Dear reviewer xkmy,

Thank you for taking the time to review our paper. We're happy you think that SHML is a novel approach with a solid theoretical foundation and that has significant potential in high-stakes applications.

To address your concerns, we've expanded our evaluation with five new datasets, additional SHML effectiveness evaluation, added a study with six new benchmarks, evaluated SHML across 10 different ML models, and performed a qualitative ablation study. We hope this information provides reasons to consider increasing your score. Our responses (A-H) follow:

(A) More datasets

While our paper's main contribution is the SHML theoretical framework, with viability studies as proof-of-concept, we agree more real-world examples improve the paper. We've added five new datasets common in concept drift adaptation, varying corruption levels at test time to evaluate adaptation methods. Please find this in Table 1 in the response pdf.

Takeaway: $\mathcal{H}$-LLM improves upon reason-agnostic baselines in five real-world datasets when the reason for performance drop can be diagnosed, even in test-time scenarios. This highlights the potential of self-healing ML.

We'll include these results with more experimental details in the updated paper.

(B) Greater empirical analysis of self-healing ML

To address your concern about limited empirical analysis, we've conducted an additional viability study across the five datasets. We vary problem difficulty (corruption level) at test time and evaluate self-healing benefits (Fig. 1 in the pdf).

Takeaway: $\mathcal{H}$-LLM improves performance relative to baselines across five real-world datasets, with greater self-healing effects at higher corruption levels.

We'll include these results with more experimental details in the updated paper.

(C) Limited benchmarks

We emphasize that our viability studies aim to show that any reason-agnostic strategy will perform poorly whenever it is important to understand the reason for performance degradation. This is because such strategies do not directly address the root cause.

That said, we've extended our evaluation benchmarks. We've run an additional viability study with five benchmarks and four adaptive algorithms common in concept drift adaptation (Table 2 in the response pdf).

Takeaway: $\mathcal{H}$-LLM consistently outperforms adaptation methods and adaptive algorithms that fail to address test-time corruption, demonstrating self-healing ML's feasibility in high-stakes environments where understanding model degradation reasons is important.

Additionally, we run one more viability study to evaluate whether these gains are consistent across different ML models. We show that $\mathcal{H}$-LLM consistently outperforms other adaptation methods across 10 different ML models (shown in Table 4 in the response pdf), highlighting the model-agnostic nature of the framework.

(D) Implementation of $\mathcal{H}$-LLM

We're glad you're interested in $\mathcal{H}$-LLM's practical implementation. Besides the main text, implementation details are in the appendix: components (Appendix B.1.), prompt templates (Appendix B.2), outputs (Appendix B.3) and viability studies (Appendix C.2.). That said, we acknowledge it might be difficult to find this information.

Actions taken: Move input/output examples and prompt structures from appendix to main text.

(E) Practical challenges of deploying self-healing systems

We're excited that you're interested in the practical deployment of self-healing systems. We answer this exact concern in Section 5.1. titled "unique challenges of building self-healing systems". We agree that the naming of this section could be improved.

Actions taken: We will rename section 5.1. to Practical challenges of deploying self-healing systems to improve clarity. We will also expand the discussion in the appendix.

(F) More details on experimental setup and data used

In the camera-ready version, we'll use the additional page to move some of the experimental details and dataset information described in Appendix C to the main paper.

(G) SHML comparison to other concept drift adaptation methods

Computational overhead. SHML methods have larger overhead than reason-agnostic approaches due to the self-healing system (LLM pipeline) identifying model failure reasons. Practically, it takes 20-40 seconds to implement a full pipeline and correct a model upon drift detection. This overhead is negligible for real-world systems given the benefits. Overhead may vary across systems.

Sample efficiency. No differences exist as failure detection doesn't depend on sample size, but on self-healing pipeline complexity.

Actions taken: Add an appendix discussing computational details for SHML systems.

(H) Contribution of each component

We've performed an additional qualitative ablation study, running a self-healing ML system and iteratively removing one component, inspecting its output (Table 3 in supplementary pdf).

Takeaway: This study illustrates that all components are required for the self-healing ML system to work. Having said this, we see the robustness of self-healing ML as a big research agenda for future work and hope this spurs research in this area.

Thank you

We believe these changes should greatly enhance the paper's contribution and improve its clarity. In addition to the clarifications which will be included in the camera-ready version, we have added five new datasets, additional SHML effectiveness evaluation, evaluation across 10 ML models, a study with six new benchmarks, and a qualitative ablation study.

Thank you for your help.

If we addressed your concerns, we hope you consider significantly revising your assessment of the paper's impact and the corresponding evaluation for NeurIPS. ☺️

Dear reviewer xkmy,

We thank you once again for the effort put into reviewing our paper. As there are only a couple working days left in the discussion period, we would like to ask if our response has satisfied your concerns. If so, we hope you consider revising your assessment of the paper's impact and the corresponding evaluation for NeurIPS. If any concerns remain, we are happy to discuss them further here.

Dear R-vweT,

we're glad you think our paper is important and easy to read.

(A) Classification of reasons for performance drops

You're right that we don't provide an exhaustive classification of reasons for performance drops. We aim SHML to be a meta-level approach which flexibly adapts to various scenarios.

A meta-level approach offers two main advantages: 1) It allows the system to adapt to unforeseen scenarios and new types of model degradation not captured by pre-defined lists. 2) It flexibly generates multiple hypotheses to explain the same model degradation, as different explanations may account for the same issue (illustrated in Table 2 of the main paper). Also note that SHML could incorporate pre-defined causes into the framework, as these represent deterministic if-else diagnosis functions and would result in a less complex policy function $\pi$ than explored in the paper.

Actions taken: Include the classification of performance drops in the discussion.

(B) How well the system works on large set of benchmarks

We introduce self-healing ML as a framework for understanding model degradation. Our goal is to demonstrate its viability rather than provide extensive benchmark comparisons.

That said, to address your concern directly, we've conducted additional viability studies:

• We compare $\mathcal{H}$-LLM, with baseline approaches on five real-world datasets by introducing corruption at test time (Table 1 in supplementary pdf). Takeaway: $\mathcal{H}$-LLM improves upon reason-agnostic baselines when the reason for performance drop can be diagnosed.
• We show how the importance of self-healing ML varies with the problem difficulty (Fig. 1 in the response pdf). Takeaway: The effects of self-healing are greater when the data corruption levels are higher.

Actions taken: We will include a summary of this discussion in the main paper.

(C) The effectiveness of each component

We emphasize that we see robust studies of each component as non-trivial research directions, as our paper primarily aims to introduce SHML. That said, we address your concern by performing a qualitative ablation study. Setup: We run $\mathcal{H}$-LLM and iteratively remove one component from the system, inspecting its output. This is presented in Table 3 in the response pdf.

Takeaway: We illustrate that all components are required for the SHML system to work.

(D) Moving items from the appendix to the main text

In the camera-ready version, we'll use the additional page to expand the viability section in the main text. We'll include more details on the prompts and input/output examples of $\mathcal{H}$-LLM.

(E) Is self-healing the first reason-agnostic method?

You're correct that there have been previous papers attempting to identify reasons for model degradation, discussed in Section 2 (L83 - 90 and L105-120). To clarify our contribution, we've created a comparison table highlighting the key differences between SHML and existing approaches:

[1] Gama et al. (2004), [2] Lu et al. (2018), [3] Goncalves et al. (2013), [4] Alippi et al. (2013), [5] Budhathoki et al. (2021), [6] Zhang et al. (2022), [7] Cruz et al. (2018)

Actions taken: We will highlight that SHML is the first framework where the diagnosis and adaptation are not fixed in advance and where the diagnosis informs the adaptation.

(F) Defining conditions in known benchmarks that could be used to extensively test SHML

We agree that having an extensive set of known conditions could be extremely useful for evaluating SHML systems that would likely gain a lot of traction in the ML community. We think there are multiple conditions that could be used to simulate real-world conditions in existing benchmarks (such as data corruption, systematically changing the DGP, introducing external shocks such as covid-19) which would require significant testing and validation. We see this as promising future work.

(G) Failures in sequential systems and how that affects self-healing ML

You're right that failure in one component can affect the entire system's performance.

a) Component sensitivity analysis. To provide initial insights, we refer to the qualitative ablation study presented earlier in our rebuttal. The study showcased that each component working is required for the system's overall performance.

b) Built-in safeguards. SHML includes mechanisms to mitigate cascading failures with the testing component (step 4). Actions that do not improve performance over baseline are discarded.

We see designing fail-safe self-healing systems as a promising research direction.

(H) Expanding limitations

In the camera-ready version, we'll expand the discussion of the limitation sections.

1. We'll add a paragraph discussing specific challenges with root cause identification (expanding on and moving items from section 5.1 lines 208 - 2012)

2. We'll add a paragraph discussing the challenges with choosing an appropriate action (expanding on and moving items from section 5.1 lines 213-216)

3. We'll add a challenge explaining when understanding the root cause is easy/difficult based on our experience.

Thank you

You have helped us improve our paper. Given these changes, we hope you consider revising your assessment of the paper's impact and the corresponding evaluation for NeurIPS. ☺️

Dear Reviewer LnCp,

Thank you for your thoughtful feedback on our work on self-healing machine learning. We appreciate your recognition of the importance and novelty of our approach. We'll address your concerns in two parts, corresponding to both weaknesses.

(A) Clarity of Section 3

The goal of Section 3 is to describe how a deployed ML model ($f$), such as a logistic regression classifier, can be "healed" with a self-healing system ($\mathcal{H}$). In this case, healing refers to restoring the performance of $f$ after it drops.

A natural question is: how do we know what actions we should take to heal $f$ once it drops in performance? We say that $f$ should be healed by the healing-system $\mathcal{H}$ which follows a policy $\pi$. The policy $\pi$ outputs actions $a$ (such as $a_1$: retrain a model or $a_2$: remove corrupted features) which are then implemented onto $f$. Therefore, $\mathcal{H}$ follows a policy $\pi$ which helps to determine optimal actions $a$ that change/modulate the deployed ML model $f$.

We will improve the paper's writing in the following ways:

a) We will add a clear definition of $f$ at the beginning:

$f$ represents the deployed machine learning model that we aim to heal. It is the function that makes the predictions on input data and whose performance we're trying to maintain and improve.

In our viability studies, a logistic regression model represents $f$.

b) We'll clarify the relationship between $f$ and $\pi$.

While $f$ is the model making predictions, $\pi$ is the adaptation policy --- a function that determines what actions to take to modify $f$ based on the diagnosed reasons for its performance degradation.

c) We will include an illustrative example.

For instance, if $f$ is a diabetes prediction model and $\pi$ diagnoses that $f$'s performance has degraded due to concept drift, $\pi$ might suggest an action to retrain f with more recent data or to adjust feature weights.

d) We will link the theory of the policy $\pi$ to the experiments.

In our viability studies with $\mathcal{H}$-LLM, the policy $\pi$ is instantiated with an LLM (GPT-4) which uses the diagnosed reasons for model failures (also achieved with an LLM) to propose concrete actions.

e) We will explain how the policy impacts SHML:

Self-healing ML is formalized as "an optimization problem over a space of adaptation actions." This means we aim to find the optimal actions to take each time the model $f$ degrades. These actions are chosen by the policy $\pi$ of the self-healing system $\mathcal{H}$ (Fig. 3). For instance, two different policies $\pi_1$ and $\pi_2$ might propose different actions to take in response to $f$'s performance drop.

(B) Requirement for optimal adaptation actions

You're correct that the assumption of optimal adaptation actions would be very strong. Optimal adaptation actions would assume that we always pick the most optimal action in any situation. However, this is unrealistic --- we rarely really know what is the optimal action to take to improve a model's performance because there are many possible reasons which could have led to the decrease in performance. So, achieving model optimality is often impossible.

However, we believe there may have been a misunderstanding. Our framework doesn't require optimal actions. Rather, we aim to make more informed decisions to try to approximate optimal actions (which are never known in practice). We illustrate this with the model degradation example in Sec. 3.2.

To improve the clarity, we will make the following changes:

a) We will add a paragraph (lines 154-156) explaining that SHML approximates an ideal rather than achieving or assuming perfect optimality.

In real-world ML problems, it is often impossible to determine the optimal action due to the complexity of the problem. SHML attempts to approximate an ideal optimal adaptation strategy instead of achieving perfect optimality. This is done by selecting actions to take by taking into account diagnosis information instead of relying on deterministic actions (such as model retraining).

b) In Section 5.2, we'll add this practical example.

Consider a scenario where both data corruption and concept drift occur simultaneously. A traditional method might simply retrain the model on new data (potentially incorporating corrupted values). In contrast, SHML would diagnose both issues and suggest a two-step adaptation strategy: (a) clean the corrupted data and (b) retrain the model on the drift-adjusted datasets. In our experiments, this has improved the model by about 18% in model accuracy, despite not necessarily being the theoretically optimal action (Sec. 6.1).

c) We'll expand the discussion section with:

While SHML attempts to find optimal adaptations, it does not theoretically guarantee that the adaptations are indeed optimal. While we see this as a substantial improvement over reason-agnostic methods, future research could explore how to obtain theoretical guarantees of optimality within self-healing machine learning.

Thank you

You have helped us improve our paper. Given these changes, we hope you consider revising your assessment of the paper's impact and the corresponding evaluation for NeurIPS. ☺️

Dear reviewer 4TYN,

Thank you for carefully reading our paper. We're glad that you think we address the problem of maintaining machine learning models autonomously and that this lowers the barriers for others. We respond to each point below.

(A) Moving information from the Appendix to the main text

In the camera-ready version, we'll use the additional page to expand the viability section in the main text. We'll include more details on the prompts and input/output examples of $\mathcal{H}$-LLM. We will also highlight which design choices are most important to consider such that adopters can more easily replicate and build their own self-healing systems.

(B) Improving the clarity of Section 3 on self-healing ML

We agree with you that the main section could be more concise and focused. To improve clarity, we will make the following changes:

a) We'll start with a concise overview of SHML:

Self-healing machine learning is a framework for autonomously detecting, diagnosing, and correcting performance degradation in deployed ML models. It aims to maintain model performance in changing environments without constant human intervention.

b) In the box "Self-Healing Machine Learning in a nutshell.", we will add a more intuitive explanation at the beginning.

SHML contains four components: (a) monitoring: continuous assessment of model performance; (b) diagnosis: identification of root causes for performance degradation; (c) adaptation: suggesting possible corrective actions to take in response to degradation, and (d) testing: empirically evaluating the effect of actions on the model's performance. After these steps, the best action is implemented on the ML model. This is illustrated in Fig. 2.

c) We will contrast SHML with traditional approaches at the end of Sec. 3.3 (lines 152-154):

The primary insight of SHML is that the best action to take in response to model degradation depends on the reason for that degradation. This contrasts with standard approaches, which often assume the best approach is independent of the degradation reason. For example, a standard drift adaptation method might continuously retrain the model, which could be suboptimal if the new dataset is corrupt.

We will also make other smaller changes, such as add a clear definition of $f$ at the beginning, clarify the relationship between $f$ and the policy $\pi$, and include an illustrative example.

(C) Question: Is the diagnostic step continuously operational?

The diagnostic step is not continuously operational. It's triggered only when the monitoring component detects performance degradation, as shown in Figures 2 and 3. This design choice is due to the computational intensity of the diagnostic step. In our implementation of $\mathcal{H}$-LLM, we perform multiple language model calls to hypothesize possible model failures, propose actions and implement them. In practice, we find this loop takes about 20 seconds to a minute to finish.

Continuous diagnostics might be feasible in two scenarios:

a) Batch prediction settings: If your model makes predictions in large batches (e.g., every 15 minutes) rather than continuously (e.g., every second), a diagnosis step could potentially run for each batch, even without detected degradation.

b) Future research could explore novel architectures where LLM-based monitoring systems run concurrently and continuously with monitoring. However, to the best of our knowledge, no such systems currently exist.

Actions taken: We will clarify when the diagnosis is triggered in the main paper more clearly and expand on when it might be continuously operational.

(D) Question: performance with large-scale models

The SHML framework is agnostic to the model $f$ (L139-141 in the main text). Therefore, self-healing ML shows superiority over benchmarks whenever understanding the reason for degradation is important, regardless of the learner. We can intuitively explain this with an example: suppose that a new batch of data is corrupt. Using a reason-agnostic approach, such as retraining your model on new data, is suboptimal regardless of the model used. This is because the optimal action requires to remove the corrupted data and then retrain the model on the de-corrupted data.

That said, we've conducted an additional viability study comparing SHML's performance against other adaptation methods for 10 popular ML models commonly used in practice. We include models used for large-scale industry applications, e.g. XGBoost or Random Forest. We simulate real-world degradation by corrupting data at test time and evaluating the performance of each adaptation approach. The results are included in Table 4 in the response pdf. Takeaway: We show that $\mathcal{H}$-LLM outperforms other adaptation methods. This showcases that self-healing ML can benefit any downstream ML model whenever understanding the reason for degradation is important (such as test time corruption).

(E) Limitations

In the camera-ready version, we will expand our limitations.

• We will move some challenges of building self-healing systems from Section 5.1 (L206-216) to the discussion to highlight limitations.
• We will expand on the challenges in diagnosis (L208-212) by giving an illustrative example.
• We will expand on the challenges in adaptation (L213-216) by explaining when root cause identification might be easy/difficult.
• We will outline research directions that can help overcome said challenges.

Thank you

You have helped us improve our paper. Given these changes, we hope you consider revising your assessment of the paper's impact and the corresponding evaluation for NeurIPS. ☺️