Final Report Summary - SOUNDSOURCE_RAT_AC (Sound Localization by Neuronal Populations in the Rat Auditory Cortex)
When we determine that our friend is in the room next door based on the sound of his voice standing out of the noise in our own room, we are making a perceptual decision: we are making the judgement that it is her voice (and not someone else’s) on the basis of ambiguous sensory evidence. Over the last 30 years, the study of well-controlled simple sensory discriminations like this one, has laid the groundwork for our current understanding of the mechanisms used by the brain to convert evidence into categorical choices. According to the current view, when subjects are asked to make a categorical judgment between two options based on ambiguous sensory information, they do this by accumulating evidence about the extent to which the two options are differentially present in the environment. When the accumulated difference is sufficiently positive or negative, larger than a critical bound, a choice is made.
Our work has sought to advance our understanding of this basic paradigm in three different directions: (1) We have studied the role of the overall stimulus magnitude on perceptual decision accuracy and reaction time. Consider discriminating the length of two lines of 10 vs. 10.1 cm and discriminating the length of two lines of 100 vs. 100.5 cm. Which task is easier? Is the relative or the absolute difference that counts? Whereas our intuition suggests that the latter task is harder, the role of overall magnitude in shaping choice and reaction time had not been thoroughly studied, nor was it clear how to include the overall stimulus magnitude into the theoretical framework of evidence accumulation. (2) We have studied the role of control on optimal decision-making (DM) strategies. Specifying clearly the goal of a DM agent, allows one to derive optimal strategies or policies necessary to achieve this goal. For instance, if the goal of the subject is to maximize reward in a DM task under a certain cost for processing time, it is possible to compute when the agent should commit to a choice and stop accumulating evidence. This strategy is typical in Reinforcement Learning, where one computes optimal policies given the task contingencies and processing limitations. One key element was missing, though, which is to factor into the optimization process how costly is it for the agent to adapt her behavior to the demands of the task. We refer to this as the cost of control. It wasn’t known what would be the effect of a cost of control on the optimal evidence accumulation strategy in a perceptual DM task. (3) One critical aspect of the DM framework, is that when the stimulus is ambiguous with respect to the category boundary, noisy fluctuations in the internal neural representation of the stimulus become important, and this causes significant variability both on which choices are made, as well as on the time it takes to make them, even when the stimulus is identical. Thus, understanding the nature of trial-to-trial variability in the responses of sensory neurons is critical to understand the accuracy of perceptual choices. Critically, the advent of multi-channel recording techniques allows one to measure the simultaneous activity of many (tens to hundreds, depending on the technique) neurons, which permits quantifying variability at the population level. Whereas there’s been ample theorizing on how different population variability regimes affect the information about a sensory stimulus in a population code, little was known about this question empirically. We sought to perform such a characterization, initially under anesthesia, for the sounds that we were using in our DM experiments.
We have made significant progress in each of these three questions. Regarding question (1), we first established a behavioral paradigm to estimate reliably the accuracy of sensory discriminations in freely moving rats. This in itself is an achievement, as it is difficult in studies with rodents to control what are the sources of uncertainty that constrain their choices. Through behavioral controls and manipulations, we were able to demonstrate that accuracy is constrained by sensory uncertainty in our task. Next, we confirmed that, when comparing two quantities (in our task the difference in sound level in each ear resulting from a sound delivered through headphones), accuracy is proportional to the mean level, implying a constant just noticeable difference (JND) when measuring differences in a logarithmic scale, such as dB. (Weber’s law). Unexpectedly, however, we discovered a previously unknown relationship between reaction time (RT) and the effect of stimulus magnitude, namely, the same accuracy at different levels is associated to temporally re-scaled RT distributions, leading to longer RTs for fainter sounds. This finding is critical because, in the evidence accumulation framework, RT and accuracy can be considered two sides of the same coin. How then, can one obtain different RTs with the same accuracy? Using computational modeling, we found that the answer to this puzzle requires that the neurons encoding the sound level on each ear have multiplicative trial-to-trial variability, and that they encode level as a strongly compressive function. The resulting model robustly explains our findings with remarkable precision. Our results reveal an unexpected connection between two fundamental quantities in sensory psychophysics: the constancy of the JND and the distribution of RTs.
Regarding question (2), we have made both technical as well as conceptual developments. At a technical level, we have extended the linear Markov decision process (LMDP) framework developed by Emo Todorov to allow the study of problems with sensory uncertainty. Using our extended framework, we have discovered that when the cost of control is significant, the optimal DM strategy involves stochastic evidence bounds. We have studied what are the practical implications of this control-limited regime for observable quantities such as accuracy, RT or decision-confidence. These can be thought of as predictions that allow identification of the cost of control of an agent. Preliminary results from our behavioral experiments show that rats indeed behave as control-limited optimal decision-makers. Our results suggest that the cost of control is a key variable shaping the behavior of decision-makers.
Regarding question (3), we have revealed a strong state-dependence on the representation of stimulus intensity in the auditory cortex. Our findings reveal that during synchronized states (typical of slow wave sleep and quiet wakefulness) the responses of the population conform to previous intuitions, with most neurons firing more strongly to loud, contra-lateral sounds. Surprisingly, however, during desynchronized states (typical of REM sleep and attentive wakefulness), single neurons prefer similarly often quiet as loud sounds, and ipsi- and contra-lateral sounds. This leads to a state-dependent switch in coding strategies: the desynchronized cortex uses an identity code in which different sounds evoke overall similar numbers of spikes from the whole population, and different neurons prefer different sounds. In contrast, the synchronized cortex uses an intensity code in which the nature of sound is encoded in the overall number of spikes fired by the local circuit. Because the trial-to-trial variability also modulates most neurons similarly in this state, stimulus representation is overall degraded in the synchronized cortex. These results predict that the accuracy of sensory judgements should be correlated to the level of cortical desynchronization, a prediction which we are currently testing.
In summary, we have used a multidisciplinary strategy involving behavior, theory and electrophysiology to generate qualitative advances in our understanding of outstanding issues in perceptual DM and sensory representations. Our quantitative approach should be particularly useful for identifying phenotypes of DM behavior in normal and disease states.
Our work has sought to advance our understanding of this basic paradigm in three different directions: (1) We have studied the role of the overall stimulus magnitude on perceptual decision accuracy and reaction time. Consider discriminating the length of two lines of 10 vs. 10.1 cm and discriminating the length of two lines of 100 vs. 100.5 cm. Which task is easier? Is the relative or the absolute difference that counts? Whereas our intuition suggests that the latter task is harder, the role of overall magnitude in shaping choice and reaction time had not been thoroughly studied, nor was it clear how to include the overall stimulus magnitude into the theoretical framework of evidence accumulation. (2) We have studied the role of control on optimal decision-making (DM) strategies. Specifying clearly the goal of a DM agent, allows one to derive optimal strategies or policies necessary to achieve this goal. For instance, if the goal of the subject is to maximize reward in a DM task under a certain cost for processing time, it is possible to compute when the agent should commit to a choice and stop accumulating evidence. This strategy is typical in Reinforcement Learning, where one computes optimal policies given the task contingencies and processing limitations. One key element was missing, though, which is to factor into the optimization process how costly is it for the agent to adapt her behavior to the demands of the task. We refer to this as the cost of control. It wasn’t known what would be the effect of a cost of control on the optimal evidence accumulation strategy in a perceptual DM task. (3) One critical aspect of the DM framework, is that when the stimulus is ambiguous with respect to the category boundary, noisy fluctuations in the internal neural representation of the stimulus become important, and this causes significant variability both on which choices are made, as well as on the time it takes to make them, even when the stimulus is identical. Thus, understanding the nature of trial-to-trial variability in the responses of sensory neurons is critical to understand the accuracy of perceptual choices. Critically, the advent of multi-channel recording techniques allows one to measure the simultaneous activity of many (tens to hundreds, depending on the technique) neurons, which permits quantifying variability at the population level. Whereas there’s been ample theorizing on how different population variability regimes affect the information about a sensory stimulus in a population code, little was known about this question empirically. We sought to perform such a characterization, initially under anesthesia, for the sounds that we were using in our DM experiments.
We have made significant progress in each of these three questions. Regarding question (1), we first established a behavioral paradigm to estimate reliably the accuracy of sensory discriminations in freely moving rats. This in itself is an achievement, as it is difficult in studies with rodents to control what are the sources of uncertainty that constrain their choices. Through behavioral controls and manipulations, we were able to demonstrate that accuracy is constrained by sensory uncertainty in our task. Next, we confirmed that, when comparing two quantities (in our task the difference in sound level in each ear resulting from a sound delivered through headphones), accuracy is proportional to the mean level, implying a constant just noticeable difference (JND) when measuring differences in a logarithmic scale, such as dB. (Weber’s law). Unexpectedly, however, we discovered a previously unknown relationship between reaction time (RT) and the effect of stimulus magnitude, namely, the same accuracy at different levels is associated to temporally re-scaled RT distributions, leading to longer RTs for fainter sounds. This finding is critical because, in the evidence accumulation framework, RT and accuracy can be considered two sides of the same coin. How then, can one obtain different RTs with the same accuracy? Using computational modeling, we found that the answer to this puzzle requires that the neurons encoding the sound level on each ear have multiplicative trial-to-trial variability, and that they encode level as a strongly compressive function. The resulting model robustly explains our findings with remarkable precision. Our results reveal an unexpected connection between two fundamental quantities in sensory psychophysics: the constancy of the JND and the distribution of RTs.
Regarding question (2), we have made both technical as well as conceptual developments. At a technical level, we have extended the linear Markov decision process (LMDP) framework developed by Emo Todorov to allow the study of problems with sensory uncertainty. Using our extended framework, we have discovered that when the cost of control is significant, the optimal DM strategy involves stochastic evidence bounds. We have studied what are the practical implications of this control-limited regime for observable quantities such as accuracy, RT or decision-confidence. These can be thought of as predictions that allow identification of the cost of control of an agent. Preliminary results from our behavioral experiments show that rats indeed behave as control-limited optimal decision-makers. Our results suggest that the cost of control is a key variable shaping the behavior of decision-makers.
Regarding question (3), we have revealed a strong state-dependence on the representation of stimulus intensity in the auditory cortex. Our findings reveal that during synchronized states (typical of slow wave sleep and quiet wakefulness) the responses of the population conform to previous intuitions, with most neurons firing more strongly to loud, contra-lateral sounds. Surprisingly, however, during desynchronized states (typical of REM sleep and attentive wakefulness), single neurons prefer similarly often quiet as loud sounds, and ipsi- and contra-lateral sounds. This leads to a state-dependent switch in coding strategies: the desynchronized cortex uses an identity code in which different sounds evoke overall similar numbers of spikes from the whole population, and different neurons prefer different sounds. In contrast, the synchronized cortex uses an intensity code in which the nature of sound is encoded in the overall number of spikes fired by the local circuit. Because the trial-to-trial variability also modulates most neurons similarly in this state, stimulus representation is overall degraded in the synchronized cortex. These results predict that the accuracy of sensory judgements should be correlated to the level of cortical desynchronization, a prediction which we are currently testing.
In summary, we have used a multidisciplinary strategy involving behavior, theory and electrophysiology to generate qualitative advances in our understanding of outstanding issues in perceptual DM and sensory representations. Our quantitative approach should be particularly useful for identifying phenotypes of DM behavior in normal and disease states.