6 ms·

When someone used the word “overfitting” I usually take that to mean that a model has begun to enter the phase where further improvement is leading to lower gen

When someone used the word “overfitting” I usually take that to mean that a model has begun to enter the phase where further improvement is leading to lower generalization. In fact, as far as I can tell, we are not overfitting in this sense. When I have seen papers examine whether progress on, let’s say, imagenet, actually generalizes to other categorization datasets the answer is yes. What we have been seeing is that the slope of this graph is flattening out a bit. Whereas in the past a 1% improvement on imagenet would have meant a 1% improvement on a similarly collected dataset, nowadays it will be more like .5% (not exact numbers just using numbers to illustrate what I mean by diminishing returns.) If an institution or a lab can show that progress on their dataset -better- predicts the progress on a bunch of other closely related tasks, then as researchers become convinced of that, they will switch over. Right now there isn’t a great alternative because it’s not easy to create such a dataset. Scale is critical. Imagenet really was on the right track as far as collecting images of nearly every semantic concept in the English language. So whatever replaces it will have to be similarly thorough and broad. In my opinion the biggest weakness of currently existing datasets is that they are typically labeled once per image with no review step. So I think the answer here isn’t “Let’s get researchers to use smaller datasets from smaller institutions” It would be more like “We have to figure out a way to get a bigger, cleaner version of existing datasets and then prove that progress on those datasets is more meaningful” The realistic way this plays out is that some institution in the “cartel” releases a better dataset and then lots of small labs try it out and show that progress on that dataset better predicts progress in general.

mattkrause 5y ago
I disagree: some "collective overfitting" happens when everyone evaluates on the same dataset (and often, the same test set). There's a neat set of papers by Recht et al. showing results are slightly overfit to the test partitions of ImageNet and CIFAR-10: rotating examples between the train and test partitions causes systems to perform up to 10-15% worse. https://arxiv.org/abs/1806.00451 https://arxiv.org/abs/1806.00451 https://arxiv.org/abs/1902.10811 https://arxiv.org/abs/1902.10811 There's another neat bit of work involving MNIST. The original dataset (from the mid-90s) had 60,000 test examples, but the distributed versions that virtually everyone uses has only 10,000 test examples. Performance on these held-out examples is, unsurprisingly, a bit worse: https://arxiv.org/pdf/1905.10498.pdf https://arxiv.org/pdf/1905.10498.pdf
- dontreact 5y ago
  I think it's super important to separate the following two situations, both of which I suppose are fair to call overfitting Situation A: Models are slightly overfit to some portions of the test set. But the following holds IF PerformanceOnBenchmark(Model A) > PerformanceOnBenchmark(Model B) Then PerformanceOnSimilarDaset(Model A) > PerformanceOnSimilarDataset(Model B) Therefore progress on the benchmark is predictive of progress in general. Situation B: The relation does not hold, and therefore progress on the benchmark does not predict general progress. This almost always happen if you train a deep neural network long enough: train performance goes up, but test performance goes down. If you look at figure 2 of the first paper you sent, you will note that it shows we are in situation A and not situation B. Situation A overfitting = diminishing returns on improvements on benchmark, but the benchmark is still useful. Situation B overfitting = the benchmark is now useless
  - sdenton4 5y ago
    >> "This almost always happen if you train a deep neural network long enough: train performance goes up, but test performance goes down." This is a problem that is more common for classification problems, I think. Generative and self-supervised models (trained with augmentation) tend to just get better forever (with some asymptote) because memorization isn't a viable strategy. I personally think image classification is mostly a silly problem to judge new algos on as a result, and leads to all kinds of nonsense as people try to extrapolate meaning from new results.
    - dontreact 5y ago
      Nearly all useful machine learning is supervised, still. And if you are using a neural network, it will eventually memorize. This is fine though, we have early stopping :)
      - sdenton4 5y ago
        Eh, I work in audio ml, where most of the interesting and useful work is in conditioned generative models. Compression, TTS, source separation. The bias towards classification is a side effect of people starting at imagenet too long; I really think it's holding the field back.
visarga 5y ago
Lately there have been amazing results on pairs of web text+image, no need for labelling. These datasets are hundreds of times larger and cover many more categories of objects. GPT-3 is also trained on raw text. I think ImageNet and its generation have become toy datasets by now.

mattkrause 5y ago

I disagree: some "collective overfitting" happens when everyone evaluates on the same dataset (and often, the same test set). There's a neat set of papers by Recht et al. showing results are slightly overfit to the test partitions of ImageNet and CIFAR-10: rotating examples between the train and test partitions causes systems to perform up to 10-15% worse. https://arxiv.org/abs/1806.00451 https://arxiv.org/abs/1806.00451 https://arxiv.org/abs/1902.10811 https://arxiv.org/abs/1902.10811 There's another neat bit of work involving MNIST. The original dataset (from the mid-90s) had 60,000 test examples, but the distributed versions that virtually everyone uses has only 10,000 test examples. Performance on these held-out examples is, unsurprisingly, a bit worse: https://arxiv.org/pdf/1905.10498.pdf https://arxiv.org/pdf/1905.10498.pdf

dontreact 5y ago
I think it's super important to separate the following two situations, both of which I suppose are fair to call overfitting Situation A: Models are slightly overfit to some portions of the test set. But the following holds IF PerformanceOnBenchmark(Model A) > PerformanceOnBenchmark(Model B) Then PerformanceOnSimilarDaset(Model A) > PerformanceOnSimilarDataset(Model B) Therefore progress on the benchmark is predictive of progress in general. Situation B: The relation does not hold, and therefore progress on the benchmark does not predict general progress. This almost always happen if you train a deep neural network long enough: train performance goes up, but test performance goes down. If you look at figure 2 of the first paper you sent, you will note that it shows we are in situation A and not situation B. Situation A overfitting = diminishing returns on improvements on benchmark, but the benchmark is still useful. Situation B overfitting = the benchmark is now useless
- sdenton4 5y ago
  >> "This almost always happen if you train a deep neural network long enough: train performance goes up, but test performance goes down." This is a problem that is more common for classification problems, I think. Generative and self-supervised models (trained with augmentation) tend to just get better forever (with some asymptote) because memorization isn't a viable strategy. I personally think image classification is mostly a silly problem to judge new algos on as a result, and leads to all kinds of nonsense as people try to extrapolate meaning from new results.
  - dontreact 5y ago
    Nearly all useful machine learning is supervised, still. And if you are using a neural network, it will eventually memorize. This is fine though, we have early stopping :)
    - sdenton4 5y ago
      Eh, I work in audio ml, where most of the interesting and useful work is in conditioned generative models. Compression, TTS, source separation. The bias towards classification is a side effect of people starting at imagenet too long; I really think it's holding the field back.

visarga 5y ago

Lately there have been amazing results on pairs of web text+image, no need for labelling. These datasets are hundreds of times larger and cover many more categories of objects. GPT-3 is also trained on raw text. I think ImageNet and its generation have become toy datasets by now.