TL;DR
This study challenges the traditional view of sensory cortices as passive conduits of information. By recording neural activity in mice performing a decision-making task, researchers found that even early visual areas exhibit significant amplitude modulations in sensory responses, influenced by cognitive factors like accumulated evidence and upcoming choices. This suggests that sensory processing is dynamically shaped by internal states, impacting how we make decisions.
What Problem Does This Paper Address?
Perceptual decision-making models often assume that sensory cortices provide raw, unbiased sensory input to downstream decision-making circuits. However, a growing body of evidence suggests that sensory processing can be modulated by various internal and external factors. This paper investigates how sensory information is represented in the brain during a task that requires accumulating evidence over time to make a decision. Specifically, it asks whether sensory inputs, even from early visual areas, are purely veridical or if they are already shaped by cognitive factors like the accumulated evidence and the impending choice. Understanding this modulation is crucial for building more accurate models of how the brain makes decisions.
Methods
The researchers employed a combination of advanced techniques:
- Cellular-resolution two-photon calcium imaging: This allowed them to record the activity of individual neurons in multiple posterior cortical regions of mice.
- Virtual Reality (VR) behavioral paradigm: Mice were trained on a head-fixed VR T-maze task where they navigated a corridor and received random visual cues (evidence) on either side. Their goal was to accumulate these cues to decide which arm of the maze to turn down for a reward.
- Impulse Response Model: To quantify sensory responses to individual cues, they modeled neuronal activity as a time series of amplitudes convolved with a parametric impulse response function. This allowed them to isolate and measure the response to each cue.
- Neural Population Decoding: Support Vector Machines (SVMs) were used to decode various task-related variables (view angle, running speed, accumulated evidence, choice, etc.) from the amplitudes of cue-locked neuronal responses.
- Model Comparison: They developed and compared several models (null hypothesis, SSA, choice, cue-counts) to explain the observed amplitude modulations, using AICc likelihood weighting to determine the best fit for each neuron.
Datasets
The study utilized data collected from 11 mice trained on the Accumulating-Towers task. These mice were from transgenic lines expressing calcium-sensitive fluorescent indicators (GCaMP6f). No new benchmarks were introduced; the focus was on analyzing neural data from a well-defined behavioral task.
Key Insights and Ideas
- Sensory cortices are not passive: Contrary to the traditional view, even early visual areas (like V1) show significant modulation of sensory responses, not just noise.
- Cognitive factors shape sensory input: The amplitude of responses to visual cues was not solely determined by the cue itself but was systematically modulated by the accumulated evidence and the animal's impending choice.
- Choice-related modulations are widespread: A large fraction of cue-locked neurons across posterior cortex exhibited amplitude modulations related to the upcoming choice.
- Opposing modulations aid signal clarity: The presence of both positively and negatively choice-modulated neurons might help to disentangle sensory evidence from choice signals, preventing interference.
- Evidence for feedback: The observed modulations, particularly their temporal dynamics and dependence on choice, suggest feedback mechanisms from higher-level decision-making areas influencing sensory processing.
Key Results
- Only a small fraction (5-15%) of active neurons in visual areas and ~5% in RSC showed responses time-locked to sensory cues ("cue-locked cells").
- Across posterior cortex, a significant portion of cue-locked neurons (around 50%) showed response amplitudes that depended on the animal's choice or the accumulated evidence.
- Various task variables, including view angle, running speed, accumulated evidence, and choice, were decodable from the amplitudes of cue-locked responses.
- Over 85% of cue-locked cells exhibited some form of amplitude modulation beyond basic visual properties, with a substantial portion favoring models related to choice or cue-counts over stimulus-specific adaptation.
- Choice modulations of cue-locked responses increased over time in the trial, contrasting with the decreasing behavioral influence of sensory evidence, suggesting a role for internal decision dynamics.
Limitations and Future Work
- Correlated noise vs. signal: The study focused on signal responses and did not attempt to report features like noise correlations that require a fully correct model of signal responses.
- Interspecies differences: While findings bear some similarity to non-human primate studies, there are notable differences, potentially due to task variations or methodological approaches. The authors speculate that these differences might arise from how abstract relationships between cues and choices are represented.
- Future directions: The authors suggest that future models of behavior should incorporate these task-related modulations in sensory processing. They also propose that the observed phenomena, particularly the mixed positive and negative choice modulations, might be a general coding principle for non-sensory information that allows for non-destructive multiplexing of information within neuronal populations. Further investigation into the specific feedback pathways and their role in shaping sensory representations is warranted.
Summary updated