MIT’s new neural network can explain its decisions

A child is presented with a picture of various shapes and is asked to find the big red circle. To come to the answer, she goes through a few steps of reasoning: first, find all the big things, next, find the big things that are red, and finally, pick out the big red thing that’s a circle – it’s something that sounds simple enough, unless, of course you’re a machine.

As humans we use reasoning to learn how to interpret the world which is why Google’s bleeding edge Artificial Intelligence (AI) division is focusing so much on it, if only for the fact that unless machines can learn to reason we’ll never be able to create a commercial grade Artificial General Intelligence (AGI), the world’s first groundbreaking prototype of which, known as Impala, was unveiled by same said company a couple of months ago in London.

Neural networks also learn using reason, although the way these so called narrow AI’s that we have today, that can only do one thing well rather than lots of things well like future AGI’s, use it in a different way to the way an advanced AGI eventually will. This said though one of the biggest issues we face today is that while we know these neural networks use reason to learn, we don’t know how they reach their final decisions. To all intents and purposes they’re black boxes and that’s an issue if they’re running self-driving cars, the stock market, or making a decision on which military targets to bomb, something that DARPA, the US military’s bleeding edge research arm is also trying to solve.

Now a team of researchers from MIT have developed a neural network that “performs human-like reasoning steps to answer questions about the contents of images.” Called the Transparency by Design Network (TbD-net), the model visually renders its thought process as it solves each part of the problem, allowing human analysts to understand its decision making process. And the model performs better than today’s best visual reasoning neural networks including Nvidia’s own attempt at reading the minds of these AI’s.

Understanding how a neural network comes to its decisions has been a long-standing challenge for AI researchers who’ve even recently turned to cell biology to help them try to solve the problem. As the neural part of their name suggests, neural networks are brain inspired AI systems intended to replicate the way that humans learn. They consist of input and output layers, and layers in between that transform the input into the correct output. Some deep neural networks have grown so complex that it’s practically impossible to follow this transformation process. That's why they are referred to as black box systems, with their exact goings on inside opaque even to the engineers who built them. With TbD-net the developers aim to make these inner workings transparent, literally, so that humans can interpret the AI's results.

It is important to know, for example, what exactly a neural network used in self-driving cars thinks the difference is between a pedestrian and stop sign, and at what point along its chain of reasoning does it see that difference. These insights allow researchers to teach the neural network to correct any incorrect assumptions. But the TbD-net developers say the best neural networks today lack an effective mechanism for enabling humans to understand their reasoning process.

"Progress on improving performance in visual reasoning has come at the cost of interpretability,” says Ryan Soklaski, who built TbD-net with fellow researchers Arjun Majumdar, David Mascharka, and Philip Tran.

The team was able to close the gap between performance and interpretability with TbD-net. One key to their system is a collection of "modules," small neural networks that are specialized to perform specific subtasks. When TbD-net is asked a visual reasoning question about an image, it breaks down the question into subtasks and assigns the appropriate module to fulfil its part. Like workers down an assembly line, each module builds off what the module before it has figured out to eventually produce the final, correct answer. As a whole, TbD-net utilizes one AI technique that interprets human language questions and breaks those sentences into subtasks, followed by multiple computer vision AI techniques that interpret the imagery.

"Breaking a complex chain of reasoning into a series of smaller subproblems, each of which can be solved independently and composed, is a powerful and intuitive means for reasoning," says Majumdar.

Each module's output is depicted visually in what the group calls an "attention mask." The attention mask shows heat-map blobs over objects in the image that the module is identifying as its answer. These visualizations let the human analyst see how a module is interpreting the image.

Take, for example, the following question posed to TbD-net: “In this image, what colour is the large metal cube?" To answer the question, the first module locates large objects only, producing an attention mask with those large objects highlighted. The next module takes this output and finds which of those objects identified as large by the previous module are also metal. That module's output is sent to the next module, which identifies which of those large, metal objects is also a cube. At last, this output is sent to a module that can determine the colour of objects. TbD-net’s final output is “red,” the correct answer to the question.

The researchers evaluated the model using a visual question-answering dataset consisting of 70,000 training images and 700,000 questions, along with test and validation sets of 15,000 images and 150,000 questions. The initial model achieved 98.7 percent test accuracy on the dataset, which, according to the researchers, far outperforms other neural module network–based approaches.

Importantly, the researchers were able to then improve these results because of their model's key advantage - transparency. By looking at the attention masks produced by the modules, they could see where things went wrong and refine the model. The end result was a state-of-the-art performance of 99.1 percent accuracy.

"Our model provides straightforward, interpretable outputs at every stage of the visual reasoning process,” Mascharka says.

Interpretability is especially valuable if deep learning algorithms are to be deployed alongside humans to help tackle complex real-world tasks. To build trust in these systems, users will need the ability to inspect the reasoning process so that they can understand why and how a model could make wrong predictions.

Paul Metzger, leader of MIT’s Intelligence and Decision Technologies Group, says the research “is part of MIT’s work toward becoming a world leader in applied machine learning research and artificial intelligence that fosters human-machine collaboration.”

The details of this work are described in the paper, “Transparency by Design: Closing the Gap Between Performance and Interpretability in Visual Reasoning," which was recently presented at the Conference on Computer Vision and Pattern Recognition (CVPR).