Computing a human-like reaction time metric from stable recurrent vision models

The meteoric rise in the adoption of deep neural networks as computational models of vision has inspired efforts to"align"these models with humans. One dimension of interest for alignment includes behavioral choices, but moving beyond characterizing choice patterns to capturing temporal aspects of visual decision-making has been challenging. Here, we sketch a general-purpose methodology to construct computational accounts of reaction times from a stimulus-computable, task-optimized model. Specifically, we introduce a novel metric leveraging insights from subjective logic theory summarizing evidence accumulation in recurrent vision models. We demonstrate that our metric aligns with patterns of human reaction times for stimulus manipulations across four disparate visual decision-making tasks spanning perceptual grouping, mental simulation, and scene categorization. This work paves the way for exploring the temporal alignment of model and human visual strategies in the context of various other cognitive tasks toward generating testable hypotheses for neuroscience.

Paper

References (68)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 4B3C6/10 · confidence 4/52023-06-15

Summary

This paper introduces a new measure of visual artificial network computation 'time', $\xi_{cRNN}$, as the time-averaged uncertainty of a convolutional RNN trained with an evidential deep learning loss (Sensoy et al.). The authors proceed to analyse the dynamics of this network as it solves a range of classification tasks. Importantly, beyond the network learning to solve the tasks, $\xi_{cRNN}$ is generally well correlated with human reaction times, with which it also shares several qualitative features across tasks.

Strengths

Deep network models of the visual system have often focused purely on classification accuracy and neural representations when comparing to biological data. However, this neglects an important aspect of human behaviour, namely the time it takes to arrive at a decision. The present paper tackles this important question using recent ideas from the machine learning literature and represents an interesting step in the direction of capturing the temporal variability in human behaviour. The authors analyse an impressive breadth of tasks and behavioural data and include many interesting analyses, both qualitative and quantitative.

Weaknesses

The major weakness of the submission is that while there are strong _correlations_ between $\xi_{cRNN}$ and human reaction times, there is less evidence that $\xi_{cRNN}$ is mechanistically similar to human reaction times, since it is fundamentally a measure of uncertainty rather than computation time. In particular, human reaction times generally involve a tradeoff between computation/evidence accumulation and decision making (as in common drift diffusion models). On the contrary, the cRNN has a fixed computational budget and has no need or even capacity for evaluating this tradeoff. This is in contrast to a few previous deep learning models in the literature that are capable of explicitly trading off computation and actions (e.g. Graves et al., Pascanu et al., and refs by the authors in the ML literature, and Jensen et al. in the Neuro/Cogsci literature). These considerations are important because e.g. task difficulty is likely to correlate with both uncertainty and reaction time, and this raises the question of whether a model of one is automatically a model of the other. It might the interesting to compare the current model to alternative models that more explicitly have adaptive/variable computation time, such as a cRNN that computes until $\epsilon$ reaches a certain threshold (akin to classical drift diffusion models). As the authors mention in L326, another potential weakness is that the present approach does not easily generalize beyond classification tasks, which form a small subset of the types of problems humans and animals are faced with during natural behaviour. _References:_ \ Graves et al.: "Adaptive computation time for recurrent neural networks", arXiv 2016.\ Pascanu et al.: "Learning model-based planning from scratch", arXiv 2017.\ Jensen et al.: "A recurrent network model of planning explains hippocampal replay and human behavior", bioRxiv 2023.\ Bansal et al.: "End-to-end Algorithm Synthesis with Recurrent Networks: Extrapolation without Overthinking", NeurIPS 2022.

Questions

L106: what does 'without loss of generality' mean here? L123 "our control models lack inference-time flexibility." What does this mean? L144: It might be worth comparing to or discussing recent work by Bansal et al. (NeurIPS 2022), which develops a way of stably training convolutional RNNs with long computation times that generalize from small to large problem settings. L156 "for a complete mathematical specification of this loss, refer to SI 1". $\epsilon$ does not seem to be defined in either the main text or SI beyond a qualitative description as 'uncertainty'. It would be good to provide a precise mathematical definition somewhere. L204: it would be good to briefly explain how this distance measure by Jeurissen et al. is defined (or at least refer to SI where it can be described in more detail). Figure 3b and 5c: it's not entirely clear to me how these 'activity maps' are computed? Figure 7b: it might be worth talking a bit more about the differences between the human and model data here instead of doing linear regression and calling it a day. The cRNN data is decidedly not linear and exhibits something resembling a plateau followed by a sharp drop at high discriminability. This seems rather different from the monotonic decrease in human RT, and the authors could perhaps speculate on or discuss this difference. Supplementary material: it would be worth having a separate folder with just a few example videos of each condition in addition to the larger set of videos currently provided.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors have adequately discussed the limitations of the work.

Reviewer fhSL8/10 · confidence 4/52023-07-06

Summary

The authors in this work propose a combination of model output uncertainty predictions along with stable recurrent vision architectures in order to derive a proxy for models' reaction time to process input static images. The authors use the previously published horizontal GRU architecture combined with stable RNN training methods (contractive RBP) in their experiments. The proposed work is technically sound with extensive evaluation on 4 diverse datasets (some inspired by prior work in visual psychology) where the computed proxy model reaction time ($\xi$) trends correlate positively with that of humans performing the same task. Overall, this paper contributes a new way to estimate output uncertainty of recurrent convolutional networks and to evaluate similarity between model & human perception by taking into account temporal dynamics of processing static inputs.

Strengths

+ The authors make several key contributions: (1) combining Evidential Deep Learning-based output uncertainty prediction with convolutional RNNs and using the AUC of the output uncertainties over time as a measure of the model's reaction time; (2) training stable recurrent vision models using the above EDL-based readout + objective function in order to obtain models whose temporal dynamics of processing visual inputs matches that of humans performing the same task. + The experiments performed are thorough on all datasets, it is clear that the positive correlation between model and human reaction times is present on the tasks evaluated. + If the authors were to release the code for some of these datasets (especially the incremental grouping task in Fig 2), this would make the contribution even more valuable as it would encourage further exploration of these understudied grouping problems + Clear presentation; The authors have exactly stated their contributions and have written the paper with good clarity and detail. The figures are intuitive and appreciate the video presentations showing model activations and uncertainty through time in the Supplementary.

Weaknesses

- Choice of tasks: While the presented 4 visual cognition tasks are interesting and relevant to cognitive scientists, only the scene recognition task concerns real-world stimuli and natural vision. It would be great if the authors found any interesting patterns of reaction time on natural images or more naturalistic tasks than the ones shown here. - Advantages of similarity in reaction time: It is very interesting that the authors are presenting recurrent networks that are correlated with humans in terms of reaction time (and the benefits as a model of biological visual processing are clear), but does this similarity translate to any significant advantages for machine learning? Are models with better RT correlation with humans also well calibrated, or more robust compared to others? In summary, inter-model comparison based on how well they rank on RT similarity with humans would be an interesting direction that the authors don't comment about here.

Questions

Please refer to my review above for questions.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have adequately addressed limitations of this work.

Reviewer JuhY7/10 · confidence 5/52023-07-10

Summary

The authors of the paper present a novel concept of crafting a metric for analyzing the temporal alignment between Recurrent Neural Networks (RNNs) and human behavior. The innovative approach, rooted in the estimate of task uncertainty via the application of the Derelict distribution, is both interested and pertinent. In an era where understanding the interface between machine learning models, specifically RNNs, and human behavior is of utmost importance, the authors' research is timely and bears the potential for broad impact. A commendable aspect of this study is the diverse range of tasks that the authors have attempted to cover. Their inclusive approach elevates the practical relevance of their research, thus ensuring the applicability of their findings across multiple contexts. However, there are two crucial areas that the authors need to address to improve the overall quality and comprehension of the manuscript. Firstly, in the same/different dot object task, the authors discuss a correlation between the task's difficulty, reaction time, and dot distance. It is important to note that human reaction time also ties strongly with factors such as the object's topology and occlusion. The authors assert that their network and measurements reflect human reaction time concerning these more intricate features. Nevertheless, the demonstration of this claim in the manuscript is not lucid enough. To solidify their claim, it is crucial for the authors to condition the different stimulus conditions on the distance between dots. They should explicitly illustrate how elements like the narrowness of the object outline and occlusion correlate with human reaction time and their RNN metric. Secondly, the specificity of the proposed metric to the particular cortically inspired architecture employed in this research should be clarified. It would be beneficial to understand if this metric can be generalized to other architectures. For instance, would it be feasible to use this metric if one were to implement a convolutional LSTM or another vanilla RNN? The authors' response to this concern could significantly influence the wider application of their proposed metric. In conclusion, while the paper is insightful and tackles a highly relevant topic, it is recommended that the authors address the concerns raised to improve the robustness and clarity of their research. By doing so, the potential impact of the manuscript can be further augmented, providing a more substantial contribution to the field.

Strengths

A relevant and timely paper extending the comparison between human behavior and ANNs.

Weaknesses

Not clear how well they demonstrate that their metric predicts human reaction time between the more straightforward relationship of dots distance, task difficulty and reaction time. In other words do they predict reaction time and object occlusion, and more complex topology of objects (narrowness of segmentation boundaries etc). Does their metric extend to other architectures beyond the specific one they use here.

Questions

Perform analysis to condition on dot distance to demonstrate that other more intricate relationships predict human reaction time.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer yVHQ8/10 · confidence 3/52023-07-14

Summary

This study introduces a metric for evaluating the alignment between model and human behavior wrt task complexity reflected in reaction times. The metric is easy to compute and shows qualitative correspondence with human RTs in different tasks.

Strengths

The paper is well written and easy to understand. The study attempts to go beyond choice responses to establish correspondence between model and behavior. For image-computable models, this is both novel and important. The idea of using model uncertainty as a proxy for RT is intuitive and appealing.

Weaknesses

The models are trained using an algorithm that imposes attractor dynamics which causes instantaneous uncertainty to become zero. This is critical for their proposed metric to remain bounded. But this is a bit odd because subjective uncertainty should be non-zero for difficult tasks even if given a long time. This raises the question of whether the metric is meaningful for other tasks e.g. with ambiguous stimuli. The authors make a convincing case for the utility of this metric for evaluating models. But the metric itself is artificial and might not have a correlate in neural activity. An alternative (and more brain-like) but admittedly more challenging approach would be to learn a policy that chooses left/right/wait at each moment (e.g. via RL) based on the latent representation such that you can directly get a reaction time from the model. This might even work with BPTT.

Questions

Sentence in lines 88-90: I don't follow. Please unpack. Lines 38-39. This assertion is a bit strong. There have been some preliminary attempts to explain human RTs using image-computable models so it would be good to cite them.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

As stated above, would be useful to give some examples of the challenges which may arise when applying this approach to other types of tasks.

Reviewer 4B3C2023-08-10

I appreciate the detailed response of the authors to myself and the other reviewers. ### Fixed computational budget I appreciate that network iterations might change the latent space less at later iterations for the simpler task compared to the more complex task, and that this intuitively seems to suggest that the model should be able to rely on less computation in this case. However, at the end of the day, the two settings use the same number of network iterations and therefore FLOPs (unless I've misunderstood something). It is exactly this intuition that was the motivation for suggesting an 'adaptive computation time' network, and it is interesting that this did not work since it seems like computation could be terminated earlier for simpler tasks, and the results of the authors show that the network has lower uncertainty earlier for simpler tasks. I wonder if there is another criterion that would be more suitable for 'early stopping' of the computation to develop a network that truly has a 'reaction time' rather than only an uncertainty? ### Without loss of generality I would suggest not saying "without loss of generality" and just mentioning that the framework generalizes to other cRNN architectures. ### Figure 7b I appreciate that your framework of course cannot explain all empirical results, and that being able to identify results that are/are not consistent with a given framework is part of its value. However, for this reason, it's also important to be up front about the cases where the data differs from the model.

Authorsrebuttal2023-08-11

A productive discussion on capturing uncertainty vs. time

Thanks for reading our rebuttal and your prompt response. > Computational budget Another insightful comment, and we sympathize with their FLOPs remark. Typically, techniques for early stopping involve at least one (or more) heuristic choices, such as the threshold value in ACT, that need careful finessing. We try to build a formulation that does not enforce such a selection. By training our cRNN with an attractor dynamics constraint, we wanted to achieve stable latent states that would also result in stable readouts (and thus a stable, bounded metric). One observation we like to highlight is that achieving lower uncertainty states does not immediately suggest that these states were reached quickly. In other words, models in our framework can achieve low uncertainty states eventually after a prolonged period of high uncertainty. Naturally, there are cases where the model reaches attractor states with low uncertainty rapidly (for very easy stimuli), or gradually stabilizes in states with high uncertainty. So in essence, our metric also implicitly captures aspects of time since it integrates over instantaneous uncertainty. Empirically, we also find this to be critical. This integrated metric, as opposed to just the uncertainty value at the final state, consistently aligns better with human RTs. Thinking out loud, one way to make these two (no heuristics *and* explicit measure of time) desires meet might be to learn a controller policy (with RL) that can choose "stop" as an action. Though admittedly, this is a more challenging optimization problem. We will plan to add this comment to our extended discussion section. > Without loss of generality Thanks. We will follow your suggestion here. “*The model of interest chosen in this work is the horizontal gated recurrent unit (hGRU; [19]), a convolutional recurrent neural network model that we will canonically refer to as cRNN throughout the rest of the paper. In principle, our framework applies to other cRNN architectures as well.*” > Figure 7b Fair enough! We will include this text in manuscript Section 7.2 (Scene categorization; Results): *While our cRNN models of choice here do not perfectly account for empirical patterns in human reaction times, specifically in their non-monotonic reaction time readouts for the low-discriminability stimuli, they provide a promising start.* And this text in manuscript Section 8 (Discussion): *The value proposition of this framework includes the ability to adjudicate models based on the consistency of their latent dynamics with human reaction time data.*

Reviewer 4B3C2023-08-15

Thanks a lot for the additional response! I agree that any handcrafted 'early stopping' criterion is bound to be heuristic so makes sense re. the attractor constraints. I also understand why the integrated uncertainty is mostly better than the final uncertainty - although if the system does reach an attractor state, the two would be equivalent if you took the total allowed computation time to infinity. It is also in some ways at odds with biological reaction times, where a fast RT can't depend on what the agent's uncertainty would be at some point in the future. Using RL to train agents that can choose when to stop would definitely be an interesting extension, although of course well beyond the scope of the present paper. The Pascanu et al., and Jensen et al. references in my original review do something along these lines for more 'canonical' RL tasks, and it would be interesting to see how this approach would compare to human behaviour for a perceptual decision making task.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC