Big Transfer (Bit) : General Visual Representation Learning
Big Transfer (Bit) : General Visual Representation Learning
Big Transfer (Bit) : General Visual Representation Learning
Zürich, Switzerland
{akolesnikov,lbeyer,xzhai}@google.com
{jpuigcerver,jessicayung,sylvaingelly,neilhoulsby}@google.com
1 Introduction
Strong performance using deep learning usually requires a large amount of task-
specific data and compute. These per-task requirements can make new tasks
prohibitively expensive. Transfer learning offers a solution: task-specific data
and compute are replaced with a pre-training phase. A network is trained once
on a large, generic dataset, and its weights are then used to initialize subsequent
tasks which can be solved with fewer data points, and less compute [40,44,14].
We revisit a simple paradigm: pre-train on a large supervised source dataset,
and fine-tune the weights on the target task. Numerous improvements to deep
network training have recently been introduced, e.g. [55,62,26,35,22,1,64,67,54,60].
We aim not to introduce a new component or complexity, but to provide a recipe
that uses the minimal number of tricks yet attains excellent performance on many
tasks. We call this recipe “Big Transfer” (BiT).
We train networks on three different scales of datasets. The largest, BiT-L
is trained on the JFT-300M dataset [51], which contains 300 M noisily labelled
?
Equal contribution
2 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
80
80
95
Accuracy [%]
90
80
60
90 60
40
85 80 60
40
100
100
100
full
full
full
full
full
full
full
full
full
full
full
full
full
full
full
10
25
10
25
10
25
10
25
10
1
5
1
5
1
5
5
Fig. 1: Transfer performance of our pre-trained model, BiT-L, the previous state-
of-the-art (SOTA), and a ResNet-50 baseline pre-trained on ILSVRC-2012 to
downstream tasks. Here we consider only methods that are pre-trained inde-
pendently of the final task (generalist representations), like BiT. The bars show
the accuracy when fine-tuning on the full downstream dataset. The curve on the
left-hand side of each plot shows that BiT-L performs well even when transferred
using only few images (1 to 100) per class.
images. We transfer BiT to many diverse tasks; with training set sizes rang-
ing from 1 example per class to 1M total examples. These tasks include Im-
ageNet’s ILSVRC-2012 [10], CIFAR-10/100 [27], Oxford-IIIT Pet [41], Oxford
Flowers-102 [39] (including few-shot variants), and the 1000-sample VTAB-1k
benchmark [66], which consists of 19 diverse datasets. BiT-L attains state-of-
the-art performance on many of these tasks, and is surprisingly effective when
very little downstream data is available (Figure 1). We also train BiT-M on the
public ImageNet-21k dataset, and attain marked improvements over the popular
ILSVRC-2012 pre-training.
Importantly, BiT only needs to be pre-trained once and subsequent fine-
tuning to downstream tasks is cheap. By contrast, other state-of-the-art meth-
ods require extensive training on support data conditioned on the task at hand
[38,61,63]. Not only does BiT require a short fine-tuning protocol for each new
task, but BiT also does not require extensive hyperparameter tuning on new
tasks. Instead, we present a heuristic for setting the hyperparameters for trans-
fer, which works well on our diverse evaluation suite.
We highlight the most important components that make Big Transfer ef-
fective, and provide insight into the interplay between scale, architecture, and
training hyperparameters. For practitioners, we will release the performant BiT-
M model trained on ImageNet-21k.
2 Big Transfer
The first component is scale. It is well-known in deep learning that larger net-
works perform better on their respective tasks [15,48]. Further, it is recognized
that larger datasets require larger architectures to realize benefits, and vice
versa [25,45]. We study the effectiveness of scale (during pre-training) in the
context of transfer learning, including transfer to tasks with very few data-
points. We investigate the interplay between computational budget (training
time), architecture size, and dataset size. For this, we train three BiT models
on three large datasets: ILSVRC-2012 [46] which contains 1.3M images (BiT-S),
ImageNet-21k [10] which contains 14M images (BiT-M), and JFT [51] which
contains 300M images (BiT-L).
The second component is Group Normalization (GN) [60] and Weight Stan-
dardization (WS) [34]. Batch Normalization (BN) [21] is used in most state-of-
the-art vision models to stabilize training. However, we find that BN is detrimen-
tal to Big Transfer for two reasons. First, when training large models with small
per-device batches, BN performs poorly or incurs inter-device synchronization
cost. Second, due to the requirement to update running statistics, BN is detri-
mental for transfer. GN, when combined with WS, has been shown to improve
performance on small-batch training for ImageNet and COCO [34]. Here, we
show that the combination of GN and WS is useful for training with large batch
sizes, and has a significant impact on transfer learning.
test resolution [57]. The latter is well-suited for transfer learning; we include the
resolution change during our fine-tuning step.
We found that MixUp [67] is not useful for pre-training BiT, likely due to the
abundance of data. However, it is sometimes useful for transfer. Interestingly, it is
most useful for mid-sized datasets, and not for few-shot transfer, see Section 3.3
for where we apply MixUp.
Surprisingly, we do not use any of the following forms of regularization during
downstream tuning: weight decay to zero, weight decay to initial parameters [31],
or dropout. Despite the fact that the network is very large—BiT has 928 million
parameters—the performance is surprisingly good without these techniques and
their respective hyperparameters, even when transferring to very small datasets.
We find that setting an appropriate schedule length, i.e. training longer for larger
datasets, provides sufficient regularization.
3 Experiments
We train three upstream models using three datasets at different scales: BiT-S,
BiT-M, and BiT-L. We evaluate these models on many downstream tasks and
attain very strong performance on high and low data regimes.
Table 1: Top-1 accuracy for BiT-L on many datasets using a single model and
single hyperparameter setting per task (BiT-HyperRule). The entries show me-
dian ± standard deviation across 3 fine-tuning runs. Specialist models are those
that condition pre-training on each task, while generalist models, including BiT,
perform task-independent pre-training. (? Concurrent work.)
BiT-L Generalist SOTA Specialist SOTA
ILSVRC-2012 87.54 ± 0.02 86.4 [57] 88.4 [61]?
CIFAR-10 99.37 ± 0.06 99.0 [19] -
CIFAR-100 93.51 ± 0.08 91.7 [55] -
Pets 96.62 ± 0.23 95.9 [19] 97.1 [38]
Flowers 99.63 ± 0.03 98.8 [55] 97.7 [38]
VTAB (19 tasks) 76.29 ± 1.70 70.5 [58] -
MT+fSWA MT+fSWA
CPC v2 ¦+fSWA ¦+fSWA
BiT-L S4L MOAM ReMixMatch MixMatch
72
70 80 70 94
70
70
70 65 93
60
60 68
60
60 50 92
55 66
50
50 40 50 91 64
100
100
100
128
256
100
200
400
100
10
25
10
25
10
25
13
26
64
25
1
5
Fig. 2: Experiments in the low data regime. Left: Transfer performance of BiT-L.
Each point represents the result after training on a balanced random subsample
of the dataset (5 subsamples per dataset). The median across runs is highlighted
by the curves. The variance across data samples is usually low, with the exception
of 1-shot CIFAR-10, which contains only 10 images. Right: We summarize the
state-of-the-art in semi-supervised learning as reference points. Note that a direct
comparison is not meaningful; unlike BiT, semi-supervised methods have access
to extra unlabelled data from the training distribution, but they do not make
use of out-of-distribution labeled data.
This dataset is more than 10 times bigger than ILSVRC-2012, but it is mostly
overlooked by the research community. In Table 2 we demonstrate that BiT-M
trained on ImageNet-21k leads to substantially improved visual representations
compared to the same model trained on ILSVRC-2012 (BiT-S), as measured by
all our benchmarks. In Section 4.2, we discuss pitfalls that may have hindered
wide adoption of ImageNet-21k as a dataset model for pre-training and highlight
crucial components of BiT that enabled success on this large dataset.
For completeness, we also report top-5 accuracy on ILSVRC-2012 with me-
dian ± standard deviation format across 3 runs: 98.46% ± 0.02% for BiT-L,
97.69% ± 0.02% for BiT-M and 95.65% ± 0.03% for BiT-S.
80 90 90 70
Accuracy [%]
BiT-L (1 HP)
VIVI-Ex-100% (4 HPs)
80 85 60
Sup-Rot-100% (4 HPs)
70
Sup-Ex-100% (4 HPs)
70 80 50
VTAB (19 tasks) Natural (7 tasks) Specialized (4 tasks) Structured (8 tasks)
Fig. 3: Results on VTAB (19 tasks) with 1000 examples/task, and the current
SOTA. It compares methods that sweep few hyperparameters per task: either
four hyperparameters in previous work (“4 HPs”) or the single BiT-HyperRule.
Semi-supervised learning also tackles learning with few labels. However, such
approaches are not directly comparable to BiT. BiT uses extra labelled out-of-
domain data, whereas semi-supervised learning uses extra unlabelled in-domain
data. Nevertheless, it is interesting to observe the relative benefits of transfer
from out-of-domain labelled data versus in-domain semi-supervised data. In Fig-
ure 2 we show state-of-the-art results from the semi-supervised learning.
Figure 3 shows the performance of BiT-L on the 19 VTAB-1k tasks. BiT-L
with BiT-HyperRule substantially outperforms the previously reported state-of-
the-art. When looking into performance of VTAB-1k task subsets, BiT is the
best on natural, specialized and structured tasks. The recently-proposed VIVI-
Ex-100% [58] model that employs video data during upstream pre-training shows
very similar performance on the structured tasks.
We investigate heavy per-task hyperparameter tuning in Supplementary Ma-
terial section A and conclude that this further improves performance.
4 Analysis
We analyse various components of BiT: we demonstrate the importance of model
capacity, discuss practical optimization caveats and choice of normalization layer.
0.82
0.76 0.92
ILSVRC-2012 ImageNet-21k JFT-300M ILSVRC-2012 ImageNet-21k JFT-300M ILSVRC-2012 ImageNet-21k JFT-300M
Fig. 5: Effect of upstream data (shown on the x-axis) and model size on down-
stream performance. Note that exclusively using more data or larger models may
hurt performance; instead, both need to be increased in tandem.
10 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
30 50 85
50
40 80
20 40
30 75
10 20 30 70
R50x1 R101x1 R50x3 R101x3 R152x4 R50x1 R101x1 R50x3 R101x3 R152x4 R50x1 R101x1 R50x3 R101x3 R152x4 R50x1 R101x1 R50x3 R101x3 R152x4
Fig. 6: Performance of BiT models in the low-data regime. The x-axis corresponds
to the architecture, where R is short for ResNet. We pre-train on the three up-
stream datasets and evaluate on two downstream datasets: ILSVRC-2012 (left)
and CIFAR-10 (right) with 1 or 5 examples per class. For each scenario, we
train 5 models on random data subsets, represented by the lighter dots. The line
connects the medians of these five runs.
For standard computer vision datasets such as ILSVRC-2012, there are well-
known training procedures that are robust and lead to good performance. Progress
in high-performance computing has made it feasible to learn from much larger
datasets, such as ImageNet-21k, which has 14.2M images compared to ILSVRC-
2012’s 1.28M. However, there are no established procedures for training from
such large datasets. In this section we provide some guidelines.
Big Transfer (BiT): General Visual Representation Learning 11
85 100 100
8 GPU 8 GPU weight decay: 1e-4
84 weeks months weight decay: 1e-5
90
Upstream performance
Upstream performance
ILSVRC-2012 top-1 [%] 80
83 80 60
82 70
40
81
60
80 ILSVRC-2012 20
ImageNet-21k 50
79 0
standard long longer 0.0M 0.5M 1.0M 1.5M 2.0M 0k 20k 40k 60k 80k 100k
Training schedule Training steps Training steps
5 Related Work
Large-scale Weakly Supervised Learning of Representations A number
of prior works use large supervised datasets for pre-training visual representa-
tions [23,51,30,36]. In [23,30] the authors use a dataset containing 100M Flickr
images [56]. This dataset appears to transfer less well than JFT-300M.While
studying the effect of dataset size, [51] show good transfer performance when
training on JFT-300M, despite reporting a large degree of noise (20% precision
errors) in the labels. An even larger, noisily labelled dataset of 3.5B Instagram
images is used in [36]. This increase in dataset size and an improved model
architecture [62] lead to better results when transferring to ILSVRC-2012. We
Big Transfer (BiT): General Visual Representation Learning 13
show that we can attain even better performance with ResNet using JFT-300M
with appropriate adjustments presented in Section 2. The aforementioned pa-
pers focus on transfer to ImageNet classification, and COCO or VOC detection
and segmentation. We show that transfer is also highly effective in the low data
regime, and works well on the broader set of 19 tasks in VTAB [66].
Few-shot Learning Many strategies have been proposed to attain good per-
formance when faced with novel classes and only a few examples per class. Meta-
learning or metric-learning techniques have been proposed to learn with few or
no labels [59,49,52]. However, recent work has shown that a simple linear clas-
sifier on top of pre-trained representations or fine-tuning can attain similar or
better performance [7,37]. The upstream pre-training and downstream few-shot
learning are usually performed on the same domain, with disjoint class labels. In
contrast, our goal is to find a generalist representation which works well when
transferring to many downstream tasks.
14 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
Fig. 8: Cases where BiT-L’s predictions (top word) do not match the ground-
truth labels (bottom word), and hence are counted as top-1 errors. Left: All
mistakes on CIFAR-10, colored by whether five human raters agreed with BiT-
L’s prediction (green), with the ground-truth label (red) or were unsure or dis-
agreed with both (yellow). Right: Selected representative mistakes of BiT-L
on ILSVRC-2012. Top group: The model’s prediction is more representative of
the primary object than the label. Middle group: According to top-1 accuracy
the model is incorrect, but according to top-5 it is correct. Bottom group: The
model’s top-10 predictions are incorrect.
6 Discussion
We revisit classical transfer learning, where a large pre-trained generalist model
is fine-tuned to downstream tasks of interest. We provide a simple recipe which
exploits large scale pre-training to yield good performance on all of these tasks.
BiT uses a clean training and fine-tuning setup, with a small number of carefully
selected components, to balance complexity and performance.
In Figure 8 and the Supplementary Material section D, we take a closer look
at the remaining mistakes that BiT-L makes. In many cases, we see that these
label/prediction mismatches are not true ‘mistakes’: the model’s classification
is valid, but it does not match the label. For example, the model may identify
another prominent object when there are multiple objects in the image, or may
provide an valid classification when the main entity has multiple attributes.
There are also cases of label noise, where the model’s prediction is a better fit
Big Transfer (BiT): General Visual Representation Learning 15
than the ground-truth label. In a quantitative study, we found that around half
of the model’s mistakes on CIFAR-10 are due to ambiguity or label noise (see
Figure 8, left), and in only 19.21% of the ILSVRC-2012 mistakes do human
raters clearly agree with the label over the prediction. Overall, by inspecting
these mistakes, we observe that performance on the standard vision benchmarks
seems to approach a saturation point.
We therefore explore the effectiveness of transfer to two classes of more chal-
lenging tasks: classical image recognition tasks, but with very few labelled exam-
ples to adapt to the new domain, and VTAB, which contains more diverse tasks,
such as spatial localization, tasks from simulated environments, and medical and
satellite imaging tasks. These benchmarks are much further from saturation;
while BiT-L performs well on them, there is still substantial room for further
progress.
7 Acknowledgements
We thank the whole Google Brain team in Zürich and its collaborators for many
fruitful discussions and engineering support. In particular, we thank Andrei
Giurgiu for finding a bug in our data input pipeline, Marcin Michalski for the
naming idea and general helpful advice, and Damien Vincent and Daniel Keysers
for detailed feedback on the initial draft of this paper.
References
1. Athiwaratkun, B., Finzi, M., Izmailov, P., Wilson, A.G.: There are many consistent
explanations of unlabeled data: Why you should average. In: ICLR (2019)
2. Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum,
J., Katz, B.: Objectnet: A large-scale bias-controlled dataset for pushing the lim-
its of object recognition models. In: Advances in Neural Information Processing
Systems (2019)
3. Barz, B., Denzler, J.: Do we train on test data? purging CIFAR of near-duplicates.
CoRR (2019), http://arxiv.org/abs/1902.00423
4. Beery, S., Horn, G.V., Perona, P.: Recognition in terra incognita. CoRR
abs/1807.04975 (2018), http://arxiv.org/abs/1807.04975
5. Berthelot, D., Carlini, N., Cubuk, E.D., Kurakin, A., Sohn, K., Zhang, H., Raf-
fel, C.: ReMixMatch: Semi-supervised learning with distribution alignment and
augmentation anchoring. arXiv preprint arXiv:1911.09785 (2019)
6. Borji, A.: Objectnet dataset: Reanalysis and correction. In: arXiv 2004.02042
(2020)
7. Chen, W., Liu, Y., Kira, Z., Wang, Y.F., Huang, J.: A closer look at few-shot
classification. In: ICLR (2019)
8. Chollet, F.: Xception: Deep learning with depthwise separable convolutions. In:
CVPR (2017)
9. De, S., Smith, S.L.: Batch normalization has multiple benefits: An empirical study
on residual networks (2020), https://openreview.net/forum?id=BJeVklHtPr
10. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale
hierarchical image database. In: CVPR (2009)
16 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
11. Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tul-
loch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: training imagenet in 1
hour. arXiv preprint arXiv:1706.02677 (2017)
12. Goyal, P., Dollár, P., Girshick, R.B., Noordhuis, P., Wesolowski, L., Kyrola, A.,
Tulloch, A., Jia, Y., He, K.: Accurate, large minibatch sgd: Training imagenet in
1 hour. ArXiv abs/1706.02677 (2017)
13. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised
visual representation learning. arXiv preprint arXiv:1911.05722 (2019)
14. He, K., Girshick, R., Dollár, P.: Rethinking imagenet pre-training. In: ICCV (2019)
15. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: CVPR (2016)
16. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks.
In: ECCV (2016)
17. Hénaff, O.J., Razavi, A., Doersch, C., Eslami, S., Oord, A.v.d.: Data-efficient image
recognition with contrastive predictive coding. arXiv preprint arXiv:1905.09272
(2019)
18. Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network.
arXiv preprint arXiv:1503.02531 (2015)
19. Huang, Y., Cheng, Y., Chen, D., Lee, H., Ngiam, J., Le, Q.V., Chen, Z.: GPipe:
Efficient training of giant neural networks using pipeline parallelism. arXiv preprint
arXiv:1811.06965 (2018)
20. Ioffe, S.: Batch renormalization: Towards reducing minibatch dependence in batch-
normalized models. In: NIPS (2017)
21. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by
reducing internal covariate shift. ICML (2015)
22. Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., Wilson, A.G.: Aver-
aging weights leads to wider optima and better generalization. arXiv preprint
arXiv:1803.05407 (2018)
23. Joulin, A., van der Maaten, L., Jabri, A., Vasilache, N.: Learning visual features
from large weakly supervised data. In: ECCV (2016)
24. Jouppi, N.P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates,
S., Bhatia, S., Boden, N., Borchers, A., et al.: In-datacenter performance analysis of
a tensor processing unit. In: International Symposium on Computer Architecture
(ISCA) (2017)
25. Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray,
S., Radford, A., Wu, J., Amodei, D.: Scaling laws for neural language models. arXiv
preprint arXiv:2001.08361 (2020)
26. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980 (2014)
27. Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep.
(2009)
28. Kuznetsova, A., Rom, H., Alldrin, N., Uijlings, J., Krasin, I., Pont-Tuset, J., Ka-
mali, S., Popov, S., Malloci, M., Duerig, T., Ferrari, V.: The open images dataset
v4: Unified image classification, object detection, and visual relationship detection
at scale. arXiv:1811.00982 (2018)
29. van Laarhoven, T.: L2 regularization versus batch and weight normalization. CoRR
(2017)
30. Li, A., Jabri, A., Joulin, A., van der Maaten, L.: Learning visual n-grams from web
data. In: ICCV (2017)
31. Li, X., Grandvalet, Y., Davoine, F.: Explicit inductive bias for transfer learning
with convolutional networks. In: ICML (2018)
Big Transfer (BiT): General Visual Representation Learning 17
32. Li, Z., Arora, S.: An exponential learning rate schedule for deep learning. arXiv
preprint arXiv:1910.07454 (2019)
33. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object
detection. In: ICCV (2017)
34. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P.,
Zitnick, C.L.: Microsoft COCO: Common objects in context. In: ECCV (2014)
35. Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts.
arXiv preprint arXiv:1608.03983 (2016)
36. Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, M., Li, Y., Bharambe,
A., van der Maaten, L.: Exploring the limits of weakly supervised pretraining. In:
ECCV (2018)
37. Nakamura, A., Harada, T.: Revisiting fine-tuning for few-shot learning. arXiv
preprint arXiv:1910.00216 (2019)
38. Ngiam, J., Peng, D., Vasudevan, V., Kornblith, S., Le, Q.V., Pang, R.: Domain
adaptive transfer learning with specialist models. arXiv:1811.07056 (2018)
39. Nilsback, M.E., Zisserman, A.: Automated flower classification over a large num-
ber of classes. In: Indian Conference on Computer Vision, Graphics and Image
Processing (2008)
40. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Transactions on knowledge
and data engineering (2009)
41. Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.V.: Cats and dogs. In: CVPR
(2012)
42. Peyre, J., Laptev, I., Schmid, C., Sivic, J.: Weakly-supervised learning of visual
relations. CoRR abs/1707.09472 (2017), http://arxiv.org/abs/1707.09472
43. Qiao, S., Wang, H., Liu, C., Shen, W., Yuille, A.: Weight standardization. arXiv
preprint arXiv:1903.10520 (2019)
44. Raghu, M., Zhang, C., Kleinberg, J., Bengio, S.: Transfusion: Understanding trans-
fer learning with applications to medical imaging. arXiv:1902.07208 (2019)
45. Rosenfeld, J.S., Rosenfeld, A., Belinkov, Y., Shavit, N.: A constructive prediction
of the generalization error across scales. In: ICLR (2020)
46. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: ImageNet Large
Scale Visual Recognition Challenge. IJCV (2015)
47. Shetty, R., Schiele, B., Fritz, M.: Not using the car to see the sidewalk: Quantifying
and controlling the effects of context in classification and segmentation. CoRR
abs/1812.06707 (2018), http://arxiv.org/abs/1812.06707
48. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556 (2014)
49. Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In:
NIPS (2017)
50. Sohn, K., Berthelot, D., Li, C.L., Zhang, Z., Carlini, N., Cubuk, E.D., Kurakin,
A., Zhang, H., Raffel, C.: Fixmatch: Simplifying semi-supervised learning with
consistency and confidence. arXiv preprint arXiv:2001.07685 (2020)
51. Sun, C., Shrivastava, A., Singh, S., Gupta, A.: Revisiting unreasonable effectiveness
of data in deep learning era. In: ICCV (2017)
52. Sung, F., Yang, Y., Zhang, L., Xiang, T., Torr, P.H., Hospedales, T.M.: Learning
to compare: Relation network for few-shot learning. In: CVPR (2018)
53. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. In: CVPR (2015)
54. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the incep-
tion architecture for computer vision. In: CVPR (2016)
18 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
55. Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural
networks. In: ICML (2019)
56. Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth,
D., Li, L.J.: Yfcc100m: The new data in multimedia research. arXiv preprint
arXiv:1503.01817 (2015)
57. Touvron, H., Vedaldi, A., Douze, M., Jégou, H.: Fixing the train-test resolution
discrepancy. In: NeurIPS (2019)
58. Tschannen, M., Djolonga, J., Ritter, M., Mahendran, A., Houlsby, N., Gelly, S.,
Lucic, M.: Self-supervised learning of video-induced visual invariances (2019)
59. Vinyals, O., Blundell, C., Lillicrap, T., Wierstra, D., et al.: Matching networks for
one shot learning. In: NIPS (2016)
60. Wu, Y., He, K.: Group normalization. In: ECCV (2018)
61. Xie, Q., Hovy, E., Luong, M.T., Le, Q.V.: Self-training with noisy student improves
imagenet classification. arXiv preprint arXiv:1911.04252 (2019)
62. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations
for deep neural networks. In: CVPR (2017)
63. Yalniz, I.Z., Jégou, H., Chen, K., Paluri, M., Mahajan, D.: Billion-scale semi-
supervised learning for image classification. arXiv preprint arXiv:1905.00546 (2019)
64. Yun, S., Han, D., Oh, S.J., Chun, S., Choe, J., Yoo, Y.: Cutmix: Regulariza-
tion strategy to train strong classifiers with localizable features. arXiv preprint
arXiv:1905.04899 (2019)
65. Zhai, X., Oliver, A., Kolesnikov, A., Beyer, L.: S4 L: Self-Supervised Semi-
Supervised Learning. In: ICCV (2019)
66. Zhai, X., Puigcerver, J., Kolesnikov, A., Ruyssen, P., Riquelme, C., Lucic, M., Djo-
longa, J., Pinto, A.S., Neumann, M., Dosovitskiy, A., et al.: A large-scale study of
representation learning with the visual task adaptation benchmark. arXiv preprint
arXiv:1910.04867 (2019)
67. Zhang, H., Cisse, M., Dauphin, Y.N., Lopez-Paz, D.: mixup: Beyond empirical risk
minimization. In: ICLR (2017)
Big Transfer (BiT): General Visual Representation Learning 19
Fig. 9: Blue curves display VTAB-1k score (mean accuracy across tasks) depend-
ing on the total number of random hyperparameters tested. Reported VTAB-1k
scores are averaged over 100 random hyperparameter orderings, the shaded blue
area indicates the standard error. Dashed gray line displays the performance on
the small hold-out validation split with 200 examples.
– Initial learning rate is sampled log-uniformly from the range [10−1 , 10−4 ].
– Total number of updates is sampled from the set {500, 1000, 2000, 4000, 8000,
16000}.
– Dropout rate for the penultimate layer is uniformly sampled from the range
[0.0, 0.7].
– Weight decay to the initial weight values is sampled log-uniformly from the
range [10−1 , 10−6 ] .
– MixUp α parameter is sampled from the set {None, 0.05, 0.1, 0.2, 0.4}.
– Input image resolution is sampled from the set {64, 128, 192, 256, 320, 384}.
20 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
Figure 10 shows more results on the ObjectNet test set, with top-5 accuracy
reported on the left and top-1 accuracy on the right. In Figure 10 (a), we first
resize the shorter side of the image to 512 pixels and then take a 480×480 pixel
sized central crop, similar to BiT-HyperRule.
ObjectNet is a dataset collected in the real world, where multiple objects
are present most of the time. A recent analysis shows that cropping out a single
object from the cluttered scene could significantly improve performance [6]. In
Figure 10 (b), we follow the same setup and report our models’ performance
on the cropped ObjectNet with a single object in each crop. We observe a solid
improvement in performance in this setting.
Overall, the trend of our improvements is consistent with the results on the
original ObjectNet test set. We provide our full numeric results in Table 6.
0.70 0.50
Top-5 Accuracy [%]
0.65 0.45
0.60 0.40
NASNet-A [2]
NASNet-A [2]
0.55 0.35
Inception-v4 [2] PNASNet-5L [2] PNASNet-5L [2]
0.50 0.30 Inception-v4 [2]
R152x1 [2] R152x1 [2]
0.45 0.25
R50x1 R101x1 R50x3 R101x3 R152x4 R50x1 R101x1 R50x3 R101x3 R152x4
Architecture Architecture
(a) Results on the original ObjectNet test set with resize and central crop.
0.65
0.85
0.60
0.80
0.55
Top-5 Accuracy [%]
0.75
0.50
0.70
0.45
0.65
R152x1 [6] R152x1 [6]
0.40
0.60
0.35
0.55
R50x1 R101x1 R50x3 R101x3 R152x4 R50x1 R101x1 R50x3 R101x3 R152x4
Architecture Architecture
(b) Results on the cropped ObjectNet, where individual objects are cropped
for evaluation. The bounding boxes are provided by [6].
Fig. 10: Results on ObjectNet; left: top-5 accuracy, right: top-1 accuracy.
Big Transfer (BiT): General Visual Representation Learning 21
Table 6: Results (%) on the ObjectNet test set. We report numbers for both the
standard setting, as well as for the setting where the ground-truth bounding box
is used.
Top-1 accuracy Top-5 accuracy
Resize & Crop Bounding Box Resize & Crop Bounding Box
BiT- S M L S M L S M L S M L
R50x1 30.8 35.0 37.6 35.1 41.6 42.5 51.8 56.4 59.5 58.7 64.9 66.0
R101x1 32.2 39.2 54.6 37.4 46.1 49.1 54.2 61.3 75.6 61.1 69.4 72.4
R50x3 33.7 40.3 49.1 38.4 46.2 54.7 54.7 62.4 71.1 61.5 70.1 77.5
R101x3 34.6 44.3 54.6 40.2 50.5 60.4 56.4 66.4 75.6 63.4 73.6 82.5
R152x4 36.0 47.0 58.7 41.6 52.8 63.8 57.0 69.0 80.0 64.4 76.0 85.1
22 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
Fig. 11: Detected duplicates between the ILSVRC-2012 training set and test
splits of various downstream datasets. Note that Flowers is not listed because
there are no duplicates. Green borders mark true positives and red borders mark
(rare) false positives.
24 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
Here we take a closer look at the mistakes made by BiT-L2 . Figure 8, we show
all mistakes on CIFAR-10, as well as a representative selection of mistakes on
ILSVRC-2012. Figures 12 and 13 again show all mistakes on the Pets and Flowers
datasets, respectively. The first word always represents the model’s prediction,
while the second word represents the ground-truth label. The larger panels are
best viewed on screen, where they can be magnified.
2
To be precise, the figures are obtained by an earlier version of our BiT-L model
but which reaches almost the same accuracy. We did not re-run the figures and
human evaluation with the latest model as they serve for illustration purposes and
the models perform essentially the same, modulo a few flips.
Big Transfer (BiT): General Visual Representation Learning 25
It has been shown that CNNs can lack robustness when classifying objects out-
of-context [4,42,47]. We investigate whether BiT not only improves classification
accuracy, but also out-of-context robustness. For this, we create a dataset of
foreground objects corresponding to ILSVRC-2012 classes pasted onto miscella-
neous backgrounds (Fig. 14 left). We obtain images of foreground objects using
OpenImages-v5 [28] segmentation masks. Figure 14 shows an example, and more
are given in Figure 15. Sometime foreground objects are partially occluded, re-
sulting in an additional challenge.
We transfer BiT models pre-trained on various datasets to ILSVRC-2012 and
see how they perform on this out-of-context dataset. In Figure 14 we can see
that the performance of models pre-trained on ILSVRC-2012 saturates on the
out-of-context dataset, whereas by using more data during pre-training of larger
models, better performance on ILSVRC-2012 does translate to better out-of-
context performance.
More qualitatively, when we look at the predictions of the models on out-
of-context data, we observe a tendency for BiT-L to confidently classify the
foreground object regardless of the context, while ILSVRC-2012 models also
predict objects absent from the image, but that could plausibly appear with the
background. An example of this is shown in Figure 14 left.
Foreground objects:
– Traffic light: U turn to Tophane by Istanbul Photo Guide.
– Sofa: Welcome by woot.
– Zebra: i like his tail in this one by meg and rahul.
– Starfish: Starfish by Summer Skyes 11.
– Limousine: Hummer limousine stopping at the door [nb: title translated] by
duncan su.
Backgrounds:
– Grass: Photo by zoosnow
(Pexels license; Free to use, no attribution required).
– Wood: Tree Bark Texture 04 by Jacob Gube, SixRevisions.
– Street at night: City street calm buildings by csr ch
(Pixabay license; Free for commercial use, no attribution required).
– Underwater: Photo by MaxX42
(Pixabay license; Free for commercial use, no attribution required).
– Kitchen: Interior of a modern modular home by Riverview Homes, Inc.
(CC-BY-SA-3.0 Unported license).
28 Kolesnikov,? Beyer,? Zhai,? Puigcerver, Yung, Gelly, Houlsby
triceratops starfish
palace sea urchin
solar dish knot
starfish sea cucumber
harp coral reef
1e-4 0.01 0.1 1.0 1e-4 0.01 0.1 1.0
Probs (INet-R50) Probs (BiT-L)
loggerhead limousine
coral reef coral reef
wreck scuba diver
beach wagon electric ray
limousine lionfish
1e-4 0.01 0.1 1.0 1e-4 0.01 0.1 1.0
Probs (INet-R50) Probs (BiT-L)
Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.
Alternative Proxies: