TL;DR
This paper proposes that economic choice in the brain might be more distributed than traditionally modeled. Using a simple neural network trained on a binary choice task, they demonstrate that value, spatial information, and action selection are not segregated into distinct modules but are instead broadly encoded across the network, suggesting a distributed computation of choice.
What Problem Does This Paper Address?
The field of neuroeconomics has largely focused on modular models of economic choice, where different stages like evaluation, comparison, and selection are handled by distinct brain regions or neuronal populations. These models often assume a "common currency" for value representation. However, the principles governing how distributed systems, like the brain, might make economic choices have been less explored. This paper aims to delineate some of these distributed principles by using a neural network as a model system and re-examining existing neurophysiological data through this lens.
Methods
The authors employed a feed-forward neural network with three hidden layers to simulate economic choice. The network was trained to perform binary risky choices between two gambles, each defined by a reward magnitude and probability. The training utilized a backpropagation algorithm to minimize the difference between the network's chosen activity and the obtained reward. To analyze the network's internal representations, they examined neuronal activity across layers, looking at how task variables (reward, probability, spatial position, choice) were encoded. They also performed lesion studies by nullifying a proportion of neurons in specific layers to observe the effects on performance and analyzed correlations between neuronal responses to understand how information is integrated and compared.
Datasets
Not discussed in the paper. The paper describes a simulated task where gambles are generated randomly with specified distributions for reward magnitude and probability.
Key Insights and Ideas
- No Pure Value Signals: The network did not develop neurons that exclusively encoded "pure value." Instead, value information was always accompanied by other information, such as spatial location.
- Distributed Comparison: Value comparison wasn't a distinct process occurring in a specific module. Instead, units encoded values of both offers in an anti-correlated format, contributing to comparison across the network.
- Continuous Transformation, Not Discrete Modules: Network layers represented stages in a continuous transformation from input to output space, rather than functionally distinct modules like "evaluation" or "selection."
- Spatial Information is Integral: Spatial information (left vs. right offer) was encoded throughout the network and was not segregated from value information, suggesting that space plays a role in the distributed computation of choice.
- Graceful Degradation: Lesioning parts of the network led to a gradual decline in performance, rather than catastrophic failure, a hallmark of distributed systems.
- Challenging the "Common Currency" Hypothesis: The findings suggest that a dedicated, universally recognized "common currency" for value might not be necessary for effective economic choice, especially in distributed systems. Value can emerge from the collective activity of neurons.
Key Results
- The trained neural network achieved an accuracy of 85% in choosing the gamble with the higher expected value, comparable to animal behavior.
- Over 80% of neurons in the network coded for multiple task variables (reward, probability, spatial position, choice), demonstrating broad, distributed coding.
- Encoding of task variables generally increased in abstraction and prevalence in deeper layers of the network.
- The anti-correlation between the value tuning for the two offers increased with layer depth, suggesting that comparison becomes more refined in later stages, but the process is distributed throughout.
- Lesion studies showed that performance degraded gracefully, with smaller lesions having less impact than larger ones, and the difficulty of the choice influenced the impact of lesions.
Limitations and Future Work
- Model Simplicity: The authors acknowledge that the neural network is a simplified model and lacks many features of the real brain, such as plasticity, feedback loops, and oscillatory signals.
- Task Specificity: The current model is designed for a single task. They suggest that more flexible networks designed for a wider variety of tasks might develop more modularity.
- Linear Correlation Analysis: The reliance on linear correlation for analysis might miss non-linear relationships. However, they argue that the core finding of distributed coding holds regardless of linearity.
- Future Work: The authors propose that these principles should serve as hypotheses for future experimental testing. They suggest that more biologically realistic models and advanced analytical techniques are needed to further investigate distributed decision-making in the brain. They also highlight the need to reconcile findings from different experimental methods (lesion studies, neuroimaging, unit physiology) to gain a more complete picture of brain organization.
Summary updated