Large Language Models (LLMs) like PaLM or GPT-3 confirmed that scaling transformers to lots of of billions of parameters improves efficiency and unlocks emergent talents. The greatest dense fashions for picture understanding, nevertheless, have reached solely 4 billion parameters, regardless of analysis indicating that promising multimodal fashions like PaLI proceed to learn from scaling imaginative and prescient fashions alongside their language counterparts. Motivated by this, and the outcomes from scaling LLMs, we determined to undertake the following step within the journey of scaling the Vision Transformer.
In “Scaling Vision Transformers to 22 Billion Parameters”, we introduce the largest dense imaginative and prescient mannequin, ViT-22B. It is 5.5x bigger than the earlier largest imaginative and prescient spine, ViT-e, which has 4 billion parameters. To allow this scaling, ViT-22B incorporates concepts from scaling textual content fashions like PaLM, with enhancements to each coaching stability (utilizing QK normalization) and coaching effectivity (with a novel strategy referred to as asynchronous parallel linear operations). As a results of its modified structure, environment friendly sharding recipe, and bespoke implementation, it was capable of be skilled on Cloud TPUs with a excessive {hardware} utilization1. ViT-22B advances the state-of-the-art on many imaginative and prescient duties utilizing frozen representations, or with full fine-tuning. Further, the mannequin has additionally been efficiently utilized in PaLM-e, which confirmed that a big mannequin combining ViT-22B with a language mannequin can considerably advance the state-of-the-art in robotics duties.
Architecture
Our work builds on many advances from LLMs, akin to PaLM and GPT-3. Compared to the usual Vision Transformer structure, we use parallel layers, an strategy wherein attention and MLP blocks are executed in parallel, as an alternative of sequentially as in the usual Transformer. This strategy was utilized in PaLM and diminished coaching time by 15%.
Secondly, ViT-22B omits biases within the QKV projections, a part of the self-attention mechanism, and within the LayerNorms, which will increase utilization by 3%. The diagram beneath reveals the modified transformer structure utilized in ViT-22B:
 |
| ViT-22B transformer encoder structure makes use of parallel feed-forward layers, omits biases in QKV and LayerNorm layers and normalizes Query and Key projections. |
Models at this scale necessitate “sharding” — distributing the mannequin parameters in several compute units. Alongside this, we additionally shard the activations (the intermediate representations of an enter). Even one thing so simple as a matrix multiplication necessitates additional care, as each the enter and the matrix itself are distributed throughout units. We develop an strategy referred to as asynchronous parallel linear operations, whereby communications of activations and weights between units happen similtaneously computations within the matrix multiply unit (the a part of the TPU holding the overwhelming majority of the computational capability). This asynchronous strategy minimizes the time ready on incoming communication, thus rising machine effectivity. The animation beneath reveals an instance computation and communication sample for a matrix multiplication.
 |
| Asynchronized parallel linear operation. The objective is to compute the matrix multiplication y = Ax, however each the matrix A and activation x are distributed throughout completely different units. Here we illustrate how it may be finished with overlapping communication and computation throughout units. The matrix A is column-sharded throughout the units, every holding a contiguous slice, every block represented as Aij. More particulars are within the paper. |
At first, the brand new mannequin scale resulted in extreme coaching instabilities. The normalization strategy of Gilmer et al. (2023, upcoming) resolved these points, enabling clean and steady mannequin coaching; that is illustrated beneath with instance coaching progressions.
 |
| The impact of normalizing the queries and keys (QK normalization) within the self-attention layer on the coaching dynamics. Without QK normalization (purple) gradients change into unstable and the coaching loss diverges. |
Results
Here we spotlight some outcomes of ViT-22B. Note that within the paper we additionally discover a number of different downside domains, like video classification, depth estimation, and semantic segmentation.
To illustrate the richness of the discovered illustration, we prepare a textual content mannequin to provide representations that align textual content and picture representations (utilizing LiT-tuning). Below we present a number of outcomes for out-of-distribution pictures generated by Parti and Imagen:
 |
| Examples of picture+textual content understanding for ViT-22B paired with a textual content mannequin. The graph reveals normalized chance distribution for every description of a picture. |
Human object recognition alignment
To learn how aligned ViT-22B classification selections are with human classification selections, we evaluated ViT-22B fine-tuned with completely different resolutions on out-of-distribution (OOD) datasets for which human comparability information is out there through the model-vs-human toolbox. This toolbox measures three key metrics: How nicely do fashions deal with distortions (accuracy)? How completely different are human and mannequin accuracies (accuracy distinction)? Finally, how comparable are human and mannequin error patterns (error consistency)? While not all fine-tuning resolutions carry out equally nicely, ViT-22B variants are state-of-the-art for all three metrics. Furthermore, the ViT-22B fashions even have the very best ever recorded form bias in imaginative and prescient fashions. This signifies that they principally use object form, relatively than object texture, to tell classification selections — a technique identified from human notion (which has a form bias of 96%). Standard fashions (e.g., ResNet-50, which has aa ~20–30% form bias) typically classify pictures just like the cat with elephant texture beneath in keeping with the feel (elephant); fashions with a excessive form bias are inclined to concentrate on the form as an alternative (cat). While there are nonetheless many vital variations between human and mannequin notion, ViT-22B reveals elevated similarities to human visible object recognition.
 |
| Cat or elephant? Car or clock? Bird or bicycle? Example pictures with the form of 1 object and the feel of a unique object, used to measure form/texture bias. |
 |
| Shape bias analysis (greater = extra shape-biased). Many imaginative and prescient fashions have a low form / excessive texture bias, whereas ViT-22B fine-tuned on ImageNet (purple, inexperienced, blue skilled on 4B pictures as indicated by brackets after mannequin names, except skilled on ImageNet solely) have the very best form bias recorded in a ML mannequin thus far, bringing them nearer to a human-like form bias. |
Out-of-distribution efficiency
Measuring efficiency on OOD datasets helps assess generalization. In this experiment we assemble label-maps (mappings of labels between datasets) from JFT to ImageNet and in addition from ImageNet to completely different out-of-distribution datasets like ObjectNet (outcomes after pre-training on this information proven within the left curve beneath). Then the fashions are totally fine-tuned on ImageNet.
We observe that scaling Vision Transformers will increase OOD efficiency: regardless that ImageNet accuracy saturates, we see a major improve on ObjectNet from ViT-e to ViT-22B (proven by the three orange dots within the higher proper beneath).
 |
| Even although ImageNet accuracy saturates, we see a major improve in efficiency on ObjectNet from ViT-e/14 to ViT-22B. |
Linear probe
Linear probe is a method the place a single linear layer is skilled on prime of a frozen mannequin. Compared to full fine-tuning, that is less expensive to coach and simpler to arrange. We noticed that the linear probe of ViT-22B efficiency approaches that of state-of-the-art full fine-tuning of smaller fashions utilizing high-resolution pictures (coaching with greater decision is usually far more costly, however for a lot of duties it yields higher outcomes). Here are outcomes of a linear probe skilled on the ImageNet dataset and evaluated on the ImageNet validation dataset and different OOD ImageNet datasets.
Distillation
The data of the larger mannequin may be transferred to a smaller mannequin utilizing the distillation technique. This is useful as large fashions are slower and dearer to make use of. We discovered that ViT-22B data may be transferred to smaller fashions like ViT-B/16 and ViT-L/16, reaching a brand new state-of-the-art on ImageNet for these mannequin sizes.
Fairness and bias
ML fashions may be prone to unintended unfair biases, akin to selecting up spurious correlations (measured utilizing demographic parity) or having efficiency gaps throughout subgroups. We present that scaling up the dimensions helps in mitigating such points.
First, scale presents a extra favorable tradeoff frontier — efficiency improves with scale even when the mannequin is post-processed after coaching to regulate its degree of demographic parity beneath a prescribed, tolerable degree. Importantly, this holds not solely when efficiency is measured when it comes to accuracy, but in addition different metrics, akin to calibration, which is a statistical measure of the truthfulness of the mannequin’s estimated chances. Second, classification of all subgroups tends to enhance with scale as demonstrated beneath. Third, ViT-22B reduces the efficiency hole throughout subgroups.
 |
 |
| Top: Accuracy for every subgroup in CelebA earlier than debiasing. Bottom: The y-axis reveals absolutely the distinction in efficiency throughout the 2 particular subgroups highlighted on this instance: females and males. ViT-22B has a small hole in efficiency in comparison with smaller ViT architectures. |
Conclusions
We have offered ViT-22B, at present the biggest imaginative and prescient transformer mannequin at 22 billion parameters. With small however essential modifications to the unique structure, we achieved glorious {hardware} utilization and coaching stability, yielding a mannequin that advances the state-of-the-art on a number of benchmarks. Great efficiency may be achieved utilizing the frozen mannequin to provide embeddings after which coaching skinny layers on prime. Our evaluations additional present that ViT-22B reveals elevated similarities to human visible notion with regards to form and texture bias, and presents advantages in equity and robustness, when in comparison with current fashions.
Acknowledgements
This is a joint work of Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin Fathy, Elsayed Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetić, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers Jeremiah Harmsen, and Neil Houlsby
We want to thank Jasper Uijlings, Jeremy Cohen, Arushi Goel, Radu Soricut, Xingyi Zhou, Lluis Castrejon, Adam Paszke, Joelle Barral, Federico Lebron, Blake Hechtman, and Peter Hawkins. Their experience and unwavering assist performed a vital function within the completion of this paper. We additionally acknowledge the collaboration and dedication of the gifted researchers and engineers at Google Research.
1Note: ViT-22B has 54.9% mannequin FLOPs utilization (MFU) whereas PaLM reported
46.2% MFU and we measured 44.0% MFU for ViT-e on the identical {hardware}. ↩