Self-organizing maps are an old idea (first published in 1989) and take strong inspiration from some empirical neurophysiological observations from that time. The original paper “Self-Organizing Semantic Maps” by Ritter and Kohonen (pdf) has a nice discussion that took me back to some questions I was looking at in another life as a neurophysiologist. The discussion centers around cortical maps, which are most pronounced in the clearly sensory and motor areas of the brain.
Very early experiments (in their paper Lashley‘s work is referenced) led some people to propose the brain was a complete mishmash, with no functional separable internal organization. Each part of the brain did exactly what the other parts did, a bit like how a lens cut in half still works like the original lens (just dimmer), or how a hologram cut in half, still shows the original scene.
Later experiments looked more closely and found a wealth of specialization at all scales in the brain. The cortex (the most evolutionarily recent part of the brain) shows an especially fascinating organization at the smallest scales. This is most apparent in the visual cortex, where the neurons are often organized in two or higher dimensional maps. A famous example is the set of orientation maps in early visual cortex:
This figure is taken from Moser et al, Nature Reviews Neuroscience 15,466–481(2014) but you’ll run into such figures every where. It shows physical space along the cortex, color coded for the orientation of bar that the neurons at that location best respond to. One can see that the colors are not randomly assigned but form contiguous regions, indicating that neighboring neurons respond to similar orientations.
I could go on about cortical organization like this (and indeed I have elsewhere) but this is a good segway into self-organizing maps. Researchers studying machine learning got interested in this feature of the brain and wondered if it held some use for classification problems. Kohonen drew up an algorithm for what he called a self-organizing map that has been used on and off mainly for visualization via un-supervised learning.
The Kohonen map is constructed as follows.
- Construct a sheet of “neurons” with the dimensionality you want. Usually this is limited to one, two or three dimensions, because we want to visualize a complex higher dimensional dataset.
- Each neuron is a template – it has the same dimensions as the input data and can be overlaid on each instance of the input data.
- We initialize the templates randomly – so the templates look like white noise.
- We “present” each input example to the map. Then we find the template neuron that most closely matches this input. We then morph the template slightly so that it even more closely matches the input. We do this for all the neighbors of the neuron too.
- We keep repeating this for all the input we have.
What does this result in? I have two videos for you, you product of the MTV generation, that shows how this works.
I took the famous NIST handwritten digits data set and fed it to a two dimensional SOM.
The first video represents the self-organizing map as a set of templates that change as inputs are supplied to it.
This is easy to interpret as you can identify the templates as they evolve. Note how neurons form neighborhoods that naturally link together similar stimuli. It is easy to see with this data set because the data are two dimensional visual patterns and we can judge similarity intuitively.
The second video represents each neuron’s stimulus preferences, with it’s most preferred stimulus digit appearing as the largest.
You can see of the network evolves as examples are thrown at it and segregates out into zones that respond to or represent a particular form of the stimulus. You can see how zones travel and try to spread-out, ending up at the corners of the map. This is due to the competition created by the winner-take-all operation we use (closest matching template wins) combined with the effect of adjusting the winning neuron’s neighbors too.
SOMs are fun things to play with, but are a little fiddly. You may have picked up from the videos that we had two parameters called sigma and eta that changed as the learning went on. The numbers got smaller as the training went on. Sigma is the size of the neighborhood and eta is the learning rate – the smaller the value of eta the less the neurons change in response to inputs. These are two things you have to fiddle with to get decent results. Also as you can imagine, this whole process is sensitive to the random templates we start with.
Numerous attempts have been made to try and figure out objective ways to determine these numerous parameters for a SOM. It turns out they are pretty data dependent and most people just do multiple runs until they get results they are happy with.
As I was reading Bishop I ran into his version of the SOM which he calls Generative Topographic Mapping which Bishop claims is more principled and mathematically grounded than SOMs. The GTM, which I really haven’t heard about at all, seems like a fun thing to study, but as a ex-neurophysiologist SOMs are definitely more intuitive to understand.
Code (and link to data source) follows: