Recently I picked up Oliver Sacks' The Man Who Mistook His Wife for a Hat for the second time. The first chapter, from which the book gets it's name, details the case of Dr P. The story of Dr P is the story of a man with visual agnosia. This means that although he can see and understand basic shapes and structures he finds it difficult to see things as wholes. Dr P can see noses, eyes, ears, and mouths but he cannot recognises the faces of those he knows (a phenomenon called prosopagnosia when related specifically to faces). In the book Sacks describes seeing Dr P mistake his foot for the shoe he had taken off, and amusingly mistake his wife's head for the hat he put on the rack when he came in.
When I got to the end of this chapter I found myself pausing to think about image recognition and what I knew about the way that artificial neural networks are being used to process complex scenes and objects. I quickly started to draw some parallels between Sacks' description of agnosia and phenomena I had read describing the workings of deep neural nets. This sparked an idea for an extraordinarily simple hypothesis for the root cause of agnosia in people. An idea we could use to build a theoretical model of agnosia.
This article is my way of writing down my thoughts. Hopefully I will find time to test this hypothesis myself, or it will inspire someone else to test it. I write this with the intention of making sure that my thoughts are clear, and that they make some degree of sense. It is also an opportunity for me to fact check some of my assumptions, and maybe even get others to check them over. I will begin with where my thoughts started, with neural networks.
In 1957 Frank Rosenblatt created the first artificial neural network, named the Mark 1 Perceptron. It was designed for image recognition and consisted of only a single intermediate layer of 512 units (or neurons). The Mark 1 Perceptron was capable of differentiating between 8 basic abstract 2D shapes, such as squares, circles, triangles, stars, etc. Although it took a few years, multilayer neural networks with many more neurons started being made that were capable of more robust image recognition. A significant barrier was overcome when neural nets were made that had greater than 8 layers, which we now refer to as deep neural networks. Enormous modern deep neural networks are quite capable of complex scene and object recognition, for example being able to tell that an image is of a cowboy riding a horse rather than just saying that it is a person.
At a glance it seems as though the larger the number of layers and neurons (the depth and breadth of a neural network) the more complex the concepts they are able to describe. Smaller networks are only capable of recognising more abstract concepts. At this point my thought pattern should be obvious. Does a generalised loss of functioning neurons in an area of the brain associated with visual processing lead to a loss of ability to process complex models? This sounds like a massive oversimplification, but perhaps it is not. Because the nervous system is plastic, functions lost through the cutting of certain pathways or the dying of small numbers of neurons can be learned somewhere else. But, tumours that take up significant space, or permanent loss of perfusion to those areas limit the depth and breadth of the nervous network there.
In situations in which a significant number of the neurons in that area have been permanently removed plasticity may still play a part. As that area re-learns to interpret visual information it only has the resources to learn to interpret the more abstract concepts. It may be possible to create a computer model to explore this possibility. By creating a convolutional neural network that is able to classify both complex and simple abstract concepts, then progressively retraining it with fewer and fewer neurons, we may be able to plot the accuracy of both complex and more abstract assertions. If this hypothesis is correct then the accuracy of classifying similarly complex objects should start to fall as the number of neurons do, whereas the accuracy of classifying simpler objects should hold steady until the number of neurons falls enough to start affecting them.
If this computer model fits the hypothesis it would be time to start asking whether we can observer this behavior in sufferers of agnosia. The question then simply becomes, does the severity of their agnosia scale appropriately with the reduction of functioning neurons? I know very little about our capacity to count neurons in human subjects, so this may be infinitely more difficult than it sounds. If this hypothesis appeared to be true on both sides we could start asking the question, what can we learn from the computer model? Is breadth or depth of a network more important? Is the structure of inter-neuronal connections not important, only the volume of them? It could also tell us something about thresholds for accuracy. How many more neurons would our patient need to get some sense of complex objects back? Answers to these questions could also re-affirm the importance of having machine learning algorithms that are based on our understanding of human anatomy.
At this stage, I don't believe I have thought this through in the level of detail it deserves. I am also not familiar enough with the literature to know how plausible this sounds, or even how important. Perhaps this idea is trivial and generally well understood. I have recently been trying to implement a variety of machine learning algorithms from scratch, and need to finish writing up an article on my implementation of an anomaly detection system. After that is finished, it seems like it might be a good time to start writing a neural net. Now I have something that feels important to write it for.
If you work in neurology, neuropsychology, neurophysiology, machine learning, or with neural networks I would like to hear your opinion. My relationship with academia is as a hobbyist, and so I have no formal contacts in specialist fields to bounce ideas off. I hope you've found this interesting.