Chiara Bartolozzi (IIT Genova) started off after the coffee break by asking, what do we need to make our systems work better?
Rodney Douglas asked Matthew Diamond to distinguish perception from deciding. He responded that in primates, it is possible by providing information and training or specifying a delayed response. That way, the percept can somehow be split in time to allow their separate study.
In rats, it is possible to train to perceive using whisking, but we are never sure about this. For example, we can control the result of an experiment to train context dependent responses, e.g., choosing a water spout rather than a button, but this is really hard. So in animals the responses are more closely bound to stimuli. E.g. rough texture means turn right. Anything past this is really hard to train.
Abisheck Banerjee asked if cross-modal behavior can be trained. Matthew gave concrete example, if rats is trained to turn right for strong vibration, left if weak. Then in only 10 trials they can generalize to strong vibration vs weak vibration.
Giacomo pointed out that we generally go from continuous input to discrete symbols, but where is this shift occurring?
Matthew responded that they only use graded stimuli, because they are interested in sensory coding, so to characterize the responses they need this analog input to measure the response functions.
He finally had a two word answer: "Atrractor dynamics", to lead to one of the next presenters.
Rodney asked are the decisions localized to posterior parietal cortex (PPC)?
Matthew answered that PPC is involved in the rat discrimination of angle task by whisking of ridges. Are the ridges at
angle greater or less than 45 degrees. PPC is located at the red labeled spot.
Some PPC cells actually center their responses around the decision point for the reward task. Others respond to the stimulus, as shown in this sketch. PPC is not a whisking or visual area but rather an integration area.
There followed quite a long discussion in the audience about connectivity and how van Essen's original cortical hierarchy has changed since its original presentation.
Matthew pointed out that the understanding of a thalamic-cortical-thalamic loop makes it really hard to understand how gain control, feature selectivity, etc are controlled. There is a reticular formation around the thalamus that could produce cross modal inhibition and interaction, but only 1/50 of people work on thalamus compared to cortex.
Rodney pointed out that Denis yesterday pointed out that layer 5 collaterals connect in such way that they could support integration.
The rest of the session discussed other issues, including temporal precision in audiotory system. Chiara asked about temporal precision in audition. Andre pointed out that audition does "where" before "what" since where involves interaural time difference changes of <10us. And this processing must therefore be done right at the auditory sensory input, since preserving such timing later on would be too costly energetically.
Barry Richmond pointed out that only cochlear nucleus, LGN, substantia niagra (SN), and a few others turn black with deoxy labeling showing they burn a lot of power. Why SN needs this is unknown. But only a few places need this high bandwidth.
Matthew talked about temporal precision in relation to his own group's work on mouse barrel cortex. The question is, what is the code for the location of the whisker stimulus, when multiple whiskers are simultaneously stimulated, but some of them press more? One part is rate code: Neurons that fire most are the ones that are touched. But there is also a time code. Its basis is that stimulated neurons fire e.g. 10ms earlier, so there is also information in the timing. Information measurement done with Stefano Panzeri showed this timing is equally informative with rate. It corresponds to about 5ms jitter for the temporal precision. It is not that surprising that a neuron that is stimulated more fires earlier, but it shows that these neurons are not Poisson generators.
With active behavioral tasks, they studied how rats discriminate textures by active whisking (as opposed to the passive responses from earlier experiments). By training rats to distinguish textured rough surfaces, is that, in paper from 2015, that the neurons carry information by multiplexing. The quantity is carried by rate and pattern. Mostly in rate but also in pattern.
Specifically the pattern of activity is related to the roughness and stick/slip responses of the whiskers. So even if the neurons responds with the same number of spikes, then the pattern of spikes still carries information about the surface in the pattern.
Finally Matthew showed Stefano's interesting "intersection information" analysis of stimulus-neuronal response and neuronal response-decision information. What is the overlap? There is overlap in both rate and pattern. I.e. rate information is converted to choice but so does pattern. This "intersection information" viewpoint and method of analysis from the 2015 Zuo paper could be a big step forward to understanding neuronal coding.
Final experiment asked question, what if we give animal stimulus that produces identical long term spike count of neuronal responses but different modulation, then the animal can still distinguish the patterns that must only be carried by spike timing. This was also shown by intersection information analysis. Perhaps not surprising since auditory system does this all time, but this analysis puts information bounds on the experiment.
Rodney Douglas asked Matthew Diamond to distinguish perception from deciding. He responded that in primates, it is possible by providing information and training or specifying a delayed response. That way, the percept can somehow be split in time to allow their separate study.
In rats, it is possible to train to perceive using whisking, but we are never sure about this. For example, we can control the result of an experiment to train context dependent responses, e.g., choosing a water spout rather than a button, but this is really hard. So in animals the responses are more closely bound to stimuli. E.g. rough texture means turn right. Anything past this is really hard to train.
Abisheck Banerjee asked if cross-modal behavior can be trained. Matthew gave concrete example, if rats is trained to turn right for strong vibration, left if weak. Then in only 10 trials they can generalize to strong vibration vs weak vibration.
Matthew responded that they only use graded stimuli, because they are interested in sensory coding, so to characterize the responses they need this analog input to measure the response functions.
He finally had a two word answer: "Atrractor dynamics", to lead to one of the next presenters.
Rodney asked are the decisions localized to posterior parietal cortex (PPC)?
Matthew answered that PPC is involved in the rat discrimination of angle task by whisking of ridges. Are the ridges at
angle greater or less than 45 degrees. PPC is located at the red labeled spot.
Some PPC cells actually center their responses around the decision point for the reward task. Others respond to the stimulus, as shown in this sketch. PPC is not a whisking or visual area but rather an integration area.
There followed quite a long discussion in the audience about connectivity and how van Essen's original cortical hierarchy has changed since its original presentation.
Matthew pointed out that the understanding of a thalamic-cortical-thalamic loop makes it really hard to understand how gain control, feature selectivity, etc are controlled. There is a reticular formation around the thalamus that could produce cross modal inhibition and interaction, but only 1/50 of people work on thalamus compared to cortex.
Rodney pointed out that Denis yesterday pointed out that layer 5 collaterals connect in such way that they could support integration.
The rest of the session discussed other issues, including temporal precision in audiotory system. Chiara asked about temporal precision in audition. Andre pointed out that audition does "where" before "what" since where involves interaural time difference changes of <10us. And this processing must therefore be done right at the auditory sensory input, since preserving such timing later on would be too costly energetically.
Barry Richmond pointed out that only cochlear nucleus, LGN, substantia niagra (SN), and a few others turn black with deoxy labeling showing they burn a lot of power. Why SN needs this is unknown. But only a few places need this high bandwidth.
Matthew talked about temporal precision in relation to his own group's work on mouse barrel cortex. The question is, what is the code for the location of the whisker stimulus, when multiple whiskers are simultaneously stimulated, but some of them press more? One part is rate code: Neurons that fire most are the ones that are touched. But there is also a time code. Its basis is that stimulated neurons fire e.g. 10ms earlier, so there is also information in the timing. Information measurement done with Stefano Panzeri showed this timing is equally informative with rate. It corresponds to about 5ms jitter for the temporal precision. It is not that surprising that a neuron that is stimulated more fires earlier, but it shows that these neurons are not Poisson generators.
With active behavioral tasks, they studied how rats discriminate textures by active whisking (as opposed to the passive responses from earlier experiments). By training rats to distinguish textured rough surfaces, is that, in paper from 2015, that the neurons carry information by multiplexing. The quantity is carried by rate and pattern. Mostly in rate but also in pattern.
Specifically the pattern of activity is related to the roughness and stick/slip responses of the whiskers. So even if the neurons responds with the same number of spikes, then the pattern of spikes still carries information about the surface in the pattern.
Finally Matthew showed Stefano's interesting "intersection information" analysis of stimulus-neuronal response and neuronal response-decision information. What is the overlap? There is overlap in both rate and pattern. I.e. rate information is converted to choice but so does pattern. This "intersection information" viewpoint and method of analysis from the 2015 Zuo paper could be a big step forward to understanding neuronal coding.
Final experiment asked question, what if we give animal stimulus that produces identical long term spike count of neuronal responses but different modulation, then the animal can still distinguish the patterns that must only be carried by spike timing. This was also shown by intersection information analysis. Perhaps not surprising since auditory system does this all time, but this analysis puts information bounds on the experiment.
Comments
Post a Comment