Philosophy 371: Minds & Brains/Cognitive Science Lab 1997
Lab # 8: Neural Network Computation at the Outskirts of the Brain
Preamble:
Computation has traditionally been equated with the operations of the kind of computer conceived by Turing, made real by Von Neumann, and planted on your desk by Apple or IBM. In our widgets lab, we eased into a new way of thinking about computation, inaugurating a long excursion in "net thinking." Net thinking sets aside the idea of programmed machine operating on data stored in memory registers, and instead invites us to think about the complex interactions of simple processing "units" -- each one somewhat like the biological neurons that compose our brain and nervous system. Over the next weeks we will consider several dimensions of this new way of looking at computation in the brain and mind.
Top-down and bottom-up:
Here’s a lab to think about: build a wing. Let the wing be suitable for keeping a two pound weight aloft in a glide for ten seconds. There are two broad approaches to the problem: 1) You might begin by thinking about the task, about the lift required (two pounds) and the surface area a wing might need to generate that lift, the shape it should have, and other aerodynamic themes. After you figure out how the wing needs to work, you might then turn to its materials -- you’ll need something strong and light, etc. Or: 2) You might begin by asking how the problem has already been solved by nature. Consider a bird’s wing. You might start with the feathers -- how do they work? How do they interact in flight? How are they mounted? As you understand each of the parts of a working wing, you may be able to duplicate it, and slowly work up to a "model wing" that works well enough to fly.
The first approach could be called "top-down." It begins by thinking about the task to be accomplished (the "top" of the problem), and works down toward the means to solve it. In this approach, we don’t worry too much about whether the solution is much like nature, although any wing we build will certainly teach us about the principles that govern all wings, including birds’. The second approach could be called "bottom-up." It begins by looking at the details of how evolution has already solved the problem. By looking at the parts, and building up from that "bottom," this approach also solves the problem and ultimately illuminates the general principles at work in the problem area.
The top-down and bottom-up approaches both make sense when we turn to the mind (especially when we think of it as a network of biological neurons). We could start at the top with some of the abilities of the brain, as explored in several labs over recent weeks, and ask how neural networks could embody (simplified) forms of intelligent behavior. Along this path we would worry less about the biological realism of the networks we might invent to do what brains do. Instead, we would focus on the capacities of neural networks in general, thinking about our network as a suggestive analogy that tells us what to look for in the biological brain. Or we could start at the bottom, studying some real neurons. In this case, we’d ask what these real neurons are doing, how they interact with other cells, and eventually how all the interactions add up to a system that does the things brains do.
Both approaches are valuable, and most research in either the mind or brain is a combination of the two. We will explore our brains using both approaches as well. The first road begins at the bottom, with a biologically realistic simulation of retinal ganglion cells. We will use an elegant and award-winning software package called MacRetina. As described by its author, "MacRetina simulates an experiment in which students record from retinal ganglion cells, the cells in the retina that send information from the eye to the brain. By sampling neural activity while stimulating with small spots of light, students can see the dynamic excitatory and inhibitory responses of these neurons, and map the organization of the retinal region that drives each cell (the cell's receptive field). Ten cells are modeled. Their overlapping receptive fields show how a visual stimulus is detected by many neighboring ganglion cells."
Background to the lab:
Before you begin, review the following notes on retinal ganglion cells, also written by MacRetina’s author, Richard Olivo:
MacRetina simulates an experiment to record action potentials from retinal ganglion cells while stimulating them with dark or light spots projected onto the retina. The retinal ganglion cells are the last in a sequence of five types of nerve cells in the retina; their axons group together to form the optic nerve, carrying information about light-dark contrast to the brain. There, extensive additional processing occurs that eventually leads to the perception of shape, pattern, movement, and the identification of objects.
The eye has about one million retinal ganglion cells, each one of which responds to light in its own small region of the retina. This region is called the ganglion cell's receptive field. The receptive fields of nearby ganglion cells overlap, so that any stimulus projected onto the retina will fall onto the receptive fields of dozens or even thousands of retinal ganglion cells (depending on how big the stimulus is). Any one point in visual space is sampled by a group of neighboring retinal ganglion cells.
The receptive fields of these cells differ, however. Many ganglion cells are excited best by a bright spot against a darker background. These cells are also activated well by appropriately positioned bright lines, and by edges between light and dark. Most of the remaining ganglion cells are the opposite: excited by dark spots against a brighter background, and also by dark lines or the dark part of an edge. Thus, two parallel systems exist, one for finding bright regions and one for finding dark regions in the visual world.
The first group, excited by bright spots, is referred to as "on-center" ganglion cells because the cells fire action potentials when a properly positioned spot of light is turned on. The second group, called "off-center" ganglion cells, was so named because the cells stop firing when a bright spot is placed on the center of the receptive field, but they then give a burst of action potentials when the bright spot is turned off (an "off-response," hence the name "off-center"). We now realize that a more easily understood stimulus for these cells is a dark spot (or a dark line), which excites the cell when the dark stimulus is present. Thus, the off-center cells are analogous to the on-center cells except that they are excited by a dark spot instead of a bright spot. Together, these two groups of ganglion cells analyze the visual image, detecting contrasting regions of light against dark (on-center) or dark against light (off-center).
Uniform fields of light are poor stimuli for either type of retinal ganglion cell. The ganglion cells' receptive fields are not just simple circles responding to light or dark, but are surrounded by a region (like a doughnut) that has the opposite response from the center. On-center cells have a surround that gives an off-response (and fires in response to dark stimuli), while off-center cells have an on-surround (firing in response to bright stimuli). Moreover, the two regions antagonize each other: light falling on the receptive field's surround cancels the response to light falling on the center. For an on-center ganglion cell, a bright spot filling the center of its receptive field (but not spilling onto the dark surround) is the best stimulus. If the spot becomes bigger, so that it covers part of the surround, the ganglion cell's response decreases. If the spot completely covers the surround (equivalent to a region of uniform illumination), the response is canceled. Lines and edges are adequate stimuli if they cover the center but not the surround. If a satisfactory edge is moved so that the entire receptive field is now in the light or the dark, the ganglion cell will stop responding. (The edge will now be well positioned, however, to stimulate some other, neighboring ganglion cells, so it will still be seen -- but in a different location.)
The synaptic network that drives retinal ganglion cells to give them their center-surround receptive fields is constructed from the four other cell types in the retina. Light is captured by the rods and cones, which are the only retinal cells that are directly sensitive to light; they in turn drive horizontal and bipolar cells (the stage at which the antagonistic center/surround receptive field is created). The bipolar cells then drive retinal ganglion cells (directly, and also indirectly through amacrine cells). Further details are beyond the scope of MacRetina, but sources of additional information are given in the "References" help file.
Finally, one last important aspect of retinal ganglion cells is related to whether the cells are good at detecting moving patterns. Some ganglion cells respond well to stationary stimuli, producing a sustained train of action potentials as long as the stimulus is present. In the cat retina, these sustained-responders are called X-type ganglion cells. They have relatively small receptive fields, they are more prevalent in the center of the retina, and they seem to be in the business of detecting the details of fine patterns. About half the X-cells have on-centers, and the other half have off-centers.
Other ganglion cells respond transiently to stationary stimuli, giving a brief burst of action potentials when the stimulus is first presented, but then falling silent. They respond well to moving stimuli, since such stimuli remain in a cell's receptive field only briefly. These transient responders are named Y-type ganglion cells. Their receptive fields are larger, they are more prevalent in the periphery of the retina, and their business seems to be detecting moving edges. They too come in both on-center and off-center varieties. Thus, any particular ganglion cell can be classified as on-center or off-center, and sustained (X) or transient (Y). MacRetina models a region of retina where both X and Y type ganglion cells are present, so a given cell might be On-X, Off-X, On-Y or Off-Y. The best way to distinguish the X and Y types is to aim a stimulus spot at the center of the receptive field (once it has been mapped), turn on the stimulus for several seconds, and see if the firing is sustained (X) or transient (Y).
MacRetina models ten ganglion cells from a 0.5 by 1.0 mm region of retina, a region in which almost 100 ganglion cells are present. Even by mapping only a fraction of the total cells, one can see how extensively receptive fields overlap, and thus how any particular visual stimulus will excite many ganglion cells.
The lab:
This lab is best done with a partner. Also, try to find a partner from the other course. Pat, please pair up with Chris.
1. Locate the MacRetina folder in the Class Software file area (located in Network servers). PLEASE Drag the MacRetina Folder out of Class Software to the surrounding Desktop. (You may make a personal copy of the software, if you want.)
2. Open the MacRetina folder and click on the MacRetina 2.1 icon. MacRetina will open and urge you to follow the tutorial program before getting started. Do it.
3. That is, pull down the Help menu and highlight Tutorial. This will show you everything you need to know.
Collecting data and writing it up:
1. Map the ten ganglion cell receptive fields simulated by MacRetina. You need not save or hand in your maps, but for each cell, list whether it is an on-center or off-center cell, and whether its response is sustained (X) or transient (Y). Accuracy matters here, so if you are uncertain about a cell, consult with another group and see if you can resolve your uncertainties.
The remaining questions are all bottom-up extensions of this raw data on the ganglion cells.
2. The area of the retina simulated in MacRetina would contain over 100 ganglion cells. As you can see from even the ten that you sampled, ganglion cell receptive fields overlap greatly. Why? You might think of the question this way: On Rigel IV, a species of humanoid has evolved that is just like us except for one tiny detail: their ganglion cell receptive fields do not overlap. Do we enjoy any advantages over Rigelians when it comes to perception? If you have more than one idea here, fine!
3. The receptive fields you mapped all have a characteristic donut shape, with an excitatory center surrounded by a region of inhibition. Why? What advantage in perception does the inhibitory surround provide? You might think of the question this way: On Tau Ceti, a species of humanoid has evolved that is just like us, only their ganglion cell receptive fields do not have inhibitory surrounds. (Their ganglion cells are either dark-spot detectors or light-spot detectors.) How might their vision of objects in their world differ from ours? (If you have more than one idea, fine!)
4. The cells also differ in their responses -- some are sustained, and some are transient. Why?
5. In the spirit of the horny fly exercise, how could ganglion cells be wired together to detect an straight edge that is light on one side and dark on the other? Note that your edge detector will also be orientation-sensitive. (It will detect only vertical edges, for example.) You can reply to this with a diagram or with a brief verbal description.
Make sure that both names are on your lab report. If you hand in your report on Docex, make sure that both names appear in the filename, as in "Smith and Jones Lab 7," for example. Drop files in Docex/Phil. 371/Drop.