Pith. sign in

REVIEW 4 major objections 5 minor 34 references

The Alpha-Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The Alpha-Alternator introduces a learned per-time-step gate, driven by the Vendi Score, that switches trust between the current observation and the latent memory, and the paper reports gains over Alternators, Mamba, and other baselines…

desk verdict A simple adaptive gate plus masking helps an Alternator in practice, but the claim that the Vendi Score here measures noise rather than just temporal change does not hold up; the forecasting abstract also overreaches. read the letter →

arxiv 2502.04593 v2 pith:QJ7F2GHJ submitted 2025-02-07 cs.LG cs.AIcs.NEstat.ML

Mohammad Reza Rezaei , Adji Bousso Dieng This is my paper · ORCID
classification cs.LGcs.AIcs.NEstat.ML
keywords dynamicsequencemodelingVendiScoreAlternatorsstate-spacemodelsneuraldecodingtime-seriesforecastingmissingvalueimputationnoiserobustness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most sequence models treat every time step as equally noisy, and this paper argues that fixed assumption hurts on neural and financial data where noise fluctuates. The Alpha-Alternator therefore computes a local diversity score, the Vendi Score, over two shifted windows around each observation and feeds it through a learned sigmoid to set a per-time-step weight that balances the new observation against the latent memory. The sign of the learned linear coefficient tells the model whether high diversity means noisy input, in which case it should trust history, or informative input, in which case it should trust the observation. Training with random masking simulates missing or corrupted entries, and the paper reports that the resulting model beats Alternators, Mamba, VRNN, SRNN, and Neural ODE baselines on three neural decoding datasets and four forecasting benchmarks. The paper's own ablation says both ingredients, adaptive weighting and masking, are needed for the best results.

What carries the argument

The load-bearing object is the adaptive gate $\alpha_t$, a scalar in $[0,1-\sigma_z^2)$ that interpolates the mean of the latent update between $g_\phi(\tilde{x}_t)$ and $z_{t-1}$. The gate is driven by the Vendi Score, the exponential of the Shannon entropy of the normalized eigenvalues of a kernel similarity matrix, computed on the two shifted windows of Eq. (7). One learned pair $(w,b)$ in Eq. (8) converts $VS_t$ into the gate, and Bernoulli masking in Eqs. (5)-(6) creates training-time noise variation. Together these pieces let the model decide at every step whether an observation is signal or noise.

What would settle it

Record a dataset where noise is injected only inside known windows; if the computed Vendi Score does not rise inside those windows, or if a fixed-alpha Alternator trained with the same masking matches the Alpha-Alternator's errors, then the adaptive mechanism is not what causes the reported gains.

Watch

Extended reading notes

Core claim

The paper introduces a generative sequence model whose observation-versus-history gate is a function of local sequence diversity. For each step $t$ it sets $\alpha_t = \sigma(w\, VS_t + b)(1-\sigma_z^2-\varepsilon_0)$, where $VS_t$ is the Vendi Score of the two shifted windows $\tilde{x}_{t-L:t}$ and $\tilde{x}_{t-L+1:t+1}$, and $w$ and $b$ are scalars shared across the dataset. With $w<0$, steps whose windows are diverse are treated as noisy and the latent update leans on $z_{t-1}$; with $w>0$, diverse steps are treated as informative and the update leans on $g_\phi(\tilde{x}_t)$. The same $\alpha_t$ also weights the observation-reconstruction term in the Alternator loss, so the model shifts between short-term reactivity and long-term memory. Random Bernoulli masking during training exposes the model to varying effective noise, and the paper reports that this mechanism yields lower MAE and MSE and higher correlation than fixed-gate Alternators and state-space baselines in neural decoding, with first- or second-place forecasting results on Electricity, Exchange, Solar-Energy, and Weather datasets.

Load-bearing premise

The argument rests on the premises that the Vendi Score of two shifted windows tracks how noisy or informative a time step is, that one learned scalar pair encodes that rule for every sequence in the dataset, and that the two fixed noise variances stayed at their chosen values are adequate.

Editorial extensions

If this is right

  • If the central claim is right, models with this gate should keep lower error in sequence segments where noise spikes, because the gate drops the influence of those observations and relies on latent history.
  • The same masking schedule that simulates noise during training should make the model tolerate missing data at test time, as the imputation experiments with 10% to 95% missing rates are claimed to show.
  • Datasets with low temporal diversity should show smaller gains from the adaptive gate, since the Vendi Score has less dynamic range; the paper's Hippocampus result is offered as exactly that case.
  • The learned sign of $w$ gives a per-dataset interpretation: a negative sign means high diversity is noise for that dataset, while a positive sign means high diversity is informative.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the same Vendi-Score-gated interpolation could be attached to other recurrent or state-space hidden states, turning the mechanism into a general noise-robustness retrofit.
  • Because the shifted windows in Eq. (7) reach one step into the future, a real-time deployment would need to lag the window by one step; the paper does not discuss this latency.
  • Dataset-shared $(w,b)$ means the model learns one noise-interpretation rule per dataset; a per-sequence or per-region pair is a natural test of whether that sharing is the bottleneck on low-diversity data such as the Hippocampus recordings.
  • Using order $q=0.2$ in the Vendi Score emphasizes rare features, so whether the gate's sign or accuracy is sensitive to $q$ is an open empirical question the paper does not address.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes the α-Alternator, an extension of the Alternator sequence model in which the gating weight α_t is made time-dependent: α_t = σ(w·VS_t + b)·(1−σ_z^2−ε0), where VS_t is the Vendi Score of two shifted windows of the (masked) sequence, Eq. (7), and w and b are learned scalars shared across sequences. The model is trained with random observation masking and an Alternator-style loss, Eq. (9). The authors evaluate the method on three neural decoding datasets (motor cortex, somatosensory cortex, hippocampus), on missing-value imputation across 10–95% missingness, and on four time-series forecasting benchmarks, and they include an ablation study of the adaptive gating and masking components. The central claim is that VS_t detects time-varying local noisiness and that the learned sign of w determines whether high VS means noise or informativeness, which in turn yields robustness improvements over fixed-gating Alternators and state-space models.

Significance. The proposed mechanism is potentially useful, and the paper has several strengths: the Alternator background and the loss in Section 3 are clearly presented; the neural decoding experiments span multiple brain regions; the ablation in Table 1 and the imputation study in Figure 4 are informative; and masking-based training is a sensible robustness augmentation. However, the central identification of VS_t with local noisiness is not established and, as defined, Eq. (7) reduces to a temporal-difference magnitude for the RBF kernel. The sampling algorithm is also causally ill-defined. The current empirical evidence supports the weaker claim that an input-dependent gate plus masking improves robustness; it does not support the stronger claim that the Vendi Score specifically detects noise.

major comments (4)
  1. [Section 3, Eq. (7)] The quantity VS_t is not a diversity or noise measure as claimed. For two elements {a,b} with a positive semidefinite kernel normalized so k(a,a)=k(b,b)=1, the Vendi Score depends only on the scalar k(a,b); with the RBF kernel used here it is monotonically increasing in ||a−b||. Hence Eq. (7) is a reparameterized temporal-difference magnitude, not a measure of local noisiness. A large VS_t can equally signal an informative transition (movement onset, weather front) or a noisy observation, and the single shared pair (w,b) can only choose one global direction for this conflation. The paper provides no quantitative link between VS_t and the true noise level: Figure 1 confounds high-frequency signal components with additive noise, Figure 3 reports only dataset-level average VS, and the Table 1 ablation removes the whole adaptive-alpha mechanism rather than isolating the VS-based noise detector. I would like to see, at minimum, a synthetic experiment with known noise variance comparing VS_t with the true noise level and with the local signal derivative, or a redefinition of VS_t over a window with more than two elements together with such a validation.
  2. [Algorithm 2 and Eq. (7)] There is a causal inconsistency in the generative procedure. Eq. (7) defines VS_t using the two windows ending at t and at t+1, so α_t depends on x_{t+1}. In training this is computable from the full sequence, but Algorithm 2 draws x_t and then computes α_t before drawing z_t; at that point x_{t+1} does not exist. As written, the sampling algorithm is not well-defined. The authors should either redefine VS_t using only past and current information (e.g., windows ending at t and t−1) or explain how the future element is obtained during sampling. This also affects the forecasting evaluation if the model was used in an autoregressive mode.
  3. [Abstract and Section 4.2, Table 2] The abstract claims that the α-Alternator outperforms both Alternators and state-of-the-art state-space models across neural decoding and time-series forecasting benchmarks. Table 2 does not support the forecasting part of this claim: on Weather, S-Mamba has better average MSE (0.251 vs 0.254) and MAE (0.276 vs 0.278); on Solar-Energy, iTransformer has better average MSE (0.233 vs 0.234) and MAE (0.262 vs 0.264); on Exchange, DLinear has lower average MSE (0.354 vs 0.358) while the α-Alternator has the best MAE. In neural decoding, Mamba has lower MAE on the Hippocampus dataset (Figure 2). The claim should be qualified to 'best or second-best in most settings', and the cases where the model is not best should be discussed.
  4. [Section 3, Eq. (9)] Because w and b are learned by minimizing the same loss that they gate, the learned sign of w is a fitted interpretation, not an out-of-sample test of whether the Vendi Score detects noise. The external benchmarks and ablations show that a learned, input-dependent gate plus masking can improve performance, but they do not identify the mechanism. Please report the learned values and signs of w and b for each dataset, and add an ablation that replaces VS_t with a simple temporal-difference norm such as ||x_{t-L:t} − x_{t-L+1:t+1}|| to demonstrate that the Vendi Score itself, rather than the distance, is responsible for the gains.
minor comments (5)
  1. [Section 4.2] The forecasting setup omits key hyperparameters (masking rate p_mask, window length L, Vendi Score order q, optimizer, epochs, number of seeds) and reports no confidence intervals or significance tests; please add these for reproducibility.
  2. [Section 5] The claim that Mamba's architecture requires hidden states h_t ∈ R^d with the same dimensionality as the data is inaccurate; Mamba's state dimension is a free hyperparameter that need not equal the input dimension.
  3. [Section 4, first paragraph] The text contains the typo 'α-Altenator'; it should read 'α-Alternator'.
  4. [Figure 2 and Table 2] The color-based best/second-best highlighting is hard to read in grayscale; please use explicit markers or boldface.
  5. [Algorithm 2] Algorithm 2 uses VS_t without defining how it is computed during sampling; if VS_t is meant to be computed from the currently generated sequence, this should be stated explicitly.

Circularity Check

2 steps flagged · score 5.0 of 10

The noise-adaptation claim is partly self-definitional: Eq. (7) defines noisiness as the Vendi Score of two windows (which is just a temporal-difference magnitude), and the learned sign in Eq. (8) is a fitted interpretation; however, the benchmark comparisons are independent, so the paper is not wholly circular.

  1. self definitional [Section 3, Eq. (7); Section 4.1 (RBF kernel setting)]
    "the noisiness of ˜x(i)t, which we denote by VS(i)t, is defined as the VS of two shifted versions of ˜x(i)1:T, VS(i)t = VS({˜x(i)t−L:t, ˜x(i)t−L+1:t+1}; k)"

    Noisiness is stipulated to be the VS of exactly two windows rather than an independently measured noise quantity. For n=2 and any kernel normalized to k(x,x)=1, VS depends only on the scalar k(a,b); with the RBF kernel used in Section 4.1, VS is a monotone increasing transform of ||x_{t-L:t} − x_{t−L+1:t+1}||. So Eq. (7) is a reparameterized temporal-difference magnitude. The statement that a high VS means a noisy element, and that the model therefore adapts to noise, follows from this definition rather than from a quantitative link between VS and true noise levels; Fig. 3 compares only dataset-level average VS, and Table 1 ablates the full αt mechanism, not the VS-as-noise assumption.

  2. fitted input called prediction [Abstract; Section 3, Eq. (8) and Eq. (9); Algorithm 1]
    "This influence is captured by a parameter that is learned and shared across all sequences in a given dataset. The sign of this parameter determines the direction of influence. A negative value indicates a noisy dataset, where a sequence element that increases the VS is considered noisy..."

    The scalars w and b in Eq. (8) are optimized on the Alternator loss in Eq. (9), the same objective in which αt gates both the latent update and the observation reconstruction term. Consequently, the fitted sign of w is a post-hoc reading of parameters chosen to minimize that loss, not an out-of-sample test of whether the Vendi Score detects noise. Because αt multiplies the observation reconstruction error in Eq. (9), the optimizer can lower training loss by reducing αt for high-VS, hard-to-reconstruct inputs; the claim that such behavior constitutes noise adaptation is therefore a restatement of the fitted gate rather than an independently validated mechanism. The external benchmark results still provide independent evidence for predictive accuracy.

full rationale

The α-Alternator's predictive performance is tested against Mamba, Alternator, VRNN, SRNN, NODE, and forecasting baselines on held-out splits, so the main empirical claims do not reduce to the paper's own definitions. The Alternator and Vendi Score citations are prior work with stated mathematical definitions and are not used to forbid alternatives, so there is no load-bearing self-citation chain. The circularity is confined to the mechanism narrative: the paper defines 'noisiness' as the two-window Vendi Score in Eq. (7), which for two points is just a monotone function of the temporal difference between the windows, and then interprets the learned sign of w in Eq. (8) as evidence about dataset noisiness even though w is fit to the same loss it gates. The ablation (Table 1) further shows that adaptive αt alone does not improve MAE on Motor Cortex or Somatosensory compared to the no-adaptation baseline, so the noise-adaptation story is weaker than the benchmark story. The paper's own limitation about fixed σz and σx is a genuine caveat but not a circular step. Overall: partial circularity of the noise-adaptation interpretation, with independent external grounding for the performance claim, giving score 5.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central mechanism depends on learned scalars w and b and on hand-chosen hyperparameters L, q, sigma_z, sigma_x, and p_mask. The most fragile addition is the assumption that a Vendi Score over two shifted windows measures local noise and that one shared (w,b) encodes the direction of that relationship; no independent evidence is provided outside the model's own objective. No new physical or conceptual entity is introduced beyond the adaptive gating function itself.

free parameters (7)
  • w = not reported
    Learned scalar in Eq. (8); its sign is claimed to decide whether high VS is noise or information, and its magnitude sets adaptation strength.
  • b = not reported
    Learned bias in the sigmoid in Eq. (8); it shifts the operating point of the gating.
  • sigma_z = 0.1
    Hand-chosen in the neural decoding setup; fixed latent variance used in alpha_t and in the transition mean.
  • sigma_x = 0.2
    Hand-chosen observation variance used in the observation distribution and in the loss weighting.
  • Vendi Score order q = 0.2
    Hand-chosen in the empirical setup; controls sensitivity to rare versus common features in the diversity score.
  • window length L = 10
    Hand-chosen; defines the two shifted windows in Eq. (7).
  • masking rate p_mask = unspecified
    Introduced in Eq. (5) as a load-bearing training ingredient but never assigned a value in the empirical setup.
assumptions (4)
  • ad hoc to paper The Vendi Score of two shifted windows of a sequence is a valid per-time-step measure of local noisiness or informativeness.
    Introduced by Eq. (7) and used to gate the latent update in Eq. (8); no derivation or independent validation links this diversity score to noise.
  • ad hoc to paper A single scalar pair (w,b) can determine, for all sequences in a dataset, whether a high Vendi Score means noisy or informative.
    Eq. (8) shares w and b across all sequences; the paper provides no evidence that the sign relationship is stable across time or sequences.
  • domain assumption The variances sigma_z^2 and sigma_x^2 are constant across time and sequences.
    Stated in Section 2 and listed as a limitation in the conclusion; the fixed variances enter the adaptive alpha and the transition mean.
  • standard math The Vendi Score's kernel k is positive semidefinite with k(r_i,r_i)=1, so the score is well-defined.
    Assumed from Friedman and Dieng (2023), Section 2.2; the paper uses an RBF kernel in experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Alpha-Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance." pith.science (2026). https://pith.science/paper/QJ7F2GHJ

@misc{pith2026250204593,
  author       = {Pith},
  title        = {Pith review of: The Alpha-Alternator: Dynamic Adaptation To Varying Noise Levels In Sequences Using The Vendi Score For Improved Robustness and Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJ7F2GHJ}},
  note         = {Machine review of arXiv:2502.04593}
}
abstract

Current state-of-the-art dynamical models, such as Mamba, assume the same level of noisiness for all elements of a given sequence, which limits their performance on noisy temporal data. In this paper, we introduce the $\alpha$-Alternator, a novel generative model for time-dependent data that dynamically adapts to the complexity introduced by varying noise levels in sequences. The $\alpha$-Alternator leverages the Vendi Score (VS), a flexible similarity-based diversity metric, to adjust, at each time step $t$, the influence of the sequence element at time $t$ and the latent representation of the dynamics up to that time step on the predicted future dynamics. This influence is captured by a parameter that is learned and shared across all sequences in a given dataset. The sign of this parameter determines the direction of influence. A negative value indicates a noisy dataset, where a sequence element that increases the VS is considered noisy, and the model relies more on the latent history when processing that element. Conversely, when the parameter is positive, a sequence element that increases the VS is considered informative, and the $\alpha$-Alternator relies more on this new input than on the latent history when updating its predicted latent dynamics. The $\alpha$-Alternator is trained using a combination of observation masking and Alternator loss minimization. Masking simulates varying noise levels in sequences, enabling the model to be more robust to these fluctuations and improving its performance in trajectory prediction, imputation, and forecasting. Our experimental results demonstrate that the $\alpha$-Alternator outperforms both Alternators and state-of-the-art state-space models across neural decoding and time-series forecasting benchmarks.

Figures

Figures reproduced from arXiv: 2502.04593 by the authors.

Figure 1
Figure 1. The α-Alternator is robust to varying noise levels compared to a Mamba and an Alternator. The Alternator is more robust to noise than the Mamba. across sequences and within individual sequences. State-space models, such as Mambas, have emerged as a popular framework for sequence modeling (Wang et al., 2025; Gu and Dao, 2023). They have demonstrated strong performance in various applications, including speech recogni… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. VS over time for the Motor Cortex, Hippocampus, and Somatosensory Cortex datasets. Lower VS values in the Hippocampus indicate less diverse observa￾tions across time steps, leading to a diminished effect of the adaptive mechanism in the α-Alternator compared to the Mamba. In contrast, for the Motor Cortex and the Somatosensory datasets, the α-Alternator effectively leverages VS-based adaptation, outperforming the Ma… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of performance on neural imputation across different brain regions. The α-Alternator consistently outperforms the baselines in imputing missing values across Motor Cortex, Somatosensory, and Hippocampus datasets. Results are averaged across missing value rat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    Askari Hemmat, R., Hall, M., Sun, A., Ross, C., Drozdzal, M., and Romero-Soriano, A. (2024). Improving geo-diversity of generated images with contextualized vendi score guidance. In European Conference on Computer Vision , pages 213--229. Springer

    work page 2024

  3. [3]

    A., Leos-Barajas, V., Mills Flemming, J., Nielsen, A., Petris, G., et al

    Auger-M \'e th \'e , M., Newman, K., Cole, D., Empacher, F., Gryba, R., King, A. A., Leos-Barajas, V., Mills Flemming, J., Nielsen, A., Petris, G., et al. (2021). A guide to state--space modeling of ecological time series. Ecological Monographs , 91(4):e01470

    work page 2021

  4. [4]

    Berns, S., Colton, S., and Guckelsberger, C. (2023). Towards mode balancing of generative models via diversity weights. arXiv preprint arXiv:2304.11961

    work page Pith review arXiv 2023

  5. [5]

    T., Rubanova, Y., Bettencourt, J., and Duvenaud, D

    Chen, R. T., Rubanova, Y., Bettencourt, J., and Duvenaud, D. K. (2018). Neural ordinary differential equations. Advances in neural information processing systems , 31

    work page 2018

  6. [6]

    C., and Bengio, Y

    Chung, J., Kastner, K., Dinh, L., Goel, K., Courville, A. C., and Bengio, Y. (2015). A recurrent latent variable model for sequential data. Advances in neural information processing systems , 28

    work page 2015

  7. [7]

    H., Siegel, M., Fries, P., and Engel, A

    Donner, T. H., Siegel, M., Fries, P., and Engel, A. K. (2009). Buildup of choice-predictive activity in human motor cortex during perceptual decision making. Current Biology , 19(18):1581--1585

    work page 2009

  8. [8]

    Dutordoir, V., Saul, A., Ghahramani, Z., and Simpson, F. (2022). Neural diffusion processes. arXiv preprint arXiv:2206.03992

    arXiv 2022

Show all 34 references
  1. [9]

    K., Paquet, U., and Winther, O

    Fraccaro, M., S nderby, S. K., Paquet, U., and Winther, O. (2016). Sequential neural models with stochastic layers. Advances in neural information processing systems , 29

    2016

  2. [10]

    and Dieng, A

    Friedman, D. and Dieng, A. B. (2023). The Vendi Score: A Diversity Evaluation Metric for Machine Learning . Transactions on Machine Learning Research

    2023

  3. [11]

    I., Benjamin, A

    Glaser, J. I., Benjamin, A. S., Chowdhury, R. H., Perich, M. G., Miller, L. E., and Kording, K. P. (2020). Machine learning for neural decoding. Eneuro , 7(4)

    2020

  4. [12]

    I., Perich, M

    Glaser, J. I., Perich, M. G., Ramkumar, P., Miller, L. E., and Kording, K. P. (2018). Population coding of conditional probability distributions in dorsal premotor cortex. Nature communications , 9(1):1788

    2018

  5. [13]

    Gregor, K., Danihelka, I., Mnih, A., Blundell, C., and Wierstra, D. (2014). Deep autoregressive networks. In International Conference on Machine Learning , pages 1242--1250. PMLR

    2014

  6. [14]

    and Dao, T

    Gu, A. and Dao, T. (2023). Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    2023 arXiv

  7. [15]

    B., Bhattacharyya, P., and Dave, S

    Kannen, N., Ahmad, A., Andreetto, M., Prabhakaran, V., Prabhu, U., Dieng, A. B., Bhattacharyya, P., and Dave, S. (2024). Beyond aesthetics: Cultural competence in text-to-image models. arXiv preprint arXiv:2407.06863

    2024 arXiv

  8. [16]

    Lin, B., Bouneffouf, D., and Cecchi, G. (2022). Predicting human decision making in psychological tasks with recurrent neural networks. PloS one , 17(5):e0267907

    2022

  9. [17]

    Lin, L., Li, Z., Li, R., Li, X., and Gao, J. (2023). Diffusion models for time series applications: A survey. arXiv preprint arXiv:2305.00624

    2023 arXiv

  10. [18]

    B., and G \'o mez-Gualdr \'o n, D

    Liu, T.-W., Nguyen, Q., Dieng, A. B., and G \'o mez-Gualdr \'o n, D. A. (2024). Diversity-driven, efficient exploration of a mof design space to optimize mof properties. Chemical Science , 15(45):18903--18919

    2024

  11. [19]

    and Khalili, N

    Mousavi, M. and Khalili, N. Vsi: An interpretable bayesian feature ranking method based on vendi score. Available at SSRN 4924208

  12. [20]

    and Dieng, A

    Nguyen, Q. and Dieng, A. B. (2024). Quality-Weighted Vendi Scores And Their Application To Diverse Experimental Design . In International Conference on Machine Learning

    2024

  13. [21]

    P., Bencomo, G

    Pasarkar, A. P., Bencomo, G. M., Olsson, S., and Dieng, A. B. (2023). Vendi sampling for molecular simulations: Diversity as a force for faster convergence and better exploration. The Journal of chemical physics , 159(14)

    2023

  14. [22]

    Pasarkar, A. P. and Dieng, A. B. (2024). Cousins Of The Vendi Score: A Family Of Similarity-Based Diversity Metrics For Science And Machine Learning . In International Conference on Artificial Intelligence and Statistics , pages 3808--3816. PMLR

    2024

  15. [23]

    S., Seeger, M

    Rangapuram, S. S., Seeger, M. W., Gasthaus, J., Stella, L., Wang, Y., and Januschowski, T. (2018). Deep state space models for time series forecasting. Advances in neural information processing systems , 31

    2018

  16. [24]

    R., Arai, K., Frank, L

    Rezaei, M. R., Arai, K., Frank, L. M., Eden, U. T., and Yousefi, A. (2021). Real-time point process filter for multidimensional decoding problems using mixture models. Journal of neuroscience methods , 348:109006

    2021

  17. [25]

    Rezaei, M. R. and Dieng, A. B. (2024). Alternators for sequence modeling. arXiv preprint arXiv:2405.11848

    2024 arXiv

  18. [26]

    R., Gillespie, A

    Rezaei, M. R., Gillespie, A. K., Guidera, J. A., Nazari, B., Sadri, S., Frank, L. M., Eden, U. T., and Yousefi, A. (2018). A comparison study of point-process filter and deep learning performance in estimating rat position using an ensemble of place cells. In 2018 40th Annual ...

    2018

  19. [27]

    R., Hadjinicolaou, A

    Rezaei, M. R., Hadjinicolaou, A. E., Cash, S. S., Eden, U. T., and Yousefi, A. (2022). Direct discriminative decoder models for analysis of high-dimensional dynamical neural data. Neural Computation , 34(5):1100--1135

    2022

  20. [28]

    R., Jeoung, H., Gharamani, A., Saha, U., Bhat, V., Popovic, M

    Rezaei, M. R., Jeoung, H., Gharamani, A., Saha, U., Bhat, V., Popovic, M. R., Yousefi, A., Chen, R., and Lankarany, M. (2023). Inferring cognitive state underlying conflict choices in verbal stroop task using heterogeneous input discriminative-generative decoder model. Journal...

    2023

  21. [29]

    Wang, Z., Kong, F., Feng, S., Wang, M., Yang, X., Zhao, H., Wang, D., and Zhang, Y. (2025). Is mamba effective for time series forecasting? Neurocomputing , 619:129178

    2025

  22. [30]

    Xu, B., Lu, Y., Inoue, Y., Lee, N., Fu, T., and Chen, J. (2024). Protein-mamba: Biological mamba models for protein function prediction. arXiv preprint arXiv:2409.14617

    2024 arXiv

  23. [31]

    Zhang, X., Zhang, Q., Liu, H., Xiao, T., Qian, X., Ahmed, B., Ambikairajah, E., Li, H., and Epps, J. (2024). Mamba in speech: Towards an alternative to self-attention. arXiv preprint arXiv:2405.12609

    2024 arXiv

  24. [32]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  25. [33]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  26. [34]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.