By Anne Urai and Peter Murphy.

In our lab journal club before the winter break, we read Wimmer et al. (2015). Sensory integration dynamics in a hierarchical network explains choice probabilities in cortical area MT. Nature Communications, 6, 6177.

The phenomenon of choice probability was first described by Britten and colleagues (1996), who measured firing rates of neurons in a region of visual cortex known as area MT, that selectively responds to visual motion. In this influential study, when a macaque monkey looked at a cloud of moving dots in an effort to determine its dominant direction of motion, the firing rates of single MT neurons could be used to predict the decision that the monkey would eventually make. Interestingly, this was the case even when identical stimuli were presented multiple times. Choice probability, which quantifies how well the monkey’s choices can be predicted from such neural activity alone, thus reflects a decision-maker’s variable, subjective judgments about sensory information in the outside world.

How can neurons at early stages of the visual processing hierarchy ‘know’ about the final decision the monkey will make? Visual neurons like those in area MT are activated by sensory input in a variable, inherently stochastic manner, while higher-level areas involved in the decision-making process ‘read-out’ this noisy information from lower levels of the hierarchy to form a choice. In this bottom-up view, choice probability arises because fluctuations in the firing of MT neurons contribute to the decision that’s ultimately made. If neurons in a population in MT fluctuate randomly, but together over trials (a commonly observed phenomenon termed ‘noise correlation’, that likely has multiple origins1), this activation would affect downstream decision neurons and, in turn, influence choice.

However, we know that the brain is not a simple feedforward system and that low-level visual neurons receive dense feedback projections from upstream regions involved in decision-making. Another explanation of choice probability is thus that the dynamics of decision formation that play out in high-level association cortex (e.g. in lateral intraparietal cortex, where firing rates are thought to closely track the monkey’s unfolding decision) also affect activation patterns in visual cortex through recurrent feedback loops. In this top-down scenario, choice probability in areas like MT does not reflect a causal influence on the decision, but rather results from the decision process that takes place further up the cortical hierarchy.

Consistent with such an account, an important study (Nienborg & Cumming, 2009) found that in macaque visual cortex (V2), the time course of choice probability gradually increased over time. On the other hand, the so-called ‘psychophysical kernel’, showing when fluctuations in the stimulus influence behaviour the most, was observed to peak early and decrease as the trial unfolded. Thus although neural activity towards the end of the trial predicted the upcoming choice, this was not due to information in the stimulus at that time in the trial. This dissociation indicates that choice probability reflects more than a bottom-up, causal influence of sensory information on behaviour.

To quantitatively tease apart bottom-up and top-down influences on choice probability, Wimmer and colleagues took an existing neurobiologically principled model of the decision-making process (Wang, 2002) that replicates essential features of neuronal spiking data during a motion discrimination task2. Wimmer et al extended this model by incorporating a layer of sensory neurons, representing area MT, that receives feedback connections from the decision layer (area LIP). Empirically, the data on which the model is based show a pattern of sustained choice probability in MT neurons throughout the duration of the task trial.

Graphical representation of the model, which contains a layer of sensory neurons (with red and blue denoting populations selective for choice ‘left’ and ‘right’, respectively) and a layer of integrator or decision neurons. Importantly, the strength of the feedback from decision to sensory regions (in green), which is mediated by fast AMPA receptors in the model, can be systematically varied to assess the effect of top-down signals on choice probability and behaviour.

By simulating neural and behavioural data from a variety of models in which the relative dominance of bottom-up and top-down components was systematically varied, Wimmer et al. found that this sustained choice probability likely arises from a combination of both factors. Early in the trial, bottom-up fluctuations in the firing of MT neurons exert a causal influence on choice and lead to a fast-rising choice probability. Later in the trial, top-down influences take over.

A combination of early bottom-up and late top-down influences leads to a sustained choice probability signal, as observed in neural recordings2.
A combination of early bottom-up and late top-down influences leads to a sustained choice probability signal, as observed in neural recordings2.

The model was also leveraged to generate predictions about the ways in which bottom-up and top-down factors should affect both the temporal stability of choice probabilities and the structure of correlations between pairs of MT neurons. The authors verified that, remarkably, each of the predicted patterns is present in the classic electrophysiological recordings2. Moreover, they corroborated aspects of these results using ‘reverse correlation’ analyses of newly collected empirical data to estimate psychophysical kernels which, as mentioned previously, pinpoint the times at which variation in the stimulus influences the final decision most.

Using these psychophysical kernels, one important empirical observation that Wimmer et al. replicated is that monkeys tend to commit to their decisions relatively early in the trial, and disregard subsequent sensory evidence once they have done so. This ‘primacy effect’ has been previously observed in the behaviour of monkeys (Kiani 2008; Nienborg & Cumming, 2009) and in some human studies (Tsetsos et al., 2012; Zylberberg et al., 2012). Importantly, it is a natural consequence of a neural network architecture that, like the original model that Wimmer et al. extend, encourages competitive, winner-take-all ‘attractor’ dynamics. By incorporating feedback connections into this model, Wimmer et al. neatly reconcile this primacy effect with the choice probabilities observed in empirical data. These, as mentioned above, are sustained throughout the trial duration and hence extend well beyond the time at which the stimulus actually influences behaviour.

Beyond this elegant explanation of choice probability, incorporating feedback connections into the model also allowed the authors to make several novel, exciting observations. For example, they detail how top-down feedback connections first serve to increase the rate at which the competition between alternative choices is resolved, and subsequently reinforce the initially winning choice at the expense of a more protracted and accurate decision process3. As they highlight, this feedback-led acceleration of decision dynamics may point to a novel role for feedback connections in what is known as the speed-accuracy tradeoff. Adjusting one’s emphasis on speed versus accuracy is a key requirement for adaptive decision-making in different contexts. Various candidate mechanisms for the regulation of this tradeoff have been proposed in the literature – from a simple lowering of the threshold amount of information required for commitment, to a change in the gain of neural processing. In this model, increasing the strength of top-down feedback offers a biologically plausible mechanism for instantiating a decision-making regime in which early input exerts a particularly strong influence on behavior – leading to quick choice commitment. This feature of the model leads to the novel, and testable, prediction that individuals with stronger feedback connections should exhibit stronger primacy and a greater emphasis on speed over accuracy.

Commitment to a choice corresponds to a well in the energy-landscape of the decision process. A well represents a stable state which is maintained unless a lot of additional input is added to a system to shift it out of that state.
Commitment to a choice corresponds to a well in the energy-landscape of the decision process. A well represents a stable state which is maintained unless a lot of additional input is added to a system to shift it out of that state.

Interestingly, it turns out humans can be rather good at relatively sustained integration, and are even able to adaptively adjust the time window over which they integrate information (Ossmy et al, 2013). It would be informative to probe in which regimes the model by Wimmer et al. displays such longer integration time constants, and to what extent this model can account for differences in behavioural strategies between species and individuals.

The feedback-accelerated choice resolution implicates that feedback connections might also influence decision confidence: when feedback is strong, the difference in ‘neural evidence’ between choice alternatives tends to be high, which in turn equates to higher confidence in the eventual choice. Coupled with the possible role of feedback connections in the emphasis of speed over accuracy, this reasoning suggests that, in some settings at least, faster and less accurate decision-making should be accompanied by greater confidence. This is surprising, since accuracy and confidence are almost always observed to be positively correlated. It will be interesting to see whether future studies that measure or manipulate the strength of top-down feedback connections can find support for this prediction.

Lastly, our lab is excited about the scope that this kind of model offers for bridging across different levels of analysis, revealing the precise mechanistic significance of physiological signals that are easily measurable in human subjects. Such models are particularly appealing in this regard because they employ biophysically realistic principles to generate actual neuronal spiking data, which can then be averaged within and across populations of neurons to derive predictions about what might be observed at coarser spatial scales. For example, it should be possible to calculate measures of the interaction or coherence between distinct neuronal populations in the model and investigate equivalent signals using human scalp electrophysiology (M/EEG), a field in which the analysis of inter-areal coherence or information transfer is commonplace. In the specific case of the Wimmer et al. model, one might derive a signature of top-down processing that can then be specifically assessed in humans. Such an approach holds obvious promise for greatly augmenting our understanding of the neural dynamics underlying human decision-making, in both health and disease.

Overall, this exciting study presents convincing model-based and empirical evidence for distinct contributions of bottom-up and top-down effects on choice probability, and provides a very interesting starting point for thinking about more flexible decision-making in different behavioural regimes.


1. One important implication of this view is that the magnitude of the choice probability should be closely related to the amount of correlated noise among sensory neurons.
2. These classical data from the Newsome lab are freely available at
3. The extreme case of a more deliberative decision process is  ‘perfect integration’, where all samples of evidence presented throughout the trial are weighted equally.

Further reading

Modelling choice probability dynamics in visual decision-making

Leave a Reply