Frequently Asked Biological Questions About LISSOM

Q.
The long-range connections in LISSOM are inhibitory, yet most of the long-range horizontal connections connect excitatory neurons to other excitatory neurons. How can you explain that?

A.
LISSOM models the cortex at a slightly higher level than individual neurons and their connections, modeling only each vertical column (as a single unit) and the interactions between the columns. In addition, the focus is on self-organization, which in a Hebbian model is driven primarily by high contrast inputs. In this restricted regime, the lateral interactions have been clearly demonstrated to be inhibitory, as a result of interactions with local inhibitory interneurons, even though at low contrasts the effects are excitatory (Hirsch and Gilbert 1991, Weliky et al 1995). If we start to study low-contrast inputs in the future, we will need to extend LISSOM to use a more complicated recurrent local circuit that would have the correct behavior at low contrasts, but because that would slow down the already-long run time drastically and make the model more complex, we have avoided it for now.

Q.
The published LISSOM simulations start out with a large excitatory lateral connection radius and then gradually reduce it to cover a much smaller area. There is no evidence for this massive shrinkage, so...?

A.
The shrinking radius was inherited from SOM, and is not a required part of the model. It does help to make sure that the entire map becomes globally ordered even when the starting point for self-organization is extremely disorganized. However, there are several other more-plausible ways to achieve the same effect, such as gradually increasing the effect of inhibition relative to excitation. Moreover, there is no evidence that there is ever a time that the cortex is quite as disorganized as the starting points we use in the simulations.

Q.
So, why do you use such an implausible starting point?

A.
We are studying computation, and we wish to know the true capabilities of the algorithms we are investigating. We deliberately make the task harder for the algorithm than is likely to occur in nature, to show that it is robust. The state of the model at any point during training can be chosen instead as a starting point, and compared to the biological case from that point forward. We do not take this step explicitly ourselves because we are not modeling any particular species, and the known "starting point" clearly differs between species.

Q.
There is little evidence of any inhibitory interactions in the very earliest stages of development, yet LISSOM has inhibitory connections from the start. Why?

A.
Immature cortical neurons appear to be very difficult to excite already, and thus probably do not require any further inhibition to prevent runaway (epileptic) responses. We hypothesize that inhibitory influences arise automatically as soon as the neurons begin responding more strongly, thus maintaining a balance between excitation and inhibition. The model does not currently contain any mechanism for this balancing process, and instead just sets excitation and inhibition to be approximately equal by fiat, which is probably much simpler to do in a model than in a real organism. A complete model would need to incorporate such a control mechanism, but we try to keep the model as simple as we possibly can unless we are looking at a phenomenon dominated by a particular mechanism.

Q.
Isn't the orientation map already organized at eye-opening? How can you explain that in an input-driven Hebbian model?

A.
Intrinsic nervous-system activity can probably account for this, although it is not yet clear which particular source is responsible. One such possibility is retinal waves, patches of activity known to exist in the retina before eye opening (Meister et al 1991, Wong et al 1993). It has been claimed that retinal waves are not sufficiently oriented to account for orientation selectivity (Miller 1994). However, we have shown that even entirely unoriented patterns can drive such selectivity if the patterns have sharp edges (Bednar and Miikkulainen 2003). It is not known if retinal activity reaches the cortex in time to drive development, but there are also a number of other candidates, including spindle waves and the PGO waves of REM sleep. See (Bednar 2002) for more details.

Q.
Work by Chapman,Stryker,and Bonhoeffer (1996) showed that the orientation map remains fairly static as it becomes sharpened by visual experience, but in LISSOM the orientation patches can be seen to vary in location as self-organization progresses. Why?

A.
The initial chaos is due to having a high learning rate and to having a shrinking lateral excitatory radius; once those conditions cease, the map is stable. Simulations without a shrinking radius and with only a short period with a high learning rate (not yet submitted for publication) show a stable map once orientation selectivity has been established.

Q.
My question or critique was not answered by this FAQ. What should I do?

A.
Send email to jbednar@inf.ed.ac.uk, or to the corresponding author of the paper you read! We much appreciate getting feedback.

References