acceptodds
Under review as a conference paper at ICLR 2027

Backward Truncation Changes the Target Geometry of Local Credit Assignment

Abstract

Adaptive states can keep updating in the forward pass while their derivatives are truncated. We show that the local supervision matched to the full task gradient depends on these retained derivatives: the adjustment can be directional rather than a global loss rescaling. We derive a direction-wise matching rule and test its consequences with controlled interventions. Changing only retained state derivatives reverses the average preference between fixed supervision rules. Directional controls separate the correction from scalar weakening; Quotient reduces mean test error by 9.4% relative to Point. With known dynamics, we also construct supervision before training, obtaining mean gains across two tracker variants and two optimizers. In a coherent optical receiver using recordings from a 1125-km link, gradient-matching preferences vary with horizon and trainable space, while training effects depend on configuration and retained derivatives. With dispersion, nonlinear, and residual compensation trainable under truncation, fixed Complex Quotient supervision improves receiver quality factor (Q-factor) over Point at all 20 tested power–channel workpoints (five powers, four channels), with gains in all 40 paired training-capture comparisons and a mean of 0.045 dB. These findings establish backward differentiation policy as an explicit design variable for local supervision in adaptive systems.

Primary area:
learning on time series and dynamical systems
Keywords:
credit assignment, adaptive systems, truncated differentiation, local supervision, target geometry, gradient matching, differentiable signal processing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_87gHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how backward truncation changes the target geometry of local supervision. The authors derive matching rules showing that retained state derivatives can require directional corrections rather than simple loss rescaling. Controlled interventions, prospective supervision construction, and experiments using recordings from a 1125-km optical link examine the mechanism and its consequences.

Critical strengths
  1. The central insight is meaningful. The theoretical matching rules explain why changing retained derivatives can alter the required supervision direction. The controlled interventions preserve the same forward computation while changing only the retained derivative paths, directly supporting this mechanism.

  2. The controlled interventions are carefully designed. The optical receiver experiments provide additional evidence: Complex Quotient supervision improves Q-factor across all 20 workpoints and 40 paired comparisons when dispersion, nonlinear, and residual compensation are trainable under Detached closure.

Critical weaknesses
  1. Local matching does not consistently predict training outcomes. The authors report finite-step reversals, while receiver gains are small and configuration-dependent. The closure interaction varies across configurations (26 positive, 28 negative), so the effect of retained derivatives on the Complex Quotient advantage is not uniform.

  2. The connection between theory and practical supervision remains incomplete. Exact matching relies on known operators under specific assumptions, while the receiver experiments use fixed structured supervision candidates rather than deriving the matching metric.

Questions for authors
  1. Under what conditions can Complex Quotient supervision approximate the theoretically matched metric? Can the matching analysis predict when its advantage over Point becomes larger or smaller under Detached versus AF-Full?

  2. Can the authors summarize the absolute Q-factor and computational cost of AF-Full with Point, Detached with Point, and Detached with Complex Quotient under matched training conditions?

Reviewer_ze99OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The article is aiming to address the impact of truncation in backward propagation of gradients in time. However, it is not clear from the main article, which context and setup the article is addressing. Based on some treatment in appendix B.6, apparently there is a neural network fΘf_\Theta (noted as GΘG_\Theta in the main text) with parameters Θ\Theta and input xx (noted as yy in the main text), whose output is to track some reference signals ycy_c and yry_r (where yy was used as input in the main text) the bias in error in tracking these signals are to be tracked and eliminated by a simple recurrent system in HH updates. the loss is defined as the weighted l2l_2 norm of the [ecer][e_c e_r] vector, and that is where the propagating gradients backward in time enters into the picture.

Critical strengths

Since the article is not properly written, it is not possible to judge its contributions.

Critical weaknesses

The article's main weakness is its presentation. The inference setup, the corresponding notation (inputs, reference, target, error,residual) and the related problem that the article is focusing on have not been properly defined. I share some of my notes/comments from the main part of the article, where it is simply not possible to follow the problem setup and the analysis approach. The tracker which is the main object of the article, is never porperly defined.I believe the article is not suitable for review in its current form.

Below are some notes from the article:

Abstract It feels like the abstract starts without a proper entrance statement. It does not clearly specify what setup the paper is adressing and what are the contributions of the article in this setup.

Inroduction Similar comment to abstract can be made about the Abstract. It does not clearly define the context of the problem being addressed. It is apparent from the title of the article and some statements in the introduction that the article addresses backward pass truncaction in a backpropagation through time setting. However, the exact system model under consideration was never properly posed.

Section 2.1

Equation (1): y_t is not defined. It seems to be the input of the dynamical iterations, however, it is really suprising that it is not formally defined anywhere in the paper. Last two paragraphs of this section requires, better organization and clarification of the closure concept. I would expect referal to Figure 1.a which is supposed to define the tracking learning model but it is never refered in the article. The caption is also not describing the system under consideration which I assume is represented by this Figure, instead it states a special condition.

Overall, the clear mathematical setup for the problem addressed by the paper is never described here. What are inputs, what are labels, how is error calculated, what are residuals. These should be clearly defined and then the learning challenge regarding this setup should be clearly put forward.

Section 2.2

The section also has many undefined variables and concepts. residual rr is not defined. What ee here is not clear. It is better the general quotient form for W is specified here. what is e in terms of the previously defined variables? which part of Figure 1 are we refering to? The discussion here seems to be addressing weighted l2l_2-norm as loss function but whether rr or whatever ee is the error is not clear.

Section 3 Suddenly paper appears to introduce a toy problem. Θ\Theta was previously introduced as the parameters of the system GΘG_\Theta that uses the input(?) yty_t. Now it looks like error ece_c and ere_r are defined as increments and decrements of this scalar parameter. So there seems to be no input, a constant bias like parameter is under consideration. The meaning of subscripts e_c and e_r are also unclear, and also why is error comparing to 1 and −1-1 is also unclear.

Questions for authors

All of the omited/unclear definitions, notations, descriptions could be asked here as separate questions. However, in my opinion article should be rewritten from the start.

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

This paper studies an upstream predictor followed by an adaptive module, such as a tracker or an adaptive equalizer. It argues that detaching gradients through the adaptive state, while retaining its forward updates, changes the local supervision needed to match the full task gradient. This change depends on direction and cannot generally be captured by a single scalar rescaling. The authors derive matching weights for the scalar and shared-eigenbasis settings and characterize exact positive semidefinite matching through a matrix commutativity condition.

Experiments include two interventions on a semi-synthetic MIT Push task, showing a preference reversal and a 9.4% improvement of Quotient over Point; an a priori construction under known dynamics, yielding improvements of 0.37–0.55%; and real optical receiver data from a 1,125 km link. In the optical experiment, CQ achieves a reported Q-factor gain of 0.045 dB over Point under the D+N+R configuration, covering 20/20 operating points and 40/40 paired comparisons.

Critical strengths

The main observation is clear: keeping the forward computation fixed while changing the backward closure changes the appropriate local supervision. The two-direction example in Fig. 1 makes this easy to see.

The comparisons in P1 and P2 are carefully controlled. P1 changes only the retained derivatives, keeping the task, data order, initialization, forward recursion, and supervision pairs fixed. P2 includes a same-rank rotated subspace and a calibrated Scalar Point baseline to check directional and scalar effects, respectively.

The optical experiments use public recordings from a 1,125 km link, with results covering five power settings and four channels.

I appreciate the detail in the reporting. The authors label the post-primary controls and include negative results, such as the −0.00639 dB final Q-factor interaction for target switching in Appendix D.3. They also list the 300 tasks that could not be run because the raw data were missing.

Critical weaknesses

The theoretical novelty seems limited. Eq. 6 and Prop. 1 largely follow from the chain rule and coefficient matching. The main addition appears to be the setting where forward adaptation remains active; the distinction from synthetic gradients and prior work on modified local losses needs to be clearer.

The evidence for the matching rule itself is limited. P2 uses a fixed mixing weight rather than the derived weights, and Quotient performs best. On the receiver data, CQ's gains are not consistently linked to the backward closure. The small gains in Section 4.3 and the optical experiments, together with the few seeds and correlated measurements, also leave questions about robustness.

A direct comparison with retaining the derivatives is missing. Given the reported 2.4–2.9-fold cost of AF-Full, I would like to see CQ+Detached compared with Point+AF-Full or TBPTT under the same compute budget. This would help establish when changing the supervision is practically worthwhile.

Questions for authors
  1. In P2, Partial uses a fixed mixing weight of 0.5. Have you tried using the per-direction weights from Eq. 8 directly? I’d be interested to see how that compares with Quotient.

  2. How do CQ+Detached, Point+AF-Full, and Point+TBPTT-k compare if we give them the same wall-clock time or FLOPs budget?

  3. For the receiver, I’m not sure how much of CQ’s gain comes from phase/gain invariance and how much comes from truncation. Since I is positive and negative about equally often, how do you interpret this result?

  4. Section 4.3 assumes the operators are known. How would you apply this construction to the receiver, where they are unknown? Is there a practical way to estimate them from the data?

  5. Have you considered testing the same idea in an RNN with truncated BPTT or a differentiable Kalman filter? A more familiar ML example would help show whether the observation extends beyond the two systems studied here.

Comments on the paper

1 comment
Area_Chair_F3BeOpenReview ↗

Reminder: halfway through discussion period

This is a friendly reminder to the reviewers that we are halfway through the discussion period. The authors have posted responses to the reviews, so please acknowledge the rebuttal, and discuss their response and remaining questions.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.