- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies whether distractor connections slow neural-network learning. Motivated by classical linear LMS analysis, the authors construct multi-MNIST, where independent classification problems are combined and useful connectivity is known in advance. Experiments compare dense and block-sparse networks under stationary and changing label mappings. The central empirical finding is that dense connectivity impairs adaptation in the non-stationary setting, and increasing network width does not eliminate this gap within the tested range.
Critical strengths
- Studying connectivity from the perspective of learning speed, beyond parameter efficiency, is worthwhile
- The controlled construction makes relevant and irrelevant connections identifiable, enabling informative comparisons.
Critical weaknesses
- The practical relevance of the setting remains limited. All experiments use constructed independent subproblems with known ideal connectivity and, in the non-stationary setting, abrupt label permutations. Real tasks often involve correlated inputs and shared representations, where useful and distractor connections are less clearly separated.
- The theoretical justification is limited to the linear setting. Section 2 motivates the argument through classical LMS analysis, but its extension to nonlinear networks with evolving representations is primarily heuristic.
- The experiments do not fully isolate the proposed mechanism. In Section 4.2, enabling additional connections changes optimization geometry and potentially stability, even though initial predictions are preserved. The transition experiment does not rule out those optimization effects.
- Empirical results are conducted on the MLP architecture and MNIST dataset only, which limits the generality of conclusions.
Questions for authors
- Does the distractor effect persist when subproblems share useful features?
- Does the principal connectivity gap persist with tuned alternative optimizers?
- Why are the step sizes chosen on a very small scale? Does increasing those step sizes validate the argument of distractors slow down learning?
4 comments
Update on Experiments
We have completed the optimizer experiments promised in response to Q2 and added them to Appendix D. We tune the step-size separately for each network and optimizer. Adam and RMSProp perform worse than SGD for both dense and block-sparse networks.
We also test batch sizes from 1 to 32 with a fixed sample budget. The connectivity gap persists across these batch sizes.
Appendix D contains further discussion on why neither of these modifications help.
Thank you for the response. It partially addressed my concerns, in particular, it clarified the motivation for the controlled benchmark and the choice of step-sizes. Several concerns remain:
- I understand the risk of falsely concluding that distractor connections are unimportant in real problems. However, even in the controlled setting with multi-MNIST, within n samples, there might be some samples that share useful features (or even they have the same label). So, in that case, do distractors remain harmful?
- Regarding the transition experiment, preserving predictions at the transition does not isolate the proposed noise mechanism. In particular, adding trainable connections also changes the optimization dynamics and available gradient directions. Tuning step-sizes does not rule out those effects.
- I appreciate the authors for running additional experiments with other optimizers. In Figure 6, training a dense network with RMSProp shows better performance than block-sparse trained with RMSProp even though the performance is bad in both cases (around 50%). This somewhat contradicts the claim in this experiment.
Please clarify if I misunderstood something.
1. We can see two possible meanings to your question, and I’ll try to answer both. Our first interpretation of your question was that different images within a sample could benefit from the same feature detector (i.e. a detector for the number 8). CNN-style weight sharing allows the same detector to be applied separately to different images because they generate multiple activations with a single set of weights, but our networks do not use weight sharing. In our case, even if there is neuron that detects when an image is an 8, it would have to send that signal to all of its outgoing connections, which would be problematic given that images for each subproblem are sampled independently. We therefore expect the benefit of this kind of sharing to be limited in our setting. More likely, we think your question is referring to the case where multiple subproblems genuinely benefit from the same feature activations. For example, suppose two of three subproblems always use identical images. Those two could share useful features, but connections between them and the independent third subproblem would still be distractors. There would be fewer distractors than if all three subproblems were independent, so our mechanism predicts a lower overall cost from distractors. You could imagine a spectrum where with all problems independent there would be the highest number of distractors, which would cause the most issues. And if all of the subproblems were identical there would be almost no distractor connections, so they would cause no issues. The main point is that useful sharing and harmful distractors can coexist. Sharing can reduce the number of distractors, but connections carrying irrelevant information can still cause issues. We expect the common case to fall somewhere between complete sharing and complete independence, with the cost depending on how many distractor connections remain.
2. We agree that our previous claims were too strong and not specific enough, and we have weakened this a little in the newest revision (line 269). The transition experiment shows that adding distractor connections prevents the network from maintaining its previous performance, even when it starts from the block-sparse solution. Thus, the gap cannot be explained solely by a failure to find that solution. However, we agree that that does not generally tell us everything about how the relevant optimization dynamics that affect learning.
3. We expect distractor connections to cause problems with other SGD-based optimizers as well, but that does not mean a network with fewer distractors will always perform better, since other optimizer effects can also influence performance. We do not know why RMSProp performs so poorly in both networks, or why block-sparse is affected more. The decline after an initial peak in block-sparse might suggest loss of plasticity, but at this point that is just a hypothesis. I wish we could offer a more satisfying answer, but understanding the interactions other optimizers add goes beyond the scope of what we explored in our experiments.
Thank you for your review!
W1, W4, Q1: Your main concern seems to be how far our results generalize to other problems (W1), architectures, and datasets (W4). We address these together.
We too want to understand how large a role distractor connections play across problems. But isolating their cost requires either knowing which connections are distractors or having an algorithm that reliably removes them, neither of which is available for general problems. The methods we test, SET and DEEP R, barely find the structure even in multi-MNIST, where the problem is as clean as it gets (Section 4.4). Had we tested real problems with these methods, we might have falsely concluded that distractor connections are unimportant simply because the algorithms failed to remove them. This is why we constructed a problem where good connectivity is known in advance, allowing us to study its effects directly.
However, we can still study which types of problems would be most affected. Our experiments vary properties of the problem, specifically the number of independent predictions (Figure 3) and the frequency of non-stationarity (Figure 5, added in the revision). We find that distractors become increasingly costly when a network must learn more independent things and adapt to more frequent changes. Together with the mechanism in Section 2, these results explain when and why unnecessary connectivity slows learning, beyond this particular problem. The mechanism is also not specific to abrupt changes, as the linear analysis in Section 2 considers gradually drifting targets.
Our broader motivation is to improve algorithms that learn many predictions and/or policies from a single datastream in parallel, as in Sutton et al., 2011, Veeriah et al., 2018, and Modayil & Abbas, 2023. As your review points out, some of the things a network learns may benefit from shared representations, while others require different information. This distinction is present in multi-MNIST (Q1), as the ten outputs within each classification benefit from shared features but not from the features of other subproblems. We deliberately chose classification instead of regression because it naturally has this property.
This motivation also informs how we think about your point on architectures (W4). The mechanism we propose is not specific to any particular architecture, but different architectures may contain different numbers of distractor connections. CNNs, for example, may avoid some unnecessary connections by exploiting known spatial structure. Understanding which existing architectures best avoid distractor connections is an interesting question, but it is not the one motivating our work. Our goal is to understand and ultimately address the problem when useful connectivity cannot be specified in advance. This is particularly relevant to agents that pose their own learning targets (e.g. Veeriah et al., 2019), where the designer cannot anticipate which features each prediction will need. In such settings, we cannot rely on inductive biases to determine architectures; the agent must learn which connections are useful for itself.
We realize this motivation should be clearer in the paper and will revise the introduction accordingly before the discussion deadline.
W2: We agree. Our aim is not to theoretically characterize the complete non-linear setting, nor do we claim to. The linear analysis is there to explain the mechanism and make predictions, and the experiments are there to test them in networks.
W3: The changes in optimization caused by unnecessary connections are precisely what we are investigating, so we are not sure which alternative mechanism you have in mind. We account for differences in step-size sensitivity by tuning the step-size independently for each method, and the performance gap persists without divergence. We would welcome clarification of any specific concerns about optimization geometry.
Q2: We are running Adam and RMSProp on the Figure 1 experiment and will report the results here when they are completed. As we discuss in Section 5, we expect momentum to reduce the gap somewhat because it averages out some of the noise, but not to remove it.
Q3: Step-sizes are independently tuned through sweeps for every method and setting, so increasing the step-sizes for any method in the paper would result in worse performance. The best step-sizes for dense networks are smaller than those for block-sparse networks, which is consistent with our explanation that distractors increase the cost of larger updates, resulting in smaller optimal step-sizes.
We hope these revisions and responses address your concerns and would appreciate it if you would reconsider your score. We are happy to address any remaining questions.