ_w, θ'_w, θ''_w, D), rather than appearing only in the prose.","section":"§4, Eq. (4.1)"},{"comment":"The statement that setting δ_i,t=0 'takes no time' is trivially true but does not clarify how the active-inference mapping Φ_move_{i,t-1} → Φ_move_{i,t} is realized in discrete time; this deserves a more precise explanation.","section":"§5, Remark 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a theoretical or computational neuroscience journal, but its central claims are not yet sufficiently supported. The main risks are the external distance function in §4 and the circular construction of the non-Bayesian interpretation in §5. The simulations are honest but not convincing. I recommend major revision rather than rejection, as the algorithmic core may be salvageable with a clearly stated weaker set of claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the Kawakami paper. Short version: it's a clean, honest extension of the Numenta inference algorithm that adds two capabilities—similarity search and surprise-triggered active inference—but the validation is thin and the probabilistic interpretation is reverse-engineered rather than derived. The algorithms themselves are the real content.\n\nWhat's new: replacing the sensory input W_t with a neighborhood N_d(W_t) in the similarity algorithms is a simple but well-defined modification that lets the model keep multiple object hypotheses alive, with a reactivation mechanism (Algorithm 4.2) that handles the two distinct ways an object can drop out. Algorithm 5.1's reset of rho^out to force continued activity after a prediction failure is a neat trick for implementing a minimal form of active inference. The paper is transparent about the restricted setting and about the fact that the non-Bayesian update (Definition 5.1) is built to match the algorithm, not the other way around. That honesty earns credit.\n\nSoft spots, in order. (1) The similarity claim is conditional on an externally supplied distance D on feature SDRs. The paper explicitly assumes the brain already knows blue and purple are similar. That is not an implementation of similarity; it's assuming it. The experiments don't break this circularity: they only test whether a pre-chosen object stays active, not whether the returned objects are similar by any independent criterion. (2) Table 4.1: for Gamma=1, success rates go from 7% to 0.1% as T increases. That is a very weak demonstration of the central capability. (3) The Bayesian/non-Bayesian correspondence is mostly conceptual; no theorems connect Algorithm 5.1 to the update rule beyond the hand-defined gamma_t. (4) No code or data are provided, only a Java program description, and the simulations have no error bars.\n\nWho this is for: people working directly on the Numenta model or on kinematic theories of cortical inference will find the variants useful to think with. It is not a paper that reorganizes the field.\n\nRecommendation: this deserves a serious referee, mainly because the algorithms are precisely specified and the conceptual claims are falsifiable. I would ask for code, confidence intervals, a non-circular similarity evaluation, and either a derivation of the non-Bayesian update or a sharper statement of what it means for the algorithm to 'correspond' to it. As it stands, it's a conditional accept.","headline":"Clean, honest extension of the Numenta inference algorithm with two new capabilities, but the similarity result is conditional on an assumed distance function and the toy validation is weak.","tokens_in":22313,"tokens_out":2023,"would_cite":false,"duration_ms":22064,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Small changes to a neocortex inference algorithm give it similarity search and surprise response, both framed as Bayesian or non-Bayesian updates.","keywords":["thousand-brains model","neocortex","similarity search","sparse distributed representation","active inference","surprise","non-Bayesian updating","association"],"falsifier":"Implement Algorithms 4.1 and 4.2 on a dataset of objects whose pairwise similarity is independently known (for example, human judgments). If, for a fixed $D$, the objects returned are those with $D$-similar features rather than those judged similar by the independent measure in a significant fraction of runs, the claim that the algorithm 'finds similar objects' would be falsified in the sense of similarity as a brain-relevant relation. For Algorithm 5.1, feed a sensory sequence with a known Bayesian posterior: if the algorithm's active-object trajectory diverges from the posterior of the non-Bayesian process (5.6) with the same $\\gamma_t$ sequence, the interpretation as Bayesian/non-Bayesian updating fails.","tokens_in":21317,"feed_emoji":"🧠","tokens_out":8805,"duration_ms":84140,"temperature":0.7,"pith_summary":"The paper sets out to show that the thousand-brains neocortex model, built around prediction, can also handle association: given an object the model has already recognized, it can search the stored objects for ones whose features are similar. The proposed Algorithms 4.1 and 4.2 accomplish this by replacing the sensed feature's sparse distributed representation with a neighborhood of similar representations under a fixed feature distance, so that objects sharing similar features remain active as the model moves along the object. The paper further claims that a third algorithm, Algorithm 5.1, lets the model actively respond to surprise—sensory input most of its mini-columns did not predict—by either rewriting its prior state or zeroing its movement, and that this response can be written as a non-Bayesian updating process. The author is explicit that the setting is restricted and not validated against brain recordings; the contribution is the algorithmic bridge showing these capabilities are compatible with the model's existing machinery. If the claim holds, the model gains two of the brain's most basic cognitive operations without abandoning its core inference routine.","feed_headline":"Neocortex model finds similar objects and reacts to surprise","feed_subtitle":"Slight edits to the standard inference algorithm add association and active inference to the thousand-brains model.","key_machinery":"The load-bearing objects are: (1) the cortical column with its three layers—location (grid-cell modules acting as reference frames), sensory (mini-columns whose sparse activity is the SDR of a feature), and output (whose sparse activity represents objects); (2) the inference Algorithm 3.2, which iterates prediction and observation until one object's output representation stays active; (3) the feature-distance neighborhood $N_d(W)$ on SDRs, which is the sole change that turns inference into similarity search; and (4) the non-Bayesian updating equation (5.6) with switching coefficient $\\gamma_t$, which expresses Algorithm 5.1's response to surprise as a mixture of standard Bayesian updating and a reset of the prior. The argument works by showing that each proposed algorithm is a literal edit of Algorithm 3.2's steps, so convergence and recognition behavior of the original carry over essentially unchanged.","core_discovery":"On the paper's own terms, the central discovery is that the inference algorithm of the thousand-brains model can be reused almost unmodified for two further cognitive operations. For similarity, the paper defines a distance $D$ on the SDRs of sensory features and replaces each sensed feature $W_t^{\\mathrm{in}}$ in the inference cycle with a neighborhood $N_d(W_t^{\\mathrm{in}})=\\{W': D(W,W')\\le d\\}$. Objects whose paths of movement vectors visit locations whose features fall in these neighborhoods remain active, so the model converges not to the observed object but to a similar stored object; Algorithm 4.2 adds a reactivation rule that handles the two ways a candidate can drop out (missing location or missing feature). For surprise, the paper flags an observation as surprising when the fraction of unpredicted mini-columns exceeds $\\theta_w$ (condition (5.1)); Algorithm 5.1 then either sets the previous output activity $\\rho^{\\mathrm{out}}_{k,t-1}=1$ so the surprising input is treated as the new prior (type I), or keeps the movement vector at zero so the next observation re-fixates on the surprising location (type II). The paper formalizes this as a non-Bayesian updating process with a switching coefficient $\\gamma_t$, connecting the model's inference to the free-energy principle's active-inference story.","pith_inferences":["The fixed distance $D$ is doing the real cognitive work; a more complete model would learn $D$ from co-occurrence statistics of features across objects, making the similarity judgment emerge from experience rather than being injected.","The neighborhood trick could be applied in the location layer as well, turning the search into a location-tolerant matching that would allow analogies between objects with different spatial layouts.","Algorithm 5.1's reset can be read as a minimal model of attentional capture: the surprising input overrides the current motor plan, which is a testable behavioral prediction if the algorithm is embedded in a simulated agent.","The distance function's thresholds ($d$, $\\theta_w$, $\\Gamma$) control a precision/recall trade-off; measuring that trade-off on a standard object dataset would give a clean benchmark for how much 'similarity' the model can extract per unit of path length."],"forward_implications":["Similarity search is a drop-in extension: with $d=0$, Algorithm 4.1 reduces to ordinary inference, so the model does not lose its original recognition behavior when the new capability is added.","Association between objects can be recorded in the location layer of a higher cortical column by arranging similar objects close together, giving the model a learned 'similarity' relation rather than just a one-shot query.","Algorithm 4.2's reactivation makes similarity search robust to the two failure modes (missing location, missing feature), so candidates are not permanently lost on the first mismatch.","Algorithm 5.1 gives a concrete implementation of two types of active inference: updating one's model of the world (the coffee-pot example) and orienting toward the surprising stimulus (the eye-movement example), both expressed through the same reset mechanism.","The non-Bayesian formulation provides a probabilistic counterpart to the algorithm's internal states, so the model's surprise response can in principle be compared with Bayesian posterior updates on the same sensory sequence."],"supporting_citations":[{"why":"Supplies the core learning and inference algorithms (Algorithm 3.1 and 3.2 backbone) that Algorithms 4.1, 4.2, and 5.1 modify.","marker":"[9]"},{"why":"Supplies the location-layer grid-cell machinery and the movement-vector formalism that Algorithms 4.1 and 4.2 operate on.","marker":"[14]"},{"why":"Provides the first software instantiation of the thousand-brains model and the model-based/model-free motor policies that Algorithm 5.1's type-II response uses.","marker":"[1]"},{"why":"Provides the non-Bayesian updating rule (equation (5.4)) that Definition 5.1 generalizes to the surprise-switching case.","marker":"[12]"},{"why":"Establishes the free-energy principle and its treatment of surprise, which Algorithm 5.1 is designed to implement.","marker":"[4]"},{"why":"Further develops the free-energy principle as a unified brain theory, framing the active-inference interpretation of Algorithm 5.1.","marker":"[5]"},{"why":"Provides the modern active-inference formulation (free-energy principle in mind, brain, and behavior) that the paper connects to its Algorithm 5.1.","marker":"[18]"}],"fun_headline_variants":["Reusing inference to add similarity and surprise to brain model","Brain model's inference engine handles similarity and surprise","Similarity and surprise via reused neocortex inference","Neocortex model does more with same inference loop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The similarity search presupposes a fixed, pre-specified distance function $D$ on feature SDRs that already agrees with true object similarity; if the brain (or the model) has no way to learn or calibrate that distance, the 'similar objects' it returns are only as valid as that externally injected measure.","fun_headline_variants_meta":{"raw":{"variants":["Reusing inference to add similarity and surprise to brain model","Brain model's inference engine handles similarity and surprise","Similarity and surprise via reused neocortex inference","Neocortex model does more with same inference loop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001226,"raw_usage":{"total_tokens":5094,"prompt_tokens":1056,"completion_tokens":4038,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":3975}},"tokens_in":672,"tokens_out":4038,"duration_ms":26858,"temperature":1.0,"reasoning_tokens":3975,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:48:08.823070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement Algorithms 4.1 and 4.2 on a dataset of objects whose pairwise similarity is independently known (for example, human judgments). If, for a fixed $D$, the objects returned are those with $D$-similar features rather than those judged similar by the independent measure in a significant fraction of runs, the claim that the algorithm 'finds similar objects' would be falsified in the sense of similarity as a brain-relevant relation. For Algorithm 5.1, feed a sensory sequence with a known Bayesian posterior: if the algorithm's active-object trajectory diverges from the posterior of the non-Bayesian process (5.6) with the same $\\gamma_t$ sequence, the interpretation as Bayesian/non-Bayesian updating fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the core learning and inference algorithms (Algorithm 3.1 and 3.2 backbone) that Algorithms 4.1, 4.2, and 5.1 modify."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the location-layer grid-cell machinery and the movement-vector formalism that Algorithms 4.1 and 4.2 operate on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the non-Bayesian updating rule (equation (5.4)) that Definition 5.1 generalizes to the surprise-switching case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the free-energy principle and its treatment of surprise, which Algorithm 5.1 is designed to implement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Further develops the free-energy principle as a unified brain theory, framing the active-inference interpretation of Algorithm 5.1."},{"cited_title":"Friston, Active Inference: The Free Energy Principle in Mind, Brain, and Behavior , MIT Press, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the modern active-inference formulation (free-energy principle in mind, brain, and behavior) that the paper connects to its Algorithm 5.1."}],"review_version":1}