NeurIPS 2020

Multi-Stage Influence Function

Review 1

Summary and Contributions: The paper proposes a novel method to estimate the influence score for multi-stage training, which is widely used in current state-of-the-art systems. The authors also empirically show the proposed method’s effectiveness with experiments.

Strengths: The proposed method is novel and technically sound. The problem the authors trying to address is important for many state-of-the-art deep models.

Weaknesses: The empirical studies in its current form are weak. The authors only conduct experiments on synthetic/simplified datasets, namely Cifar10 with two selected classes and MNIST with four classes. Also, for the NLP task, the authors only present results with a few selected samples. Such empirical study makes the paper quite weak. I think the paper would be significantly improved if the authros could apply the proposed method to real tasks to show its effectiveness. For example, research has shown that ImageNet contains some corrupted/incorrect labelled images and thus negatively impact some systems’ performance. I wonder if the proposed method can be used to improve the performance of those systems? *********************************after rebuttal*************************** Thank you for your feedback on my review; really appreciate that. After reading the other reviews and the rebuttal. This remains for me a marginal paper.

Correctness: The method sounds interesting and correct. But I think the empirical studies are quite weak.

Clarity: The paper is generally well written and easy to follow. Typo: line 29: “. From a robust…”

Relation to Prior Work: Yes; to the best of my knowledge, the prior work is well presented and discussed.

Reproducibility: Yes

Additional Feedback: Can the proposed method be used to guide the pre-training processes to improve the current state-of-the-art systems for some real tasks using BERT pre-training or ImageNet pre-training?

Review 2

Summary and Contributions: This paper addresses a longstanding and difficult problem: how can one trace backward from a final result, through various layers of transformation, to determine the culprits in the initial training set for a poor machine learning model? This paper proposes a multi-stage influence function ‘propagation’ score to trace eventual labels back to their sources. It is an extension to multiple layers of (I presume) the same authors’ single-stage model from 2017.

Strengths: The paper is clear and logically organized. I did not have time to check the math in Section 3, but the experiments in Section 4 are well-crafted and executed and the various kinds of results seem convincing, though lacking a higher-level discussion of are properly reported.

Weaknesses: What is somewhat lacking though is a higher-level discussion of what one could/should expect in principle to be actually useful; for example, let’s say you can identify a few problematic samples in a training set, is it sufficient to simply delete them? How much damage would you do thereby to predictions for other tasks in the future? That is, how ‘stable’ is the initial training data to perturbations, and the subsequent one?

Correctness: I did not do a thorough check but seems ok on the surface.

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: Influence function is used to identify the training examples which contribute most to the prediction of the model. This paper studies multi-stage influence functions, which means that the model is trained by more than one stages, e.g., firstly pre-trained on an auxiliary task and then finetuned on the target task. The paper studies the multi-stage Influence functions under two different fine-tuning settings: (1) fine-tuning all the parameters, and (2) fine-tuning partial parameters. Experiments are conducted to validate the proposed method.

Strengths: 1. The authors introduce an interesting and understudied problem. The paper is overall clearly written and easy to follow, although I have no prior experience with influence functions. 2. Two widely used fine-tuning methods, fine-tuning all the parameters and fine-tuning partial parameters, are discussed to devise the influence functions. Experiments show somewhat positive results to validate the proposed method.

Weaknesses: 1. My main concern is that the authors did not compare the proposed method with any latent baselines. For example, we can also use to the uncertainty (the loss value or the entropy of the predicted distribution) of the model as an indicator to identify those problematic examples in the pre-training data. As the proposed method in this paper does not show very impressive results in the experiments (the Pearson’s correlation is only 0.4~0.6 in Fig. 1), it may not outperform this simple baseline. 2. In the experiments, the transfer tasks come too artificially. “At the pretraining stage, we train the models with examples from two classes (“bird" vs. “frog") for CIFAR-10 and four classes (0, 1, 2, and 3) for MNIST”. The transfer tasks in these settings may be too easy. In addition, the experiments are limited to small datasets like MNIST and CIFAR-10. I wonder how the proposed method scale to larger datasets or more challenging transfer tasks.

Correctness: Yes

Clarity: Yes

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback:

Review 4

Summary and Contributions: This paper addresses the problem of estimating the influence of specific training data points on the predictions of a model. Specifically, the work considers multi-stage training where a model is first pre-trained and then finetuned to a task. The paper considers two cases --- first, when the pretrained parameters are kept fixed; and second, when the pretrained parameters are also updated during finetuning --- and derives efficient algorithms to compute the influence scores under these settings. The experiments showed that the proposed approach can approximate the true influence of pre-training data points in downstream models and can be used to clean the training data and thus avoid negative transfer issues.

Strengths: Deep learning models induced by multi-stage training are increasingly becoming the norm in ML research and applications and yet the underlying mechanisms the govern this process are still poorly understood. This paper is a step in the direction of a creating more systematic and controlled methodologies for multi-stage training.

Weaknesses: None == Updated == I agree that the empirical is somewhat weak in that it does not reflect the difficulty and scale of modern ML/DL tasks. However, I understand that it would be very computationally demanding to make those experiments.

Correctness: The proposed approach is sound and the experimental results are convincing.

Clarity: The paper is well written and organized

Relation to Prior Work: Yes

Reproducibility: Yes

Additional Feedback: typo in line 84: perspective instead of prospective