Flowing from sense to action. Are neural integrators necessary?

Sensory memory allows us to retain ‘snapshots’ of the world. Without this, our life would lose coherence. From a neurophysiological perspective, this process is thought to be supported by neurons with levels of activity that persist for several seconds, even after the removal of the stimulus of ‘interest’. Imagine you briefly saw a number and you are now trying to remember it just long enough until you rush to register it on your smartphone or notebook. Once you are done, you forget about the number in a matter of seconds. One way to think about sensory or short‐term memory is that during this period of ‘remembering’, groups of neurons ‘hold’ a level of activity that encodes for this information and when it terminates, the memory vanishes. Neural circuits specialised in integrating information in time‐varying signals are of critical importance – not only for memory formation – but for a variety of brain functions, in particular for motor control, navigation and decision making.

A study by Koutsikou et al. (2018) in this issue of The Journal of Physiology elegantly exposes the minimal circuitry necessary for generating such sensory memory while exploring a simple ‘touch‐and‐go’ paradigm in the tadpole. Immediately prior to swimming, a population of cells in the hindbrain called reticulospinal (RS) neurons provide the descending motor commands to drive rhythmogenesis in spinal networks necessary for swimming initiation (see Grillner (2003) for review). For young tadpoles, tactile stimuli are a signal for predation and are naturally aversive. These spinal‐projecting hindbrain RS neurons integrate touch‐triggered ascending sensory inputs, thus forming a basic sensorimotor circuit that when engaged will induce swimming. The computation to be performed is a binary decision: to swim or not to swim.

To analyse the neural dynamics between sensory input and motor output, Koutsikou and colleagues used a preparation that allows for simultaneous recordings from ascending sensory afferents, RS neurons and ventral root activity in the spinal cord. They observed that sensorimotor reaction times were not stereotyped and that long‐duration delays and trial‐to‐trial variability mediated the final response, which was reflected in the activity of hindbrain neurons with persistent excitation for prolonged periods of time. This fact led the authors to hypothesize that they may also belong to a recurrent excitatory network within the hindbrain, reminiscent of an integrator, in order to account for the sustained activity even after the disappearance of the stimulus.

Theoretical studies have tested plausible networks for implementing such an integrator by placing constraints onto connectivity patterns of excitatory and inhibitory recurrent networks and biophysical properties, such as membrane time constants, specific ion‐selective channels and plateau potentials (see e.g. Lim & Goldman, 2013). To predict the inner makings of Koutsikou's hypothetical recurrent network, the authors explored the parameter space for synaptic strengths and connection probabilities by modelling the circuit using intrinsic properties of the hindbrain RS driven by realistic sensory impulses. Yet, too often too little is gained by this approach, particularly when detailed knowledge of neurons is lacking. However, the experimental settings and preparation described in Koutsikou et al. (2018) grants them the necessary accessibility to perform intracellular recordings and to monitor the subthreshold processes that lead up to state transition at the level of the individual neuron. This approach enabled them to overcome the limitations associated with extracellular recordings that detect all‐or‐nothing states, thereby unmasking the presence of accumulating synaptic inputs that promote and prolong the excitability of ‘decision‐making’ RS neurons. Presumably, they can also target interneurons belonging to the putative recurrent network to quantify their dynamic behaviour. Nonetheless, this particular sensorimotor circuit and function deserves further attention and a closer look at other architectures driving persistent activity, which might help inspire experimental design to uncover the cellular and circuit basis for sensory memory.

Here, an obvious explanation for requiring sensory memory within the hindbrain circuit is to provide a broader window of opportunity for integration of other signals descending from other sensory systems or ongoing motor commands to weigh in and exert their influence by enhancing or suppressing the final decision. Undoubtedly, neural networks reliant on reverbatory architectures could serve a multitude of purposes. Neurons displaying persistent or prolonged activity that are triggered by extrinsic inputs have also been observed in many other brain areas (see Major & Tank, 2004 for review), signifying their broader impact on other aspects of neural function. Notably, the need for mathematical integration in the sense of Newtonian calculus was conceived in the oculomotor field in order to account for the sustained activity in extraocular motoneurons after the end of the saccadic movement, which maintains eye position at eccentric positions (see Robinson (1989) for review). For instance, hindbrain neurons found in the nucleus prepositus are thought to transform eye‐head gaze displacement phasic signals transmitted from the superior colliculus and vestibular nucleus by ‘integrating them’ into tonic signals with a firing frequency that is position dependent. Based on this notion, one can presume that fundamental sensorimotor behaviours, such as during visual exploration, may also rely on the short‐term storage of incoming visual information for forming a uniform and stable understanding of our surrounding environment. It is tempting to speculate that a generic neural network defined by units with specific intrinsic properties would form a common architecture for generating persistent activity in different brain areas used for different functions across different species.