This article won the , and was also the of August 2017 KDNuggets Silver award most viral post Every time DeepMind publishes a new paper, there is frenzied media coverage around it. Often you will read phrases that are often misleading. For example, its new paper on relational reasoning networks has reporting it like futurism DeepMind Develops a Neural Network That Can Make Sense of Objects Around It. This is not only misleading, but it also makes the everyday non PhD person intimidated. In this post I will go through the paper in an attempt to explain this new architecture in simple terms. You can find the original paper . here This article assumes some basic knowledge about neural networks . How this article is structured I will follow the paper’s structure as much as possible. I will add my own bits to simply the material. What is Relational Reasoning? In its simplest form, Relational Reasoning is to understand relations between different objects(ideas). This is considered an essential characteristic of intelligence. The authors have included a helpful infographic to explain what it is learning Figure1.0 The model has to look at objects of different shape/size/color, and be able to answer questions that are related between multiple such objects. Relational Networks The authors have presented a neural that is made to inherently capture relations(e.g. Convolutional Neural networks are made to capture properties of images). They presented an architecture that is defined like so : network Equation1.0 Definition of Relational Networks Explained The Relational Network for O ( ) is a function O is the set of objects you want to learn relations of fɸ. is another function that takes two objects :o_i_ , and o_j_. The output of is the ‘relation’ that we are concerned about. gθ gθ Σ i,j means , calculate gθ for all possible pairs of objects, and then sum them up. Neural Networks and Functions It is easy to forget this when learning about neural networks, backprop ,etc. but a neural network is in fact a ! Therefore, the function that I described in Equation 1.0 is a neural network!. More precisely , there are two neural networks: single mathematical function which calculates relations between a pair of objects gθ, which takes in the sum of all and calculates the final output of the model fɸ, gθ, Both , and are multi layer perceptrons in the simplest case. gθ fɸ Relational Neural Networks are flexible The authors present Relational Neural Network as a module. It can accept encoded objects and learn relations from them, but more importantly, they can be plugged into Convolutional Neural networks , and Long Short Term Memory Networks (LSTM). The Convolutional network can be used to learn the objects using images. This makes it far more useful for applications because reasoning on an image is more useful than reasoning on an array of user defined objects. The LSTMs along with word embeddings can be used to understand the meaning of the query that the model has been asked. This is again , more useful because the model can now accept an English sentence instead of encoded arrays. The authors have presented a way to combine relational networks, convolutional networks , and LSTMs to construct an end to end neural network that can learn relations between objects. Figure 2.0 An end to end relational reasoning neural network. Figure 2.0 Explanation The image is passed through a standard Convolutional Neural network(CNN), which can extract features of that image in filters. The ‘object’ for the relational network is a vector of features of each point in the grid. e.g. one ‘object’ is the yellow vector. k The question is passed through an LSTM , which produces a feature vector of that question. This is roughly the ‘idea’ of that question. This modifies the original Equation 1.0 slightly. It adds another term which makes it Equation1.0 Relational Network conditioned using LSTM Notice the extra in Equation 1.0. That is the final state of the LSTM. The relations are now using . q q conditioned q After that, the ‘object’ from the CNN and the vector from the LSTM are used to train the relational network. Each object pair is taken, along with the question vector from the LSTM, and those are used as inputs for gθ( ). which is a neural network The outputs of gθ are then summed up , and used as inputs to fɸ(which is another neural network). fɸ is then optimsed on the answer to the question. Benchmarks The authors demonstrate the effectiveness of this model on several datasets. I will go through one of them (and in my opinion the most notable) — CLEVR dataset. The CLEVR dataset consists of images of objects of different shapes,sizes and color. The model is asked questions about these images like: Is the cube the same material as the cylinder? The types of objects(top),and the positioning scheme (centre&bottom) Figure 3.0 The authors point out that other systems are far behind their own model in terms of accuracy. This is because Relational networks are designed to capture relations. Their model achieves an unprecedented 96% + accuracy, as compared to a mere 75% (using stacked attention models) Figure3.1 Comparison between different architectures on the CLEVR dataset using pixels(i.e. not matrix encoded) Conclusion Relational Networks are extremely adept at learning relations. They do so in a data efficient manner. They are also flexible and can be used as a drop in solution when using CNN’s, LSTMs, or both. This post was about debunking the ‘AI has taken over’ hype caused by very large publications, and giving some perspective on what the current state of the art is. P.S. If you notice any errors, or would like any modifications, please let me know through responses. Your suggestions are welcome. If you liked the article, please recommend it to others by tapping the ❤ button.
