Opponent processes in visual memories: A model of attraction and repulsion in navigating insects’ mushroom bodies

Abstract

Solitary foraging insects display stunning navigational behaviours in visually complex natural environments. Current literature assumes that these insects are mostly driven by attractive visual memories, which are learnt when the insect's gaze is precisely oriented toward the goal direction, typically along its familiar route or towards its nest. That way, an insect could return home by simply moving in the direction that appears most familiar. Here we show using virtual reconstructions of natural environments that this principle suffers from fundamental drawbacks, notably, a given view of the world does not provide information about whether the agent should turn or not to reach its goal. We propose a simple model where the agent continuously compares its current view with both goal and anti-goal visual memories, which are treated as attractive and repulsive respectively. We show that this strategy effectively results in an opponent process, albeit not at the perceptual level-such as those proposed for colour vision or polarisation detection-but at the level of the environmental space. This opponent process results in a signal that strongly correlates with the angular error of the current body orientation so that a single view of the world now suffices to indicate whether the agent should turn or not. By incorporating this principle into a simple agent navigating in reconstructed natural environments, we show that it overcomes the usual shortcomings and produces a step-increase in navigation effectiveness and robustness. Our findings provide a functional explanation to recent behavioural observations in ants and why and how so-called aversive and appetitive memories must be combined. We propose a likely neural implementation based on insects' mushroom bodies' circuitry that produces behavioural and neural predictions contrasting with previous models.

A lateralised design for the interaction of visual memories and heading representations in navigating ants

“…Views facing the nest may as well be included during learning and categorised as left, right or both, explaining why most ants facing their goal usually choose to turn in one particular Revisiting current questions in insect and robot navigation such as early exploration, route following and homing 20,[46][47][48][49] ; the integration of aversive memories 8,24,50 , path integration and views ( [51][52][53][54] or other sensory modalities ( 55-58 as well as seeking for underlying neural correlates [5][6][7] -with such a lateralised design as a framework promises an interesting research agenda.…”

“…Previous studies assumed that ants memorise views while facing the goal [20][21][22] and anti-goal [23][24][25] ) directions, and that they must consequently align their body in these same directions to recognise a learnt view as familiar [26][27][28] . On the contrary, our modelling effort suggests that ants should rather recognise views based on whether the route direction stands on their 'left or right' rather than 'in front or behind'.…”

Visually guided homing of bumblebees in ambiguous situations: A behavioural and modelling study

“…Numerous variants of homing models have been described and tested in the last decades, be they using frequency-based [41], rotation invariant [26], brightness [30], skyline [27], or optic-flow representations [42], single snapshot [30], multi-snapshots [36,43], or attractive and repulsive views [38]. We sought first for parsimonious models to predict the bumblebees' search location in visual conflict situations.…”

“…From all listed models, only the model proposed by Le Möel et al 2020 [38], uses several snapshots to guide the agent home. Hence, the other mentioned models [26,27,37,41] are likely to suffer from object occlusion as the B1 or CwN1 model do.…”

The Antarium: A Reconstructed Visual Reality Device for Ant Navigation Research

“…Equally, investigations of the real-life computational properties of navigation-relevant neural circuits are currently hampered by limitations in the way visual information can be presented in electrophysiology rigs (see e.g., Table 1 ). There are currently no projection devices that can convey the full information content of the spatial, spectral, and polarization signal patterns that characterize natural navigation environments; and lastly the navigational competence of insects is based on active learning processes (e.g., Collett and Zeil, 2018 ; Jayatilaka et al, 2018 ; Zeil and Fleischmann, 2019 ) and relies on the active comparison between remembered and currently experienced input patterns (e.g., Zeil, 2012 ; Le Möel and Wystrach, 2020 ; Murray et al, 2020 ). It is thus likely that the neural machinery underlying navigation is heavily state-, context- and activity-dependent, requiring closed-loop control of the visual scene by the insect and control by the experimenter over the experience (What has been learned?…”