P272 Modeling unsigned temporal difference errors in apical dendrites of L5 neurons
Gwendolin Schoenfeld1,2,3,Matthias C. Tsai*,1,4, Walter Senn4, Fritjof Helmchen1,2,3
1Laboratory of Neural Circuit Dynamics, Brain Research Institute, University of Zurich, Zurich, Switzerland
2Neuroscience Center Zürich, University of Zurich and ETH Zurich, Zurich, Switzerland
3University Research Priority Program (URPP), Adaptive Brain Circuits in Development and Learning (AdaBD), University of Zurich, Zurich, Switzerland
4Computational Neuroscience Group, Department of Physiology, University of Bern, Bern, Switzerland
*Email: tsai@hifo.uzh.ch
Introduction
Learning goal-directed behavior requires the association of salient sensory stimuli with behaviorally relevant outcomes. In the mammalian neocortex, dendrites of pyramidal neurons are suitable association sites, but how their activities adapt during learning remains elusive. Computation-driven theories of cortical function have conjectured that apical dendrites should encode error signals [1,2]. However, little biological evidence has been found to support these proposals. Therefore, we propose a biology-driven approach instead and attempt to explain the function of bottom-up and top-down integration in a model of pyramidal neurons based on experimentally observed apical tuft responses in the sensory cortex during learning.
Methods
We track calcium transients in apical dendrites of layer 5 pyramidal neurons in mouse barrel cortex during texture discrimination learning [3]. Based on this experimental data, we implement a computational model (Fig 1a) incorporating: top-down signals encoding the unsigned temporal difference (TD) error [4], bottom-up signals encoding sensory information, multiplicative gain modulation of firing rates by apical tuft activity, and a local associative plasticity rule comparing top-down signals and somatic firing to dictate apical synapse plasticity. Finally, we test the relevance of apical tuft activity by inhibiting apical tufts during reward and punishment both in our model and experimentally using optogenetics (Fig 1b).
Results
We identify two apical dendrite response types: 1) responses to unexpected outcomes in naïve mice that decrease with growing task proficiency, 2) responses associated with salient sensory stimuli, especially the outcome-predicting texture touch, that strengthen upon learning (Fig 1c). These response types match two distinct unsigned components of the temporal difference error. Our computational model demonstrates how these apical responses can support learning by selectively amplifying the responses of neurons conveying task-relevant sensory signals. This model is contingent upon top-down signals encoding unsigned TD error components, bottom-up signals encoding sensory stimuli, and apical synapses following an associative plasticity rule.
Discussion
Our findings indicate that L5 tuft activities might transmit a salience signal responsible for selectively amplifying neuronal activity during relevant time windows. This picture is in line with theories claiming that the top-down feedback onto apical dendrites is involved in credit assignment. However, instead of transmitting neuron-specific signed errors, our work suggests that the brain could employ a two-step strategy to assign credit to individual neurons. By first solving the temporal credit assignment problem, a temporally precise top-down salience signal can be broadcast to sensory regions, which in a second step — involving local associative plasticity — can be leveraged to recognize and amplify task-relevant responses.
Figure 1. Fig 1. a, Left: Two unsigned TD error components. Middle: Model schematic. Right: State-value estimate and its temporal derivative (signed and unsigned). b, Optogenetic inhibition time during trials (top) and across training (middle). Bottom: Number of trials to reach expert performance in mice and model. c, Calcium imaging (left) and its model (right) across learning for sensory or outcome types.
Acknowledgements
This work was supported by the Swiss National Science Foundation, the European Research Council, the Horizon 2020 European Framework Programme , and the University Research Priority Program (URPP) ‘Adaptive Brain Circuits in Development and Learning’ (AdaBD) of the University of Zurich.
References
1. https://doi.org/10.1016/j.tins.2022.09.007
2. https://doi.org/10.1016/j.tics.2018.12.005
3. https://doi.org/10.1101/2021.12.28.474360
4. https://doi.org/10.1007/BF00115009
