This paper is to design a generalized multimodal architecture that can solve all Vision language tasks.
Concretely, they will pre-train their model on 4 main tasks (MLM, ITM, WRA, MRM) and will evaluate various downstream tasks (VQA, VCR, NLVR).

https://i.imgur.com/IG7suDj.png

As shown in Fig 1, UNITER first encodes image regions (visual features and bounding box features) and textual words (tokens and positions) into a common embedding space with Image Embedder and Text Embedder.
Then, a Transformer module is applied to learn generalizable contextualized embeddings for each region and each word.

The contribution is two-fold:
(1) Masked language/region modeling is conditioned on full observation of image/text, rather than applying joint random masking to both modalities
(2) Introducing a novel WRA pre-training task via the use of Optimal Transport (OT) to explicitly encourage fine-grained alignment between words and image regions. 
Intuitively, OT-based learning aims to optimize distribution matching by minimizing the cost of transporting one distribution to another. In our context, we aim to minimize the cost of transporting the embeddings from image regions to words in a sentence (and vice versa), thus optimizing toward better cross-modal alignment.

This is a mildly silly paper to summarize, since there isn't really a new mechanism to understand, but rather a number of straightforward (and interesting!) empirical results that are also quite well-explained in the paper itself. That said, for the sake of a tiny bit more brevity than the paper itself provides, I'll try to pull out some of the conclusions I found the most interesting here. 

The general goal of this paper is to better understand the contours of when self-supervised representation learning is valuable for vision (and specifically when it can compete with supervised learning), and when it doesn't. In general, the results are all using ResNet backbones, with SimCLR SSL, on image classification datasets. Some bullet-point takeaways: 

- The SSL models being tested here seem to roughly saturate at unsupervised dataset sizes of around 500K; the comparative jump from dataset sizes of 500K to 1M is fairly small.
- Once you have a supervised dataset of around 50K or more, the benefit of SSL pretraining starts to diminish, and it converges to being more similar to just supervised learning on that numbrer of labeled images. On the flip side, it's only possible to get close to "good" fully supervised performance by using 100K images or more on top of a SSL baseline.
- Even within image classification datasets, it's much better to do SSL representation on the same dataset as the one you'll use for downstream training; trying to transfer representations to different datasets leads to meaningfully worse results. Interestingly, this is even true when you add out-of-domain (i.e. other-dataset) data to an existing in-domain dataset: a dataset of 250K in-dataset images does better than a 500K dataset of images from mixed datasets, and does notably better than a 1M dataset of mixed images. In this case, adding more out-of-domain images seems to have just degraded performance
- SSL seems to perform more closely to SL on a course label set; when the label set gets more granular, the task gets harder overall, but, more specifically, the gap between SSL and SL grows
- When the authors tried different forms of dataset corruption, SSL was much more robust to adding salt-and-pepper noise than it was to removing high-frequency information in the form of reducing the images to a lower resolution.

This paper is an interesting extension of earlier work, in the TransformerXL paper, that sought to give Transformers access to a "memory" beyond the scope of the subsequence where full self-attention was being performed. This was done by caching the activations from prior subsequences, and making them available to the subsequence currently being calculated in a "read-only" way, with gradients not propagated backwards. This had the effect of (1) reducing the maximum memory size compared to simply doubling the subsequence length, and (2) reducing the extent to which gradients had to propagate backward through time. 

The authors of the Compressive Transformers paper want to build on that set of ideas to construct an even longer accessible memory. So, they take the baseline non-backpropogated memory design of TransformerXL, but instead of having tokens roll out of memory after the end of the previous (cached) subsequence, they create an extra compressed memory. Each token in this compressed memory is a function of C inputs in the normal memory. So, if C=3, you would input 3 memory vectors into your compression function to get one instance of a compressed memory vector. Depending on the scale of your C, you can turn up the temporal distance into the past that your compressed memory had to. 

https://i.imgur.com/7BaCzoU.png

While the gradients from the main loss function didn't, as far as I could tell, pass back into the compression function, they did apply a compression loss to incentivize the compression to be coherent. They considered an autoencoder loss to reconstruct the input tokens from the compressed memory, but decided against that on the principle that memory inherently has to be compressed and lossy to be effective, and an autoencoder loss would promote infeasibly lossless compression. Instead, they take the interesting approach of incentivizing the compressed representations to be able to reconstruct the attention calculation performed on the pre-compressed representations. Basically, any information pulled out of the pre-compressed memories by content-based lookup also needs to be able to be pulled out of the compressed memories. This incentives the network to preferentially keep the information that was being actively used by the attention mechanisms in prior steps, and discard less useful information. 

One framing from this paper that I enjoyed was them drawing a comparison between the approach of Transformers (of keeping all lower-level activations in memory, and recombining them "in real time," for each downstream use of that information), and the approach of RNNs (of keeping a running compressed representation of everything seen up to this point). In this frame, their method is somewhere in between, with a tunable compression rate C (by contrast, a RNN would have an effectively unlimited compression rate, since all prior tokens would be compressed into a single state representation).

The idea of the Switch Transformer is to have more parameters available for a network to use, but to only use a small subset of those parameters for each example that's run through the network. This is achieved through a routing scheme, whereby a weighting layer is applied to each token and produces a set of logits/softmax weights over the set of possible experts. The token is then sent to the expert that was given the highest weight. The network is implemented such that different experts can actually live on different devices. 

https://i.imgur.com/HEB7cJw.png

This architecture is inspired by previous Mixture of Experts work, which applied a similar scheme, but sent each token through a set of k experts rather than just a single one. This had the ostensible effect of increasing stability and performance, but the authors of this paper argue that using a single expert per token is actually preferable on both of these fronts. 

There are a lot of experiments in this paper, and I'd recommend taking a look at them in detail if you're interested, but, at a high level, they found evidence that, compared to models with a comparable amount of parameters they were indeed able to get comparable or better performance with a lower number of FLOPS. It also meant they were able to build up to a trillion-parameter model, without having unreasonable computation requirements. 

Some interesting considerations relevant to this approach: 

- To keep training speed up, you need to strike the right balance of the number of tokens sent to each expert; in this case, the authors added a loss term to incentivize the division between experts to be roughly uniform
- There was some numerical instability around the expert training procedure if you used float16 data types, so they switched to using float32, but only within the experts themselves, rather than in the rest of the network.
- To regularize this huge of a network, the authors decided to apply dropout, but only within the experts

When machine learning models need to run on personal devices, that implies a very particular set of constraints: models need to be fairly small and low-latency when run on a limited-compute device, without much loss in accuracy. A number of human-designed architectures have been engineered to try to solve for these constraints (depthwise convolutions, inverted residual bottlenecks), but this paper's goal is to use Neural Architecture Search (NAS) to explicitly optimize the architecture against latency and accuracy, to hopefully find a good trade-off curve between the two. 

This paper isn't the first time NAS has been applied on the problem of mobile-optimized networks, but a few choices are specific to this paper. 

1. Instead of just optimizing against accuracy, or optimizing against accuracy with a sharp latency requirement, the authors here construct a weighted loss that includes both accuracy and latency, so that NAS can explore the space of different trade-off points, rather than only those below a sharp threshold.  
2. They design a search space where individual sections or "blocks" of the network can be configured separately, with the hope being that this flexibility helps NAS trade off complexity more strongly in the early parts of the network, where, at a higher spatial resolution, it implies greater computation cost and latency, without necessary dropping that complexity later in the network, where it might be lower-cost. Blocks here are specified by the type of convolution op, kernel size, squeeze-and-excitation ratio, use of a skip op, output filter size, and the number of times an identical layer of this construction will be repeated to constitute a block. 

Mechanically, models are specified as discrete strings of tokens (a block is made up of tokens indicating its choices along these design axes, and a model is made up of multiple blocks). These are represented in a RL framework, where a RNN model sequentially selects tokens as "actions" until it gets to a full model specification . This is repeated multiple times to get a batch of models, which here functions analogously to a RL episode. These models are then each trained for only five epochs (it's desirable to use a full-scale model for accurate latency measures, but impractical to run its full course of training). After that point, accuracy is calculated, and latency determined by running the model on an actual Pixel phone CPU. These two measures are weighted together to get a reward, which is used to train the RNN model-selection model using PPO. 

https://i.imgur.com/dccjaqx.png

Across a few benchmarks, the authors show that models found with MNasNet optimization are able to reach parts of the accuracy/latency trade-off curve that prior techniques had not.