Choice, difficulty, and confidence in the brain
Research highlights
► Confidence is an emergent property of a decision-making attractor neuronal network. ► Decision easiness, to which confidence is related, produces higher firing rates and shorter decision times in the network. ► Decision easiness is predicted to produce a larger BOLD in decision-making networks in the brain. ► The BOLD signal in the medial prefrontal cortex area 10, implicated in decision-making between rewards, increases linearly with task easiness and therefore confidence.
Introduction
Decision-making, and our confidence in the decisions we make, are important areas in neuroscience in the new field of neuroeconomics (Behrens et al., 2007, Deco and Rolls, 2006, Glimcher, 2003, Hampton and O'Doherty, 2007, Heekeren et al., 2004, Kepecs et al., 2008, Kiani and Shadlen, 2009, Kim and Shadlen, 1999, Knutson et al., 2007, Marsh et al., 2007, Rolls, 2008, Romo et al., 2004, Shadlen and Newsome, 1996, Shadlen and Newsome, 2001, Sugrue et al., 2005). Critical issues are to understand the neural processes that underlie decision-making and decision confidence, and to establish the bases for neural signatures of these processes. Primate single neuron recordings (Kim and Shadlen, 1999) show that neuronal responses in a motion decision-making task occur earlier on easy vs. difficult trials in a decision-related brain region, the dorsolateral prefrontal cortex. In the human dorsolateral prefrontal cortex, higher fMRI BOLD signals can be observed on easy trials vs. difficult trials in a similar decision-making task, and this has been proposed as a signature of decision-making (Heekeren et al., 2004, Heekeren et al., 2008). However, exactly what neuronal mechanisms might underlie such decision-making, and whether this is a reliable signature, are not yet clear. The accumulator, counter, or race models of decision-making in which the evidence for different choices accumulates (Carpenter and Williams, 1995, Ratcliff et al., 1999, Smith and Ratcliff, 2004, Usher and McClelland, 2001, Vickers, 1979, Vickers and Packer, 1982) do not describe a neurobiological mechanism, and do not make specific predictions about how neuronal firing rates, synaptic currents, or BOLD signals will alter as decisions become easier and decision confidence increases.
Here we capitalize on recent advances in theoretical understanding of how choice decisions are made using an integrate-and-fire attractor network that makes probabilistic decisions from the spontaneous firing state into one of two or more high firing rate stable attractor states each implemented by a set of coupled neurons that receives the inputs for one of the decisions, and where the choice made is probabilistic because of the noise contributed to by the random spiking times of the neurons for a given firing rate (Deco and Rolls, 2006, Deco et al., 2009, Rolls, 2008, Rolls and Deco, 2010, Wang, 2002, Wang, 2008). Such attractor networks are implemented by excitatory connections between cortical pyramidal cells, and provide a neural architecture not only for decision-making but also for short-term memory (Goldman-Rakic, 1995, Rolls, 2008) and memory recall (Rolls, 2008). We show that the model predicts higher firing rates, synaptic currents, and fMRI BOLD (functional magnetic resonance neuroimaging blood oxygenation level dependent) signals on easy trials vs. difficult trials, and in fact that these increase monotonically (approximately linearly) with ΔI, the difference between the stimuli. Further, it is well established that subjective decision confidence increases with discriminability, ΔI (Jonsson et al., 2005, Vickers, 1979, Vickers and Packer, 1982), and thus the degree of confidence in a decision emerges from the model as being reflected by the firing rates of the neurons involved in the decision-making. Consistently, in rats too the probability that a trial will be aborted reflecting low decision confidence also increases with ΔI on error trials (Kepecs et al., 2008). Moreover, neuronal data consistent with the predictions of the model about decision confidence have been recorded by Kiani and Shadlen (2009) in the posterior parietal cortex during a perceptual decision-making task, and neurons in the rat orbitofrontal cortex that respond on error trials with firing rates that increase with ΔI, that is with confidence in the error (Kepecs et al., 2008), may also be consistent (Insabato et al., 2010, Rolls and Deco, 2010). We test these predictions of the attractor network theory of decision-making in two fMRI investigations of decision-making about the reward value and subjective pleasantness of thermal and olfactory stimuli.
Section snippets
Modelling investigations
The theoretical framework of the model used here was introduced by Wang (2002) and developed further (Deco and Rolls, 2006, Deco et al., 2009, Deco et al., 2007, Marti et al., 2008, Rolls and Deco, 2010, Wang, 2008), and the results described here apply generically to integrate-and-fire attractor network models of decision-making. In this framework, we model probabilistic decision-making by a network of interacting neurons organized into a discrete set of populations, as depicted in Fig. 1.
An attractor network model of decision-making and confidence
The attractor network model of decision-making analyzed is shown in Fig. 1. The implementation of the integrate-and-fire model with AMPA, NMDA, and GABA dynamic synapses is described in Methods, fMRI signals linearly related to the easiness of decisions, and to decision confidence.
Fig. 2a and e shows the mean firing rates of the two neuronal populations D1 and D2 for two trial types, easy trials (ΔI = 160 Hz) and difficult trials (ΔI = 0) (where ΔI is the difference in spikes/s summed across all
Discussion
The attractor neuronal network model of decision-making described here provides a firm foundation for understanding fMRI BOLD signals related to decision-making. The model provides two reasons for the fMRI signal being larger on easy than on difficult trials during decision-making. The first is that the neurons have somewhat higher firing rates on easy trials (e.g. those with a large ΔI) than on difficult trials, as shown in Fig. 3b. The reason for the faster response is that the difference in
Acknowledgments
F.G. was supported by the Gottlieb-Daimler- and Karl Benz-Foundation and by the Oxford Centre for Computational Neuroscience. G.D. received support from the McDonnell Centre for Cognitive Neuroscience at Oxford University. The fMRI investigation was performed at the Centre for Functional Magnetic Resonance Imaging of the Brain (FMRIB) at Oxford University, and we thank Peter Hobden, Siri Leknes, Katie Warnaby, and Irene Tracey for their help.
References (57)
- et al.
Probabilistic population codes for Bayesian decision making
Neuron
(2008) - et al.
Attention, short-term memory, and action selection: a unifying theory
Prog. Neurobiol.
(2005) - et al.
Stochastic dynamics as a principle of brain function
Prog. Neurobiol.
(2009) - et al.
Classical and Bayesian inference in neuroimaging: applications
Neuroimage
(2002) - et al.
Similarity effect and optimal control of multiple-choice decision making
Neuron
(2008) - et al.
Thresholding of statistical maps in functional neuroimaging using the false discovery rate
Neuroimage
(2002) - et al.
Banburismus and the brain: decoding the relationship between sensory stimuli, decisions, and reward
Neuron
(2002) Cellular basis of working memory
Neuron
(1995)- et al.
Robust smoothness estimation in statistical parametric maps using standardized residuals from the general linear model
NeuroImage
(1999) - et al.
Neural predictors of purchases
Neuron
(2007)