Rained very hard last night, continuing today.
New people: Andrew Rowley from Manchester ("Helping to use SpiNNaker effectively). Julia Buhmann, connectomics, DNNs, ("Do not eat meat"). Florian Engert, neuroscience, larval zebrafish ("For sale: fish feelings lovingly explained").
For Friday, 2h wrapup, followed by demo sessions.
Florian Engert started off talking about fish feelings, technically valence, claiming that audience would think this more interesting than mechanisms or networks.
He took larval zebrafish as example, which they study. There are 100k neurons in this vertebrate brain. They have mammalian behaviors. They are translucent so they can measure all neurons.
Then somehow a discussion of suffering started. By definition alpha fibers are stimulated. But these still need a brain to turn pain to what is defined as suffering. It is pain input as represented in the brain, not peripheral nervous system.
He then drew a scale, "Bad - Good". On this scale, pain, hunger, loneliness, cold, all are associated with bad valence. He argued that suffering is an evolutionary adaption that crosses all species and that it interesting to think about it for AI. What is the mechanistic model of feelings that might allow it to be added to machines.
He then drew this sketch of the fish and talked about oxytocin neurons (the blue ones) that represent internal noxious state or input. The cells release oxytocin. They seem to target spinal column. They have glutamate also. This system is neuromodulatory. Meister calls it a plumbing system of the brain, as opposed to the electrical system that acts quickly. Zebrafish have a a hundred of these neurons in total, bilaterally. On the sketch of the oxytocin system responses, the vertical axis is activity measured by calcium imaging. Food and sex don't count for these larva (who still live on their yolk). Cocaine does nothing. Visual and auditory likewise do nothing. Mild heat makes it go up. Mustard oil burns them. Electric shocks stimulate these cells the most.
Now they can drive these same cells with channelrhodopsin, and slowly over about a minute this stimulation has an effect that the tail starts flicking, indicating that the fish is trying perhaps to escape.
They grow up in clutches of 200 fingerlings that grow up in protected area. If fish are put in isolation they "suffer". They label paramecium with dye and this florescence can be measured and quantified by quantal jumps of florescence. The pain threshold is lower when they are isolated. Moritz posited that maybe they eat less when isolated since there is less competition. Florian agreed this was possible.
How do they know they are isolated? If they are put in glass chamber so they can see siblings. They are still lonely. But the trick is to put fish water in with them. Then they can smell their siblings and even in darkness they know they are not alone. Old fish water, goldfish water don't work. Adult fish water scares them and they eat less. The biggest enemy of larval fish are adults, who eat them.
They also observe that "happy water" lights up particular glomuleri in the olfactory bulb. Social deprivation slightly lights up oxytocin system moderately. Then they can use this activity level also to measure this "suffering". By definition, reflexive actions that do not require rumination are reactive, while others are not.
Finally, they found a specific circuit where the olfactory bulb projects inhibition onto oxytocin system to suppress its response to other stimuli, which goes to explain the elevated pain threshold to other noxious stimuli.
Next he talked about serotonin (5HT). In the context of hunger, they can make them hungry and measure how fast they eat after they are made hungry. If they then measure 5HT activity (the red blobs), then they see the trace below: 5HT goes up during fasting. They fish are exposed to food, but cannot eat it. 5HT activity drops very low. As they are allowed to eat, 5HT goes up again.
It indicates that 5HT encodes being hungry and consuming food. Knowing they can eat is good, so 5HT drops. This then explains a big mystery of serotonin measurements that they seem in humans to encode both hunger and satiation. These direct measurements show that serotonin promotes finding food when hungry, and consuming. These two modes, finding food and eating are expressed by different behavioral states.
To use this in AI, you need labeled lines that encode the noxious inputs, which the network can use. They should be modulatory, so they can control expression or probability of entering behaviors.
That brought us to the coffee break. It was still raining hard.
After some ping pong and coffee, Georg Keller continued the discussion, moved the question from "Why do we move?", i.e. what is the incentive signal, to "How do we move?", i.e. what are the cortical signals.
Georg described experiments from Adam Kampff where the rat is trained to run over unstable ladder. Then they make one of rungs rotate to slip. A normal rat instantly within 100ms changes to different motor pattern. An M1 lesioned rat freezes. It takes them 3-4s to find another strategy. So M1 is a memory for alternative motor plans that can be quickly retrieved.
The experiment the Keller lab is doing now is based on this experiment. M1 is very thin; it has mostly layer 2-3 and 5. 5 outputs subcortically and 2-3 get intracortical input. They train a mouse to run down a virtual corridor (on a mouse ball) to get a drunk and a well trained mouse runs a kind of S contour. A mouse with inhibited M1 does a complete spaghetti pattern, as though the lights were off. They do a perturbation where they instantly shift the target +-30deg. Then they see that normal mice correct for this within about 100ms. If M1 is inhibited for 2s, then they keep running straight, as soon as inhibition is removed they correct.
What does M1 activity look like? It is shown in this sketch:
The L2/3 and L5 activity is linear with turn acceleration, during spontaneous turns. Also, activity is larger when turn is towards target than when away from target. Then when they look at correlation between acceleration and layer activity after perturbation, then they see L5 is still correlated, but L2/3 activity is just flat. Indeed, it looks like L2/3 signals an error in the target, in some motor space, maybe heading direction.
Next Yulia Sandarmirskaya took over. She talked about a simplified motor system consisting of an eye, which is moved by muscles to pan and tilt. Saccadic eye movements occur at about 3Hz, open loop. There is some learning involved in these eye movements. There are several basic characteristics of saccades.
- There is a main sequence of saccades. They have a velocity on vertical axis. Larger or smaller saccades take about the same time.
- If you make a fixation target, then have users saccade to target, and if target moves during saccade toward fixation point, then users don't notice the perturbation, but make a correction saccade after the main one. After a few hundred trials, users start making shorter saccades. The adaptation is local to saccade target direction and distance.
- There are "memory saccades". The subject sees two saccades to two stimuli briefly presented. The subject must plan two saccades, the first to the first target, the second from first to second target. Babies just add the two original saccades. But adults can do this task. How do they do it?
The circuit looks a bit complicated but it is straightforward to implement. It can learn to make saccades, and can adapt as humans do to saccade perturbations.
What about memory saccades? They are handled by storing the salient points in motor coordinates.
Matthew Cook then explained that this system is a really an example of a relational network as he described yesterday. There are connections simply because the variables are represented in two different maps.
Then he explained how using such a relation it is simple to do coordinate transforms by this sketch. It allows transforming from x,y to s,t coordinates, or polar coordinates by the right connections from the 2d map to the 1d projections. For example for rho and theta from x and y, the rho outputs would be connected radially along angles, while the rho outputs would be connected circumferentially to larger and larger radii.
There followed a very nice description of learning relationships between variables and how these can be scaled by decomposition.
But what about motor control when there are many programs running simultaneously? Matt briefly outlined an intruiging plan to put behavior selection in the same relational network scheme, where higher levels send utilities of actions to lower finer grain areas (e.g. limbs) which return probability of success of the actions.
Fredrik Sandin asked does it solve the binding problem? Matt answered that after a long discussion with Christoph von der Marlsberg, he eventually convinced him that it does, if you look at it carefully.
The interesting discussion brought us to the afternoon sports break.
Comments
Post a Comment