What can artificial neural networks teach us about connectomics?

When children learn to read, they are taught to “sound out” words, or read aloud letter-by-letter. In the 1980s, Terrence Sejnowski and Charles Rosenberg sought to model this process by building a neural network called NETtalk that learned to convert written text to speech1. In other words, they taught a computer2 to read. The authors concluded that NETtalk was too simple to serve as a complete model for human learning, but it does have some important implications for connectomics. Connectomics is the study of wiring-diagrams of the nervous system, or the patterns of connections between neurons. The hope is that by mapping the connectome, we will gain fundamental insights into the function, and dysfunction, of the brain in health and disease. A quick summary of the inner-workings of NETtalk is necessary before we can understand the implications for connectomics.

NETtalk is a simple neural network that consists of 3 layers of nodes: an input layer, a hidden layer, and an output layer (Figure 1). To train the network, written text is fed into the input layer, is then propagated through the hidden layer, and is finally mapped onto a phoneme in the output layer. The letter-to-phoneme correspondence is determined by the connections between nodes, and the network learns the correct correspondence by readjusting the weights of connections between layers. A teacher unit provides feedback, and the network uses this feedback through many iterations of training to adjust its connections.

nettalk
Figure 1: Schematic of NETtalk network architecture from Sejnowski and Rosenberg (1987).

Similar findings have been observed in biological systems. Jeff Lichtman’s group has shown that the wiring pattern at the neuromuscular junction of a small muscle in the ear of a mouse varies widely across animals, and can even vary within the same animal3 (Figure 2).  Despite the same genes and environment, populations of neurons in the left and right ears of a mouse are wired up differently. Knowing the connectome of this muscle, therefore, does not help us understand the its function.

muscle
Figure 2: Connectomes of a muscle in the left and right ears of a mouse. From Lu et al., 2009.

As the field of connectomics continues to grow, it is important to reflect on what a neural wiring-diagram can tell us about brain function. The Human Connectome Project, a $30+ million project funded by the NIH, is currently mapping structural and functional connectivity in humans using multimodal neuroimaging. Is this a good use of finite scientific resources and tax payers’ money? Will a complete connectome bring us any closer to figuring out the brain? These questions are currently the subject of much controversy (see this video for a lively debate between Anthony Movshon and Moritz Helmstaeder at the 2016 Cognitive Neuroscience Society meeting). My thinking is that a wiring-diagram is useful if it is acquired at the right scale. But what is the correct scale? We don’t know. A microscale connectome that maps individual synapses between individual neurons is almost certainly too detailed a map, as the previously noted examples from artificial1 and biological3 systems have demonstrated. Conversely, a macroscale connectome, the scale at which the Human Connectome Project is being conducted, probably isn’t detailed enough. Therefore, a mesoscale description is required. Unfortunately, neuroscience has historically been constrained to micro and macro scale brain mapping by available technology. In response, a core goal of the White House BRAIN initiative is the development of new tools for mapping neuronal connections. These new technologies for acquiring wiring-diagrams at multiple resolutions will be essential for determining what, if anything, the connectome tells us about brain function.

References/Notes

  1. Sejnowski TJ, Rosenberg CR (1987) Parallel networks that learn to pronounce English text. Complex Syst 1:145–168.
  2. The computer used to run NETtalk was a VAX 11/780 from Digital Equipment Corporation. The VAX 11/780 was big and expensive: it was over 5 feet tall, 4 feet long, and cost over $150,000. It was slow too; the CPU clocked in at 5 MHz. The iPhone 7, with a 2.3 GHz quad-core processor, has over half a million times more processing power per dollar than the VAX 11. And it’s 216,000 times smaller.
  3. Lu J, Tapia JC, White OL, Lichtman JW (2009) The interscutularis muscle connectome. PLoS Biol 7:e32.
Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s