Recently, I have been attracted by the paper "Stable learning establishes some common ground between causal inference and machine learning" published in Nature Machine Intelligence Journal. After perusing it, I met with a problem regarding the comparison between transfer learning and stable learning:
The authors compare stable learning with transfer learning as shown in Fig.3 in the paper. Since I am not quite familiar with the performance of SOTA transfer learning techniques, I would question that in the real-world scenarios where the distributions of testing data are known, if we can use transfer learning to maximize Acc_n for each n as much as possible, we are sure to achieve better Avg(Acc_1...n) than stable learning. Does it mean that stable learning is not so practical and useful in this specific scenario?
Dear Wenzhao,
One can argue that both IID and transfer learning end up learning the assumptions and correlations inherent in the original dataset and not just the true connection between discriminative features and labels.
Our of distribution (OOD) generalization, like stable learning, is used when the standard statistical assumptions for IID distributions do not hold. Like many statistical tools, certain making certain assumptions can improve accuracy when those assumptions hold, but they will not be generalizable. Therefore, I believe that stable learning can be very useful for real world tasks like robotic gesture training, where the IID assumptions likely do not hold.
https://arxiv.org/pdf/2104.07876.pdf
Haseeb Javed
, Nathan Zavanelli Thanks for your answers! But my curiosity might mainly lie in a more practical setting: now we are given one task for training and another n tasks with different known distributions. For each new task i, we can apply transfer learning to maximize Acc_i as much as possible. Then we are sure to get a high Avg(Acc_{1...n}). However, stable learning doesn't focus on one specific task, which means that on that specific task, it might not achieve higher Acc_i than transfer learning. In this sense, Avg(Acc_{1...n}) output by stable learning might also fail to surpass that of transfer learning. So if this is true, when dealing with this kind of scenario, transfer learning still seems to be more practical than stable learning?
If we can assume that the actual, real life data that the algorithm will see when applied follows the same assumptions as the training and validation data, then yes, transfer learning will outperform. Sometimes, however, this might not be the case. This transfer learning approach achieves higher accuracy precisely because it bakes in these assumptions. The real world can often be a bit messier than we assume in certain instances though.
Transfer learning (TL) is a research problem in machine learning (ML) that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem. For example, knowledge gained while learning to recognize cars could apply when trying to recognize trucks.
Recently, as the applications of artificial intelligence gradually seeping into some risk-sensitive areas such as justice, healthcare and autonomous driving, an upsurge of research interest on model stability and robustness has arisen in the field of machine learning. Rather than purely fitting the observed training data, stable learning tries to learn a model with uniformly good performance under non-stationary and agnostic testing data. The key challenge of stable learning in practice is that we do not have any knowledge about the true model and test data distribution as a priori. Under such condition, we cannot expect a faithful estimation of model parameters and its stability over wild changing environments. Previous methods resort to a reweighting scheme to remove the correlations between all the variables through a set of new sample weights. However, we argue that such aggressive decorrelation between all the variables may cause the over-reduced sample size, which leads to the variance inflation and possible underperformance.
https://www.kdd.org/kdd2020/accepted-papers/view/stable-learning-via-differentiated-variable-decorrelation
https://arxiv.org/abs/1911.12580
https://www.nature.com/articles/s42256-022-00445-z
- Badges
- Science topic
- Similar topics
- Biological Science
- Bioecology
- Systems Ecology
- Landscape Ecology
- Connectivity