Why Multimodal Machine Learning models do not work. Part 2/2 – The CAUSES

Key Takeaways

[Music] disclaimer this video highlights the not so favorable aspects of integrating vision and language for information on risk and side effects please read your favorite publication and ask your machine learning researcher or engineer hey there did you watch miss coffee beans previous video if not check it out right now and then come back to this video last time she told us about the symptoms that are a clear hint that something fishy is going on when integrating vision and language with deep learning methods it seems that one modality is usually almost forgotten even though the tasks are defined to require the understanding of both modalities to deliver the right answer in this video miss coffeebean is going to tell us what she found further on her journey through the dark paper forest miss coffee bean what are the reasons for problems in vision and language fusion on my journey i encountered two sources of all evil some say the source of problems when integrating vision and language is in the data while others say it is in the neural network based models let's start with the accusations against the data the vqa data set was released in 2015 and a lot of work was invested by researchers to get on top of the challenge leaderboard then 2017 a version 2 of the vqa dataset had to be released since obvious biases in the data were detected these biases were making the undesired happen the model could focus only on the text ignoring the image it's in the title making the v in vqa matter for example far too numerous how many questions have the answer to if the question is about sport then usually the correct answer is tennis also questions like is there a special something in the image bear the answer yes since humans contributing for the data set collection usually do not ask if there is a horse in the image if there is none to be seen how would they even get the idea of a horse in the first place so these obvious biases were removed but there are suspicions that neural networks can also rely on not so obvious biases and many models still seem to ignore the image far too much what holds for vqa usually holds for visual dialogue too as visual dialogue is simply put vqa with history this paper shows a very simple statistical method not even neural networks without accessing the image or the sequence of the dialogue and this method performs on par on some metrics with extremely complex neural models of course authors of the original visual dialogue papers did not take the critique well if you want to read how researchers fight check the links to these papers and the response in the description below but whatever side you are on i think that one must take this alarm seriously it is always tricky to produce clean training data because the fingerprints of the data collectors are everywhere on it and neural networks are extremely good at exploiting every hint and bias to get to the right answer even if for the wrong reasons but apart from neural networks exploiting bias in the data it seems to be something fundamentally wrong with neural networks trying to fuse modalities so it is time for us to discuss the problems that models might inherently have we think this question is still very very poorly investigated but we have one research product to show you this paper conducts a very simple yet fundamental experiment given an image that can be described by only one word a neural network has to construct a word that is describing the image not the classification task where the network has to choose the right index of the right class to which we humans have assigned the labels it is rather a regression task where the model has to come up with the right word embedding that best fits to the picture simply put they try to translate images to single words rather than a whole sentence like in image captioning what do we expect it should be an easy task right image captioning is a much harder task and neural networks can do it so why should it not be possible to translate an image to a single word and you're right it is possible the neural network performs with an impressive accuracy only that the paper looks a little bit closer it also measures how far the neural network can transform the input space to make it look like the output space and here the results are surprising it turns out that the neural network is really bad at that it stays much closer to the input modality which is the image even though the network is trained to fully transform the image space to the textual word embedding space it preserves the neighborhoods of the input space rather than construct the output space why because neural nets are guaranteeing continuity and preserved topology we think that there is still very much to investigate along this line especially the case when both visual and textual modalities are mapped by neural network into a common space because this was only a special case where the image only was transformed but even this special case analysis is a strong alarm signal when talking about neural networks in multimodal research as they are universal approximators given infinite training time and data we expect them to be the solution to all our multimodal problems but perhaps we expect too much from them in our real setting with not so much training data or training time like in the example seen before the neural network was able to transform single data points from the image to the textual modality but was not able to sensibly transform the space in between with infinitely many training samples this would not have been a problem since the network would have had a training sample for every point in the space but even if we would have these infinite resources we have seen how problematic data and biases can be when integrating vision and language and that we can insert multi-modal fusion difficulties when choosing a model or architecture we could end our video here since this is what's miss coffee bean found on her search for hints and reasons on why current vision and language models fail where they fail but we want to spend a minute on multimodal models that are coming out right now beating the previous state of the art but have not been yet more thoroughly investigated for problems yes we are talking about the multi-modal transformers which are all variations of the same theme of processing images and text and combining them with transformer modules and cross-modal attention for more details on how this architecture works check out our previous video on wilbert link below multimodal transformers show through their better accuracy on multi-modal tasks and attention visualization that they are better at combining the visual and textual modalities but still a recent publication suggests that for multimodal transformers it seems again that quote the textual modality plays a more important role than image in making final decisions so here we are again at the same problem that existed even before multiple transformers the inequality of modalities and the curse of biased data sets towards the textual modality so what do we do now more research the relevant papers for this video are in the description below if you are interested to investigate the root of the problem yourself this video is of course not an exhaustive literature enumeration of all problems in multimodal integration it's rather just a hint of where to look when searching for trouble i mean in an academic sense anyway what do you think let us know in the comments what your thoughts and observations are see you next time okay bye [Music] [Music] you