Learning Functional Transduction

Research in machine learning has polarized into two general approaches for regression tasks: Transductive methods construct estimates directly from available data but are usually problem unspecific. Inductive methods can be much more specific but generally require compute-intensive solution searches. In this work, we propose a hybrid approach and show that transductive regression principles can be meta-learned through gradient descent to form efficient in-context neural approximators by leveraging the theory of vector-valued Reproducing Kernel Banach Spaces (RKBS). We apply this approach to function spaces defined over finite and infinite-dimensional spaces (function-valued operators) and show that once trained, the Transducer can almost instantaneously capture an infinity of functional relationships given a few pairs of input and output examples and return new image estimates. We demonstrate the benefit of our meta-learned transductive approach to model complex physical systems influenced by varying external factors with little data at a fraction of the usual deep learning training computational cost for partial differential equations and climate modeling applications.

Paper

References (79)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer w6wT5/10 · confidence 2/52023-06-19

Summary

This paper proposes a new deep learning approach for the problem of meta-learning. Inspired by the theory of reproducing kernel Banach space, the proposed method jointly trains a deep transformation as a representation, and a parametrized kernel function $K(vi, \cdot)$ of the problem instances, at the meta-training stage. Experimental results validate the effectiveness of the proposed method.

Strengths

The motivation is strong, as the problem of meta-learning is a rather important problem in the community. The link between meta-learning and the theory of RKBS discussed in this paper may bring some new ideas.

Weaknesses

My main concern comes from the fact that the proposed method also needs a meta-training procedure. At this point, the comparison in some experiments (for example, section 5.1, if there is no mis-understanding) seems somewhat unfair. Also, it would be better if there is some more comparison between the proposed method and some other meta-learning approaches.

Questions

It would be better if the authors could give a thorough discussion about their method’s superiority, as there are many meta-learning approaches that are in a similar style, that is, first training a meta model in the dataset containing sufficient information of the task space, then apply some task-specific procedures in the target task. It would be better if the authors could clearly state several specific points, at which their method is better than most of the others in the literature.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Contribution

2 fair

Limitations

No.

Reviewer f2BS7/10 · confidence 3/52023-06-27

Summary

This paper proposes a method for transductive learning based on reproducing kernel banach spaces (RKBS). The resulting model is capable of learning *in-context* in the sense that given a new instance of a learning problem or a new dataset $\mathcal{D}$, it can infer the resulting functional relationship at inference time. The authors show that transformer attention layers can be considered reproducing kernels for this model. They provide a meta training objective that allows them to learn the kernel operations $\kappa^\ell$ and nonlinearities $F^\ell$ which generalize well over the distribution $\mathfrak{D}$ of possible datasets. Once trained, the model can generate functions for new datasets. The authors test the proposed ideas to learn operators $\mathcal{O}$ for PDEs including advection reaction diffusion equation and Burgers' equation as well as climate modeling. Lastly, they test on a MNIST like task where there are pixel permuted and class permuted versions of the dataset. In all of these settings they identify benefits to their approach over several existing benchmarks.

Strengths

The paper provides a nice framework and perspective to think about transductive learning through reproducing kernels. To the best of my knowledge, this is a novel idea. The paper also provides several experiments that showcase the benefits of the approach, especially in PDE modeling.

Weaknesses

As one of the primary motivations of the model is a meta-learning algorithm capable of in-context learning, I think that an experiment involving in-context learning of language patterns would greatly strengthen the paper. Most of the experiments at this point are for PDEs but the proposal is much broader.

Questions

On the MNIST experiment, how well would a pure supervised algorithm (perhaps with a provided context signal to indicate permutation) trained on all instances of the data perform? I am wondering if the meta-learning objective outperforms standard supervised learning because it sees a larger amount of total data.

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

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors addressed their need to work with finite dimensional outputs for the PDE modeling (Fourier coefficients) and also acknowledged the computational and statistical requirements to optimize the metal learning objective.

Reviewer Sczv8/10 · confidence 4/52023-06-30

Summary

**SUMMARY AFTER REBUTTAL**: as described below, most of my comments were addressed and the authors made a significant job in including new results. I have increase my score during the rebuttal phase and I strongly vote for acceptance. --- Neural operators are neural networks that are trained to approximate a function-to-function mapping. This paper proposes an algorithm to perform few-shot learning of these models, where at inference time the network is given a set of examples of the operator, and an input function as the test point. The specific algorithm they propose is built on reproducing kernel Banach spaces, where for a given prediction, the kernel and the dictionary on which it is evaluated is computed by a recursive embedding of the few-shot examples. This is inspired by the transformer architecture. When examples are continuous, they are discretized with an FFT. They analyze a series of classical benchmarks including modeling PDEs, showing the method to perform well.

Strengths

As far as I know, the idea of performing few-shot learning of neural operators is novel. These networks are used extensively in physical modelling, so the paper can have a sizable impact there. The paper is also relatively clear (with a few exceptions, see below), especially if one considers the complexity of some underlying ideas. The connection between their method and the transformer is also interesting, but from what I understand this is extended from Wright and Gonzalez (2021). The experiments are varied and cover a wide range of use cases.

Weaknesses

1) I have found the initial discussion on the difference between transduction and inference a bit misleading. According to Vapnik, 2006 (which they cite), almost anything which is used today in ML / deep learning is inductive. Transductive methods are only those that can make predictions on a given set of test points but *cannot* operate outside those. For example, SVM is inductive in general, except for some variants such as the transductive SVMs discussed in Collobert et al., 2006. In fact, what they are calling "transductive" is what is typically called "instance-based" in ML (kNN, SVM). Their setup is a standard few-shot setup extended to operators. This also leads to some strange sentences, such as "inductive neural learning with gradient descent is compute-intensive" referred to neural networks; SVM training is also notoriously intensive (and it can also be done via gradient descent). 2) There are some "standard" methods for performing few-shot learning in the literature (e.g., MAML, prototype networks, ...). Many of these methods start from a standard neural network and adapt it to a few-shot scenario. Architectures for performing operator learning are known, and it is not clear from reading the paper why the few-shot learning methods are not immediately extensible to this setup (e.g., why can't we do MAML on a standard neural operator?), and why we need instead to resort to a more complex formulation in terms of kernels. To clarify: I think the algorithm shown here is interesting, but it's a bit hard to motivate it by reading the paper itself. 3) The computational complexity of the method is not discussed, in particular the need to execute multiple FFTs for the few-shot examples (Section "Discretization"). This also connects to recent literature on performing FFTs in an efficient way on GPUs (e.g., FlashConv). Also for testing, if I understand correctly the authors are mostly comparing their method (which is only a forward pass) with a full training of its competitors, which is a bit unfair. No baselines for few-shot learning are tested. For example, they state that their setup is "Similar to (Pathak et al., 2022)", but they do not compare.

Questions

- I think rephrasing the entire "transductive" discussion in terms of few-shot learning could significantly improve the paper. At the least, objectively incorrect sentences should be amended. - Discussing better training and test complexity is important. - I think the paper can improve readability by providing a practical example of (5)-(6) in the more familiar case of finite-dimensional input-output spaces. - There has been a limited amount of works that have explored using neural networks to perform Bayesian posterior inference on-the-fly on a large family of distributions (e.g., Prior-Data Fitted Networks (PFNs)). I would be curious to see a discussion on the connection with this work. On the related work part, there is also some relevant literature on recursive (recurrent) kernel evaluation which is not mentioned. - There is a small typo on page 3: point evalutation. - Will the code be released?

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

4 excellent

Limitations

The authors are clearly discussing the limitations of their few-shot setup. However, they are not discussing the computational complexity of the method, which I think would be the biggest limitation in scaling this up to large setups. A few sentences on this would be appreciated.

Reviewer f2BS2023-08-11

Response to Authors

I appreciate the authors' answers to my questions about the possibility of further ICL experiments and data volume. Though natural language experiments would strengthen the work, it could also be left as a future application. I will maintain my support for acceptance.

Reviewer w6wT2023-08-12

Thank you for your patience to resolve my confusions! I have now understood the experimental settings in section 5. Though I keep my point that the work has sound theoretical grounding(mostly due to the lack of understandings of DNNs), I admit that the idea of meta-learning neural operators has its novelty. I will no longer stand in the way of acceptance.

Authorsrebuttal2023-08-19

Second reply to reviewer w6wT

To reviewer w6wT, We thank you for your positive re-evaluation of our work. Moreover, we are happy that you praise the theoretical aspect of our work in your reply. To help definitely convince you, as we proposed in our rebuttal, we would like to complement our experimental results in section 5.1 (ADR equations) by providing a comparison with model-agnostic meta-learning (MAML) applied to FNO with different inner training budgets (10/50/100 gradient steps with inner learning rate at 1e-2/7e-3/5e-3), in line with you main suggestion (as well as that of reviewer Sczv). We synthesize the results in terms of RMSE and fine-tuning time in the following table: | RMSE / Adaptation time (sec.) | 10 gradient steps | 50 gradient steps | 100 gradient steps |---|---|---|---| | FNO-MAML | $6.5e^{-1} // 2.6e^{-1}$ | $3.1e^{-1} // 6.7e^{-1}$ | $1.4e^{-1}//2.1e^{0}$ | As you can see, on this task, a meta-learned parameter initialization with a limited fine-tuning budget does not improve over the previously tested fine-tuned approaches. More importantly, this approach is much less accurate and computationally more intense than our system (due to the need for sequential gradient computation). We hope that this control experiment will help dissipate your concern and contrastively demonstrate the potential value of our own approach. We remain happy to answer to any additional comment before the end of the discussion period.

Reviewer Sczv2023-08-16

Thanks for the detailed feedback. Concerning the definition of "transduction", I understand the point but I also think that "transduction" used in this sense can be misleading, especially since there are other definitions that are more common in today's literature. However, this does not impact my evaluation in any way. I will keep my score as-is waiting for the other points. For recursive kernels, the authors might be interested in this paper coming from the kernel signal processing literature, which is a bit niche but connected: https://ieeexplore.ieee.org/document/6722955

Authorsrebuttal2023-08-19

Second reply to reviewer Sczv

To reviewer Sczv, Thank you for your reply as well as the interesting reference. We are happy that our rebuttal was informative. In addition to these, as we mentioned in our rebuttal, we would like to complement our experimental results in section 5.1 (ADR equations) by providing a comparison with an other meta-learning method. While we re-emphasize that meta-learning of neural operators is a new topic with no standard approaches, we tried to your suggestion to apply gradient-based model-agnostic meta-learning on our meta-dataset of operators to the same FNO model with different inner training budgets (10/50/100 gradient steps with fixed inner learning rate at 1e-2/7e-3/5e-3). We synthesize in the following table the results in terms of RMSE and fine-tuning time that will complement table 1 of the paper. | RMSE / Adaptation time (sec.) | 10 gradient steps | 50 gradient steps | 100 gradient steps |---|---|---|---| | FNO-MAML | $6.5e^{-1} // 2.6e^{-1}$ | $ 3.1e^{-1} // 6.7e^{-1}$ | $1.4e^{-1} // 2.1e^{0}$ | As you can see, on this task, a meta-learned parameter initialization with a limited fine-tuning budget does not improve over the previously tested fine-tuned approaches. More importantly, this approach is much less accurate and computationally more intense than our system (due to the need for sequential gradient computation). These complementary results further validate the interest of our own approach. We agree that this comparison will also help the reader better situate our approach in the literature. We remain happy to answer to any additional comment before the end of the discussion period.

Reviewer Sczv2023-08-19

I really appreciate the effort that the authors have put into addressing this specific concern. As I believe this is a very strong addition to the paper, I have increased my score to a Strong Accept and most other scores to Excellent.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC