From a single decision to a multi-step algorithm
Highlights
► We discuss how the brain implements multi-step algorithms. ► We examine how they might be realized in model neural networks. ► Mental algorithms may consist in serial chains of stochastic evidence accumulation. ► Dual-task paradigms point to a serial organization of decision processes. ► Unlike a digital computer, the brain implements an approximate form of seriality.
Section snippets
Introduction: Turing's hunches and rational machines
The mathematician Alan Turing made at least two seminal contributions to computing. First, to help break the German Enigma code, Turing designed a sequential decision algorithm that consists in accumulating votes for the different options and using a threshold criterion to commit to a choice. This framework has become ubiquitous in the psychology and neuroscience of decision making [1, 2]. Second, Turing formalized a sequential computing device, the famous ‘Turing machine’. Although Turing
From isolated decisions to mental programs: the brain's Turing machine
We have proposed a framework for the neural basis of serial computations that puts together Turing's two seminal contributions [4, 5•]. Our framework (Figure 1) assumes that serial tasks (e.g. computing 23 × 14) are assembled as sequences of elementary decision steps, each involving a parallel competition between a subset of ‘productions rules’, implemented by pools of neurons that accumulate relevant evidence. When one of the competing pools encoding a production reaches a threshold, the race
Two theories on how the brain implements multiple decisions
Over the past years, Cisek, Kalaska, Shadlen and others have argued in favor of a parallel decision framework, and questioning the role of an executive system coordinating decision making. Their argument rests on the following observations. First, during a decision process, the same neuron may, at different times, encode stimulus properties, a decision signal, or a motor response [8, 9, 10•]. This argues in favor of distributed instead of centralized organization of decision variables. Second,
Evidence for serial operations in the human brain
A classical psychological observation, the psychological refractory period (PRP), illustrates the limits that human brain architecture imposes on parallel processing. When a human subject is asked to process two near-simultaneous stimuli as fast as possible, a striking seriality emerges: one stimulus is processed with no trace of interference, as if it was presented in isolation, but the response to the second is massively delayed, often by hundreds of milliseconds [17, 18•, 19, 20, 21]. This
Seriality and conscious control
Psychologists have long speculated that serial effortful performance is associated with the deployment of strategies that are consciously controllable and reportable [44, 45]. Indeed, the parieto-prefrontal network that is tied during dual tasking overlaps strongly with the ‘global neuronal workspace’ postulated to underlie conscious reportability [46•]. Furthermore, the mechanisms that lead to the postponement of the second task during the PRP are very similar to those that lead to the loss of
Chaining: a missing link
Of particular importance for building a mental program is the capacity to ‘chain’ operations, such that the output of the first one becomes the input to the second. Chaining is an understudied yet essential operation. It implies a flexible neural architecture for ‘routing’ information across any two brain processors, which is likely to involve prefrontal cortex [40, 41, 60, 61].
Using behavioral response-time measurements, Fan et al. [62] demonstrated that in humans, even for identical tasks,
Approximate seriality in a parallel brain
Unlike a digital computer, the brain only implements an approximate form of seriality. Not only do a variety of automatized perceptual and motor operations operate in parallel, but even central decisions may be occasionally subject to partial time sharing [72]. During the PRP, the response to the second item may interfere with the first, suggesting the presence of ‘cross-talk’ [22, 23]. During curve tracing, the successive stages of amplification may partially overlap in visual cortex [69••].
Where is the program? Neural mechanisms for coordinating multiple operations
The precise mechanism by which the executive system organizes serial decisions remains unknown. We emphasize that we do not advocate a homunculus or any dedicated localized structure that will perform such a role. As a first step to understand how executive control is deployed over many parallel cortical processors, we modeled how a network of spiking neurons could implement a sequence of two independent tasks (Figure 4) [5•]. In the model, seriality in dual-task performance results as a
Concluding remarks
How the architecture of the brain supports multi-step operations remains largely unexplored. We conclude by pointing to two prominent challenges for further research.
First, dissecting the electrophysiological mechanisms of seriality is hindered by the intrinsic difficulty of training non-human primates in complex multi-step tasks. Indeed, such tasks tap rostral and mesial prefrontal cortical systems [77, 84•] that are more expanded and connected in the human brain than in other primates [85, 86
References and recommended reading
Papers of particular interest, published within the period of review, have been highlighted as:
• of special interest
•• of outstanding interest
Acknowledgements
Supported by INSERM, CEA, Collège de France, and an ERC grant ‘NeuroConsc’ to S.D. The authors are very grateful to Ariel Zylberberg and Mike Shadlen for stimulating discussions.
References (94)
- et al.
Banburismus and the brain: decoding the relationship between sensory stimuli, decisions, and reward
Neuron
(2002) - et al.
Probabilistic population codes for Bayesian decision making
Neuron
(2008) - et al.
The human Turing machine: a neural framework for mental programs
Trends Cogn Sci
(2011) - et al.
Neurobiology of decision making: an intentional framework
- et al.
Neural correlates of reaching decisions in dorsal premotor cortex: specification of multiple direction choices and final selection of action
Neuron
(2005) Discrete versus continuous stage models of human information processing: in search of partial output
J Exp Psychol Hum Percept Perform
(1982)Dual-task interference in simple tasks: data and theory
Psychol Bull
(1994)- et al.
Dual-task performance with ideomotor-compatible tasks: is the central processing bottleneck intact, bypassed, or shifted in locus?
J Exp Psychol Hum Percept Perform
(2005) - et al.
An EZ-diffusion model for response time and accuracy
Psychon Bull Rev
(2007) - et al.
Central processing overlap modulates P3 latency
Exp Brain Res
(2005)