CLIP and other recent works tackle problems of robustness never thought possible before, pushing forward the frontiers of generalisation under distribution shift, with the hopes of achieving better performance at deployment. These models rely on weakly-supervised learning with large diverse datasets scraped from the web, combining images and language in an interesting way. I was a bit puzzled by the underlying mechanism that led to such powerful performance, thinking that these truly look like magic. But understanding the mechanism underlying an interesting model is the first step towards finding faults with it (and research opportunities), so I spent some time organising my thoughts on the matter. I found myself explaining this 4 times over the past week, so might as well write a blog post for others to benefit as well.
Let’s start with invariances. Say you have an image of a dog, and the dog is in the bottom left. If the dog were in the bottom right, you would still classify the image as ‘dog’. There are different ways to build such an invariance into a model. You can build your model to be translation invariant by using convolutions – layers which don’t change the representation if you shift the input left or right (with some caveats not relevant to this discussion). This is a ‘hard’ invariance – you encode the symmetry into your model. But you can also build your model to be translation invariant by using a ‘soft’ invariance. Define an orbit to be a sequence of images, where each image is the original dog image shifted by a bit (i.e., on each image you repeatedly apply the action of the group to which you want to be invariant). Evaluate your model on each such image in the sequence, and collect all the feature vectors (e.g., the representations from the penultimate layer of a neural network). Define a new model output to be the average of all these feature vectors, fed into a softmax layer which you use with a cross-entropy loss. This new model, let’s call it the ‘orbit model’, is invariant to translations: if you shift or move the dog in the original image, the orbit would stay the same, and thus the orbit model output would not change – the model is invariant to translations.
But building such models defined over orbits can be expensive (or downright impractical for infinite orbits) so instead you can just uniformly sample a bunch of shifted images from the orbit, and average only a finite number of the corresponding feature vectors (or even just use one). For certain losses, such as the cross entropy loss, the objective you’d be minimising will simply be an upper bound to the ‘correct’ objective evaluated over the entire orbit (again, with some caveats). So optimising your model parameters to minimise this biased loss is guaranteed to minimise your orbit model loss as well (you can read more about hard invariances and soft invariances—such as feature averaging and data augmentation—here).
You can actually interpret data augmentation as building these sort of soft invariances into your model: augmenting your dataset with random translations is doing exactly this. Similarly, curating huge datasets with lots of images of dogs sitting in different locations, all consistently labelled ‘dog’, will enforce a soft invariance in your model indicating that the dog’s location is irrelevant to the classification task. Here we still use normal supervised learning and we curate a large dataset where we define the labels to be invariant to the different locations of the dog (i.e., when you ask people to label your dataset, you tell them ‘classify this as dog if there’s a dog anywhere in the image’).
Note though that such soft invariances learnt through the loss don’t actually guarantee that new dog locations never-before observed at training time should map to the same embedding. In fact, our model is not really translation invariant, but something more subtle. If our dataset only had dogs in the bottom half of the image (on the ground) and never in top half (in the sky), and we had lots and lots of such images with dogs at every possible pixel shift in the bottom half of the image, the model would learn a very specific invariance which is ‘dogs are translation invariant in the bottom half of the image’ (in a very hand-wavy way). When you think about this, this invariance makes more sense than global translation invariance. Labelling consistency with a large enough dataset is defining the invariance.
What does this have to do with CLIP? Well, you can think of the huge dataset of images and captions scraped from e.g. reddit as enforcing some sort of soft invariances. But unlike the explicit soft invariance enforced through data augmentation or labelling of large datasets, CLIP’s invariances are enforced in a fairly indirect way. I will assume you already know how CLIP works (and if not, read here, but briefly, you simply use two encoders – one encoding images and the other encoding ‘captions’ scraped from the web, and you want images to encode into embeddings which are near-by to their caption embeddings, but far away from the caption embeddings of other images).
What does CLIP have to do with soft invariances? in simple terms, if you have a picture of a dog, and you have a caption saying ‘I took a picture of a dog’, then CLIP will try to embed the picture near the caption embedding. Now, say you have another picture of a dog, but this time instead of the dog sitting in the bottom-left of the image, it is sitting in the bottom-right. Your encoder might give you a different image embedding for this new image. But if the caption scraped from the web still says ‘I took a picture of a dog’, then minimising the loss will get the model to do something quite interesting: The encoder will try and push the embedding of the new picture (of a dog sitting at the bottom-right) to lie in embedding space next to the embedding of the caption from before (‘I took a picture of a dog’), and by transitivity, be near the embedding of our first picture (of the bottom-left dog).
In fact, any new picture of a dog with the caption ‘I took a picture of a dog’ should be mapped to the same point in embedding space (following our contrastive loss), and we get something very similar to what we had in the above soft invariances case – all images of dogs in our training set, as long as they have the same caption (or close enough caption in embedding space), will be mapped to the same embedding – i.e. the model learnt a soft invariance through the loss. If we have enough data, like with OpenAI CLIP or Google JFT, we probably have dogs in every imaginable location, all of which being mapped – under our objective – to the same embedding. Building a classifier on this embedding space will map all such embedded dogs at different locations to the same class ‘dog’, and will thus be invariant according to our definition above.
The interesting thing about CLIP is that we didn’t have to define the invariance by hand: neither through model architecture, nor by constructing transformations for an orbit/data augmentation, or even by deciding on a labelling scheme ‘dog class should be the same regardless of the dog location’. This invariance is defined by language. So after data, language (and the way we curate our captions or prompts) becomes the most important thing. That’s because our captions define the invariances in our model, hence which inputs should lie nearby other inputs in embedding space. The model can classify abstract concepts like sketches of bananas if the language specifies that bananas are bananas regardless of their medium (a more fine view of this is that the embedding of ‘I drew a sketch of a banana’ lies near the embedding of ‘I took a picture of a banana’).
But this is where things get complicated as well. What if the language description of an image, which might concentrate on its salient objects, ignores other objects in the image? For example imagine scraping a web picture from a dashcam sitting at the front of a car, looking at the road ahead. The caption says ‘I took a picture of a black car in front of me’. There are lots of other objects in the picture as well though: There are clouds in the sky; There is a pedestrian about to cross the road; There is a red traffic light which caused the driver to stop.
If you can see where I’m heading with this, then you very much wouldn’t want to use a CLIP-like model in your autonomous car. Let’s unpack what will happen if you do: lots of dashcam pictures in our dataset will have a black car in front of them as the main object in the scene. Many of these will have the caption ‘a black car in front of me’. Out of these, some will have pedestrians about to cross the road at the corner of the frame, while others will not. Some will have a green traffic light at the top of the frame, while others will have a red traffic light. But following the discussion above, all will be mapped to the same embedding. I.e., the model is taught soft invariances which say “the representation should be invariant to pedestrians about to cross the road”, “the representation should be invariant to the colour of the traffic light”, among many other implicit invariances. Our model is only as good as our embeddings, and our embeddings ‘throw away’ all information not encoded in the caption (since the embedding will be the same regardless of the red or green traffic light). So if you try to use such a model for red/green traffic light classification (which will be used to slow down the car in front of a red light), you will most likely get very bad performance (and in fact we tried that with CLIP, and got very bad performance).
This is basically a problem of compositionality, where images may contain many objects, each of which might be relevant to a different application you might care about (pedestrian detection, car detection, traffic light classification, cloud forecasting, etc). When we use large pre-trained models, we don’t have control over the implicit invariances of the web-scraped captions, and all we can do is hope that they capture what we care about for the task at hand. This is a severe limitation of this class of models as long as they are used as fixed feature-extractors.
But this can also be seen as an opportunity. Can we design such models and losses that can be used in compositional settings? can we design such models and losses that can be tuned post-hoc to recover lost information and to remove unwanted invariances? these questions are really interesting, and might lie at the heart of upcoming advances in machine learning. But only a few can actually work on these question given the closed-source data and immense compute required to retrain such models. Perhaps a much more pertinent question to consider is democratising such research, building open-source datasets and compute efficient tools to allow the wider community to explore these new avenues.
If you use any of the ideas above, please cite appropriate papers which came up with the ideas. If you’re using ideas proposed here, please cite
Yarin Gal, "The mechanism underlying OpenAI's CLIP, and limitations", Technical report, 2021.
The above dashcam example is a bit of a simplistic view of what would happen in a compositional setting. In practice, for large enough data and diverse enough captions, you might have very similar dashcam pictures with some captions more descriptive than others (‘there’s a black car and a red traffic light far ahead, oh and also a pedestrian about to cross the road’). You might also have the same captions describing very different images (as in the figure above). The model will then have to minimise a loss that balances the tension between all these different captions and image combinations, arriving at a representation that is more complicated than my explanation above. Again, it all boils down to language, and the soft invariances encoded by it.
The above thoughts are written informally for ease of accessibility, and to reach a wide audience. These can be trivially formalised as a theory to explain some aspects of CLIP, from which we can generate testable hypotheses to try to support or falsify this theory - giving us many more questions for future research. You can read more about this approach to machine learnign here.
Many thanks to Jan Brauner for comments on an early draft of this writeup.