acceptodds
Under review as a conference paper at ICLR 2027

A Landscape Theory of Association: Paths and Choices in Representation Space

Abstract

Association enables a cue to activate related representations, supporting memory retrieval and flexible reasoning. Understanding this process requires explaining both how an association unfolds and why one destination is selected over alternatives. However, descriptions based solely on representational similarity or energy barriers do not jointly explain these two aspects. We develop a landscape model in which modes of a representation density define associative states and calibrated gradient diffusion connects their basins. Our model separates three quantities: fixed-duration path likelihood, governed by the full Onsager-Machlup action; the weak-noise scale of direct escape, determined by exit barriers; and next-state choice, obtained by normalizing competing exit rates. By doing so, the framework links the representation landscape to associative paths and destination choices, revealing how local curvature shapes probable routes and how competition among accessible destinations governs selection. Across three controlled landscapes, free-time escape-rate slopes recover both barrier predictions within 5.45 percent; in an equal-barrier design, an action-independent moving-tube PDE recovers the complete ordering of six prespecified paths predicted by the full Onsager-Machlup action, whereas the quadratic-only truncation selects a different winner. On a five-well landscape, 160,000 first-exit trajectories recover all four barrier-difference log-odds slopes within 4.5 percent; a fixed-center energy intervention raises the targeted exit probability by 21.8 percentage points on average and reverses the dominant destination. Together, these results quantitatively connect landscape geometry to associative paths, escape scales, and next-state choice.

Primary area:
applications to neuroscience & cognitive science
Keywords:
associative dynamics, cognitive modeling, representation geometry, energy landscapes, stochastic processes

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_a5a9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a landscape theory of association. A density pp over a representation space is mapped to a potential Φ=−D0log⁡p\Phi=-D_0 \log p; non-degenerate modes of pp are taken as associative states, and association is modeled as overdamped gradient diffusion on Φ\Phi with isotropic noise of strength DD.

The central claim is that three quantities derived from the same landscape must be kept distinct: (i) fixed-duration path likelihood, ranked by the full Onsager-Machlup (OM) action including its divergence term; (ii) the weak-noise scale of escape from a basin, set by Freidlin-Wentzell direct-exit barriers Vdir (a,b)=2(Φ(sab)−Φ(a))V_{\text {dir }}(a, b)=2\left(\Phi\left(s_{a b}\right)-\Phi(a)\right); (iii) next-state choice, obtained by normalizing competing Eyring-Kramers exit rates over saddle-adjacent basins, whose log-odds slopes in 1/D1 / D equal barrier differences.

Validation consists of three simulation studies on hand-designed two-dimensional landscapes. On a scaled Müller-Brown potential, fitted Arrhenius slopes match the two barriers within 5.45% and 0.14%. On an equal-barrier two-channel landscape, a moving-tube backward-Kolmogorov PDE that does not use any action functional reproduces the full-OM ordering of six prespecified paths, while the quadratic-only truncation selects a different winner. On a five-well Gaussian mixture, 160,000 first-exit trajectories recover all four barrier-difference log-odds slopes within 4.5%; multiplying one mixture weight by 1.5 (centers and covariances fixed) raises the targeted exit probability by 21.8 points on average and reverses the dominant exit, and a non-adjacent well is never reached directly.

Critical strengths

1.A correct and clearly stated decomposition of three quantities that applied work frequently conflates. The statements are accurate: the free-time infimum of the full OM action diverges to −∞-\infty, quadratic-only or Euler-discretized objectives drop the divergence correction and can select a different path, and sourcewell depth cancels exactly in same-source log-odds. The paper demonstrates concretely, via the equal-barrier construction, that dropping the divergence term can reverse which route is selected. Applied communities that compute "most probable paths" from learned energy landscapes (for example transition-path sampling with score-based models) can genuinely benefit from this clarification, so the paper has value as a careful reference independent of its cognitive framing.

2.Exemplary numerical discipline. The experiments use frozen protocols, prespecified candidate sets, complete event accounting ( 160,000 releases with 17 censored and zero failed), whole-batch bootstrap intervals, time-step, grid, and initialization audits, and an action-independent PDE adjudication in Section 4.2. Deviations from asymptotics are reported honestly. I could rebuild the equal-barrier landscape from Equations (34) and (35) alone; my critical points and Hessians match Table 7, and my minimized OM actions reproduce Table 11 to within about 0.05 nats (for example 144.48 versus the reported 144.43 for c00). This level of reproducibility is rare and deserves recognition.

Critical weaknesses

1.The cognitive claim, which is the paper's reason for existing in this primary area, is never tested, and the controls are weaker than they look. By the authors' own account the theory is classical (Kramers and Arrhenius escape, the OM small-tube theorem, Eyring-Kramers competing rates), so the contribution must rest on the application to association. Yet there is no association data: the USF and SWOW norms are cited but never used, there is no learned density, no embedding space, and no comparison with association or memory-search models. The non-circular test the authors themselves describe in Section 2 (learn the landscape without association responses, then predict held-out behavior) is not run. All three experiments generate data from the same SDE that the theory analyzes, so agreement is expected by construction.

Moreover: (a) "no direct S→RS \rightarrow R exits" is guaranteed by basin topology for continuous paths, so it tests nothing; (b) the ≈0.0875NLL\approx 0.0875 \mathrm{NLL} penalty of the topology control equals −log⁡(1−Q(R))-\log (1-Q(R)) under the frozen distance score and is a data-independent constant, which I verified from Table 13; (c) the center-distance baseline is a strawman, since it ignores mixture weights by construction, and a trivial weight-aware score wbexp⁡(−db2/2σdist 2)w_b \exp \left(-d_b^2 / 2 \sigma_{\text {dist }}^2\right) also flips the dominant exit (E: 0.346 to 0.297, U: 0.327 to 0.422); (d) for well-separated isotropic components the barrier reduces, to leading order, to a quadratic distance term plus a log-weight bias, so the choice law approximates a bias-weighted Gaussian-similarity Luce rule restricted to neighbors (cf. Shepard 1987; Nosofsky 1986). What the landscape adds empirically over this classical family is never shown.

2.The interpretation of the path layer is contradicted by the paper's own benchmark. The OM small-tube ranking describes the mode of the path measure, not where probability mass flows. Reproducing the setup of Section 4.2: the full-OM winner spends about 7.7 of the T=8.57T=8.57 time units within 0.1 of the narrow upper saddle, where the Laplacian of Φ\Phi is about 36, because confinement there maximizes small-tube probability. Direct simulation of the same SDE (Euler-Maruyama, Δt=10−3\Delta t=10^{-3} and 5×10−45 \times 10^{-4}, same DD ) tells the opposite story: the narrow channel carries only about one third of actual transitions, both for free-time crossings (34 to 35% across runs of 4,000 to 8,000 trajectories) and for fixed- TT bridges ending near bb (34 to 37% across radii and time steps). Trajectories that do use the narrow channel linger near the saddle for only about 0.1 time units, roughly 1/70 of what the OM-optimal path does. The Eyring-Kramers asymptotic split is 26%/74%, so the paper's own third layer predicts the opposite channel preference from its first layer, a tension the paper never acknowledges. Meanwhile the tube log-probabilities are around -1600 , so these "most probable routes" are events of no observable relevance. Claims such as "local curvature shapes probable routes" therefore hold only in a technical sense that has no counterpart in association behavior. Route-level questions call for ensemble quantities such as reactive currents in transition path theory, which are not discussed.

Questions for authors

1.Real data. Can you demonstrate the framework on association data? For instance, learn pp or an energy model on word embeddings without association responses, identify modes and saddles, and predict SWOW or USF response distributions for held-out cues, against baselines such as embedding-similarity Luce choice with frequency bias and random walks on semantic networks. This is the result most likely to change my rating.

2.Meaning of the path layer. Which association observable corresponds to a fixed- TT small-tube probability, and how should TT be chosen? In your own benchmark the narrow channel carries about one third of simulated transitions while the OM winner sits at the saddle for about 90%90 \% of TT. Do you agree that the OM ranking does not describe where trajectories go, and would reactive-flux channel probabilities be the appropriate route observable?

3.Reversibility. Gradient diffusion with isotropic noise satisfies detailed balance, so the reduced jump chain obeys Kolmogorov's cycle condition and asymmetry can enter only through a scalar potential. Is this consistent with free-association norms, and how would the framework capture directional effects not explained by state depth, such as forward asymmetry in sequential recall (Kahana, 1996)?

4.Distinctive predictions. Which quantitative predictions distinguish the landscape model from a biasweighted similarity choice rule restricted to neighbors (see Weakness 1 )?

5.Clarification. For c00, both zero-radius extrapolations (140.28 and 142.70) miss the full-OM limit of 144.43 by 1.7 to 4.2 nats, far outside the stated ≈0.06\approx 0.06-nat deterministic bounds. Please report extrapolation uncertainty separately from grid uncertainty.

Suggestions for future work: relate the framework to sampling accounts of memory search (for example Zhu, Sanborn and Chater, NeurIPS 2018); enlarge Figure 2.

6 comments
AuthorsOpenReview ↗

Response to Reviewer a5a9 [Part 1/3: Q1–Q2]

Thank you for the detailed review and for independently reproducing our geometry and action calculations. Your comments prompted us to carry out additional channel measurements and calibrate the similarity controls. Below, we report these results and address your questions.

Response to Q1: Data-derived landscapes and human association

We agree that testing the model against human association data is an important next step. We plan to pursue the experiment you suggest in our next paper. We have already carried out part of the computational work on real image representations and can report those results here.

We used THINGS images and frozen CLIP ViT-B/32 embeddings to construct a spherical energy field over 1,854 concepts, retaining all 512 coordinates. Tangent-Hessian checks certified 1,847 minima, and a bounded search identified 55 concept-pair connections through index-one saddles. We did not use SWOW responses to construct the landscape or find these connections.

Starting at the certified shrimp minimum (D=0.13905D=0.13905), we used 768 new importance-weighted paths to estimate first-grid-exit event masses, averaged over a fixed preparation-age window. Within the finite observation window, the scallop-to-seafood log mass ratio was 0.663 (nominal family interval [0.222, 1.105]). We checked sensitivity to the grid with a separate 256-path refinement panel. We will include this high-dimensional example and its local geometry diagram in the appendix.

These results show that we can construct a landscape from learned representations and measure local transition events on it. They provide a computational basis for the proposed human-association study.

Response to Q2: Reference tubes, channel use, and duration

Thank you for pointing this out. The OM result compares shrinking tubes around specified time-parameterized paths. Channel fractions sum over the trajectories using each channel. An empirical test of the path observable would need repeated, time-resolved latent trajectories with a specified readout and time alignment. Zhu et al. [1] offer a related example, connecting sampling in multimodal representations to memory-search behavior through retrieval distances and temporal correlations. We examined both probability objects on the original equal-barrier two-channel landscape.

For the six predeclared reference paths, the existing action-independent absorbing-tube PDE gives the full-OM ordering under both spatial refinement and tighter solver tolerances. At the smallest tested relative radius, ε/∣b−a∣=0.03\varepsilon/\\|b-a\\|=0.03, log⁡P(c00)−log⁡P(c03)=108.085\log P(c00)-\log P(c03)=108.085 nats. Full OM selects c00c00, whereas omitting the divergence term selects c03c03 and ranks c00c00 last. The PDE calculation independently tests the effect of drift divergence on these reference-tube probabilities.

We also ran SDE simulations at the original noise level and target radius 0.20, with observation horizons TT and H=64TH=64T:

Conditioning eventUpper-channel fractionPointwise 95% interval
First arrival by 64T64T2,737/7,987 = 34.27%[33.23%, 35.32%]
Endpoint in the target disk at TT35/102 = 34.31%[25.19%, 44.37%]

The first-arrival experiment has 205 censored trajectories among 8,192 releases. The endpoint experiment uses 16,384 releases. The paired time steps give identical primary fixed-time classifications. First-arrival upper shares stay between 34.15% and 34.44% across the tested radii, crossing rules, and time steps. Within the same [0,T][0,T] window and radius-0.10 saddle neighborhood, c00c00 spends 7.707 time units, compared with a median of 0.082 among upper-channel paths conditioned on the terminal event.

These results distinguish reference-tube preference from ensemble channel use. Each candidate tube ends inside the same target disk, so conditioning on that terminal event preserves the tube-probability ratios. For equilibrium channel use, committor-based reactive flux is the appropriate observable [2]. We will put these results alongside the tube rankings in the equal-barrier experiment and clarify the abstract and Figure 2 caption. We chose T=3τ≈8.57T=3\tau\approx8.57, where τ=1/λmin⁡(∇2Φ(a))\tau=1/\lambda_{\min}(\nabla^2\Phi(a)) is the source well's slow relaxation time. Relating this duration to human response timing requires independent calibration.

AuthorsOpenReview ↗

Response to Reviewer a5a9 [Part 2/3: Q3–Q4]

Response to Q3: Reversibility and directional effects

The gradient diffusion in our model is reversible. In the metastable approximation, the jump chain Pab=kab/λaP_{ab}=k_{ab}/\lambda_a has reversible weights ρa∝πaλa\rho_a\propto\pi_a\lambda_a, where πa\pi_a is basin occupancy and λa=∑b≠akab\lambda_a=\sum_{b\ne a}k_{ab}. Opposite cue-conditional probabilities can therefore differ, while the forward and reverse probability products around a cycle remain equal.

Your question about directional retrieval points to a useful extension. Persistent net circulation at stationarity, measured on a common set of states, would require a nonreversible model, for example one with a nonconservative drift. The direction and strength of that drift require information beyond a static representation density. We use gradient dynamics here to isolate predictions from the density-derived landscape; adding such a field would be a separate modeling and validation task. Forward effects tied to study order also require assumptions about encoding or retrieval context.

Response to Q4: Weighted similarity controls and topology

We agree that weighting can reverse the winner. We therefore compared full probability predictions using four calibrated similarity-choice controls [3]: shared and inverse-noise scores, each with component-center or actual-minimum distances. The score is sb=αlog⁡wb−βdb2/(2σ02)s_b=\alpha\log w_b-\beta d_b^2/(2\sigma_0^2), optionally multiplied by Dref/DD_{\mathrm{ref}}/D. We fitted two nonnegative shared coefficients in five leave-one-noise-level-out folds, excluding both G0/G1 conditions at the test noise. All methods use the same E/U/L support. FW and EK keep their original frozen geometric predictions.

ModelMean held-condition NLL ↓Mean JS ↓
Original fixed weighted Luce (centers)1.059230.050915
Calibrated: centers, shared0.931020.014025
Calibrated: centers, inverse-noise (primary)0.917490.009960
Calibrated: actual minima, shared0.927320.013041
Calibrated: actual minima, inverse-noise0.913470.008922
FW0.882470.001082
EK0.878160.000053

Calibration closes about 80% of the original fixed-control gap to FW. Against the primary control, FW and EK reduce NLL by 0.03502 [0.03335, 0.03718] and 0.03934 [0.03813, 0.04091], respectively. Against the best listed sensitivity control, the gains are 0.03100 [0.02940, 0.03307] for FW and 0.03532 [0.03418, 0.03681] for EK. For each of the four calibrated controls, FW has lower NLL in 9/10 conditions and EK in 10/10. Both also have positive mean gains in all five held-noise folds. The intervals are nominal 95% matched-batch bootstrap intervals from five batches, with training parameters refitted in every resample. This comparison is posthoc: G1 enters training at the other noise levels.

At the middle noise level in G0, U and L have equal weights and center distances but measured probabilities of 0.308 and 0.086. The primary control predicts 0.180 for each. Using actual-minimum distances gives only a small asymmetry, 0.185 and 0.173, whereas EK predicts 0.304 and 0.091. A center-based score cannot separate two destinations with identical weight and distance inputs, regardless of the shared coefficients. Across five noise levels, the mean absolute error in the intervention-induced increase in P(U)P(U) is 12.37 percentage points for the primary control and 11.95 for the best sensitivity, compared with 1.19 for FW and 1.26 for EK. These comparisons show that the landscape-based models predict the full choice distribution and the magnitude of the intervention effect more accurately, on average, than the calibrated endpoint controls tested here.

Adding R incurs the penalty −log⁡(1−QR)≈0.0875-\log(1-Q_R)\approx0.0875 because it diverts probability mass from E/U/L. This follows from normalization. The predictive comparisons above all use the same E/U/L support.

AuthorsOpenReview ↗

Response to Reviewer a5a9 [Part 3/3: Q5 and References]

Response to Q5: Extrapolation and presentation

For c00c00 versus c04c04, linear and quadratic zero-radius fits give 140.283 and 142.703, compared with the full-OM limit of 144.434. We will report three quantities separately: propagated PDE numerical sensitivities (0.060 and 0.062 nats), extrapolation residuals (−4.152 and −1.731 nats), and sensitivity to the extrapolation specification (2.420 nats). The PDE figures quantify numerical sensitivity, not total extrapolation error. At the smallest tested relative radius (0.03), the independent PDE reproduces the full-OM ordering of all six predeclared candidates under grid refinement and tighter solver tolerances.

References

[1] Zhu, J.-Q., Sanborn, A. N., & Chater, N. (2018). Mental Sampling in Multimodal Representations. Advances in Neural Information Processing Systems, 31.

[2] E, W., & Vanden-Eijnden, E. (2006). Towards a theory of transition paths. Journal of Statistical Physics, 123, 503–523.

[3] Nosofsky, R. M. (1986). Attention, similarity, and the identification–categorization relationship. Journal of Experimental Psychology: General, 115(1), 39–57.

Reviewer_TqQ9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a landscape-based framework for modeling association in representation space. It defines associative states as modes of a representation density and models transitions between them using stochastic gradient dynamics. The framework distinguishes three aspects of association: fixed-time path probabilities characterized by the Onsager-Machlup (OM) action, escape rates governed by Freidlin-Wentzell (FW) barriers, and next-state selection determined by normalized Eyring-Kramers (EK) rates.

The authors validate these theoretical predictions through three controlled experiments involving escape-rate scaling, path selection under equal barriers, and competing transitions in a five-well landscape. The numerical results generally agree with the theoretical predictions.

Critical strengths
  1. A well-organized theoretical framework.

The paper provides a clear distinction between fixed-time path probabilities, weak-noise escape rates, and next-state selection. Integrating these concepts into a unified framework offers a useful perspective on transitions between attractor states.

  1. Careful numerical validation.

The experiments are well controlled and provide quantitative support for the theoretical predictions. In particular, the independent PDE-based verification of path probabilities and the intervention on competing exit probabilities strengthen the numerical evidence.

Critical weaknesses
  1. Evaluation is limited to low-dimensional toy models.

All experiments are conducted on carefully constructed two-dimensional energy landscapes with only a few attractors. While these experiments validate the theoretical predictions, they do not demonstrate whether the framework can characterize attractor transitions in high-dimensional representation spaces, which are more relevant to associative memory. In particular, identifying attractors, saddle points, and transition topology may become substantially more challenging in higher dimensions. Without such validation, the practical relevance of the framework to associative memory remains unclear.

  1. Limited novelty beyond established stochastic dynamics theory.

The main theoretical components, including OM path probabilities, FW escape barriers, and EK transition rates, are well-established results. The conclusions that path probabilities depend on local landscape geometry and that relative barriers govern competing transitions largely follow from these classical theories. Although their integration into a unified framework is useful, the paper does not clearly demonstrate fundamentally new theoretical insights or capabilities beyond applying existing results to controlled landscapes.

Questions for authors
  1. Can the authors provide evidence that the framework can be applied to higher-dimensional landscapes, particularly learned neural representations? It would be valuable to demonstrate whether attractors, saddle points, and transition relationships can be reliably identified beyond the current two-dimensional examples.

  2. What do the authors consider the main theoretical novelty beyond combining existing OM, FW, and EK results? Are there any new predictions or insights specific to associative memory that cannot be directly obtained from classical stochastic dynamics theory?

2 comments

Thank you for your careful assessment of our framework and controlled experiments. We address your two questions below.

Response to Q1: High-dimensional geometry and dynamics

We agree that testing the framework in learned high-dimensional representations is important. This is also part of the human-association study we plan for our next paper. We have already completed part of this computational work on representations of real images and report the results below.

We applied the model to THINGS images encoded by frozen CLIP ViT-B/32. A concept-balanced spherical density over 1,854 prototypes defines the energy on S511S^{511}, retaining all 512 coordinates. We constructed the landscape and candidate connections without using SWOW responses.

On this field, we computed tangent gradients and Hessians, located attractors and index-one saddles, and checked the saddles' downhill connections. From 3,708 image-derived starts, we identified 1,847 locally certified minima and found 55 concept-pair connections in a bounded search. To track basin identities after image replacement, we used transport-aware matching. It matched 1,667 mode identities in at least eight of nine replacement landscapes, covering 90.08% of concepts.

We also measured stochastic events independently. Starting at the shrimp minimum (D=0.13905D=0.13905), we used 768 new importance-weighted paths to estimate first-grid-exit event masses, averaged over a fixed preparation-age window. Within the finite observation window, the scallop-to-seafood log mass ratio was 0.663 [0.222, 1.105] (nominal family interval). We checked grid sensitivity with a separate 256-path refinement panel.

These results show that we can identify attractors and index-one saddles, verify local connections between basins, and measure specified stochastic transition events on the same learned high-dimensional landscape. We will include this example and its local geometry diagram in the appendix.

Response to Q2: Contribution beyond classical theory

Beyond the classical OM, FW, and EK results, we contribute an explicit association model that links representation geometry to distinct, testable predictions for path probabilities, escape rates, and next-state choice, together with controlled numerical tests of these predictions. Specifically:

We formulate association as a gradient diffusion on a calibrated landscape derived from a representation density, with explicitly defined states and consistent noise scaling. This construction connects finite-time reference-tube probabilities, metastable escape, and normalized next-state choice while keeping their distinct limits and normalizations explicit. Two concrete consequences follow: equal barriers need not imply equal finite-time tube probabilities, because landscape curvature along the paths can change their ordering; and leading-order next-state odds are governed by relative direct-exit barriers, with source depth canceling when comparing exits from the same basin and basin adjacency defining the eligible competitors.

The controlled experiments test these predictions separately. The Müller-Brown benchmark measures barrier-controlled escape-rate scaling. The equal-barrier two-channel experiment uses an independent absorbing-tube PDE to verify the full-OM ordering of six predeclared reference tubes. The five-well competing-exit experiment tests barrier-difference log-odds, curvature-prefactor calibration, and a fixed-center intervention. These studies quantify how landscape geometry affects path probabilities, escape rates, and choice distributions beyond endpoint distance alone.

Thank you for the detailed response. The additional high-dimensional CLIP experiments partially address my concerns about applicability, although systematic validation of associative transitions remains limited. My main concern regarding the theoretical novelty beyond established stochastic dynamics results remains, so I will maintain my original assessment.

Reviewer_EtMvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper develops a landscape theory of association, where modes of a representation density define associative states and calibrated gradient diffusion connects the basins od the associative states. Through PDE-based characterizations (including the Onsager-Machlup action from stochastic thermodynamics), the associative paths and destination choices are linked to the representation density. This enables a way to understand which things end up getting associated, and provides a key quantitative description of how to understand it. There is a strong mathematical development, based on previous work in the literature, as well as experimental results that seem to connect well to the mathematical predictions.

Critical strengths

The paper looks to be correct, in terms of the formulation and mathematical development essentially from thermodynamics, and the experiments also look to be comprehensive and correct (as far as I can tell from reading the manuscript).

Association is a key capability in biological/human systems, as well as emerging AI systems, including associative memories that are interesting in their own right but also as simplified versions of Transformer architectures, etc. (Hopfield is all you need, etc.). As such, the topic is of strong interest. The insights are understandable and actionable.

The physics-based methods capture some key insights that go beyond energy-based and score-based methodologies. Through a unified theory, one gets fixed-time path probabilities, free-time escape rates, and normalized next-state choice, all together and through the same common parameterization the experiments are able to reproduce some known empirical phenomena. The insight that topology and barrier differences govern next-state choice is quite insightful.

Critical weaknesses

This not so much a weakness of the paper as of the reviewer, but some of the PDE and thermodynamics arguments were dense for me, and required looking at various literature to understand (it wasn't fully self-contained), but presumably a more knowledgeable reader would not have this issue.

The connection to external empirical phenomena that have been observed by other workers in psychology, AI, etc., could be strengthened, since the goal of theory is to explain such things and to provide design principles for future systems.

Questions for authors

N/A

2 comments

Thank you for your positive assessment of the theory and experiments. We appreciate your comments on the PDE exposition as well. Below, we focus on your suggestion to strengthen the connection to empirical findings in association and memory search.

Empirical connections

We agree that connecting the framework to observed association and memory search is important. We plan to study human association responses in our next paper and have already completed part of the computational work on representations of real images.

Zhu et al. [1] link sampling from multimodal representations to retrieval distances and temporal correlations in memory search. Our framework distinguishes destination choice, basin residence, and time-parameterized paths. These provide separate targets for comparison with response frequencies, waiting times, and ensembles of time-aligned latent trajectories, once a response readout and an independently calibrated timescale are specified.

Using THINGS images and frozen CLIP embeddings, we constructed an energy landscape over 1,854 concepts while retaining all 512 coordinates. We identified 1,847 locally certified minima and 55 concept-pair saddle connections in a bounded search, and independently measured finite-window stochastic events on the landscape. The construction used no SWOW responses.

We can therefore construct a landscape from learned representations and quantify its local dynamics. This provides a computational basis for the proposed human-association study.

References

[1] Zhu, J.-Q., Sanborn, A. N., & Chater, N. (2018). Mental Sampling in Multimodal Representations. Advances in Neural Information Processing Systems, 31.

Reviewer_EtMvOpenReview ↗

human phenomena

Thank you for providing a bit more insight into the human memory/association phenomena. Looking forward to this next paper. Hopefully this next paper can also provide some more intuition behind the mathematical development, for a less-sophisticated reader.

Comments on the paper

1 comment

Dear reviewers,

We are now about halfway through the discussion period. Please read through the authors' rebuttal and respond if there are points that remain unclear (or to let them know they have addressed your concerns).

Sincerely,

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.