Neuroscience: Intelligence in the Honeybee Mushroom Body

Abstract

Intelligence, in most people’s conception, involves combining pieces of evidence to reach non-obvious conclusions. A recent theoretical study shows that intelligence-like brain functions can emerge from simple neural circuits, in this case the honeybee mushroom body.

The mushroom body, a conspicuous structure in the brain of insects, was first described by the French biologist Felix Dujardin in 1850 [1]. When examining the brains of various insects, Dujardin observed that social insects, such as honeybees, have a bigger and more complex mushroom body than their less social relatives, solitary bees, for instance. This observation inspired Dujardin to propose that the mushroom body was the seat of intelligence in the insect brain. Dujardin was remarkably insightful: after more than 150 years of research, a large body of evidence demonstrates that the mushroom body is a multisensory brain center required for the formation, storage and retrieval of associative memories [2]. The wealth of anatomical, physiological and behavioral data makes the mushroom body a powerful system for theoretical studies that use mathematical models highly constrained by data. With these models, learning can be replicated and features critical for its function can be identified. In a recent paper in Current Biology, Peng and Chittka [3] use a computer model of the mushroom body to examine its capacity to learn and reveal how that capacity depends on circuit features.

The mushroom body integrates input from multiple sensory systems, but its connections to the olfactory system have been extensively characterized in different insects including honeybees and Drosophila melanogaster [2]. Insects detect odors through sensory neurons covering their antennae. As demonstrated in Drosophila melanogaster, most of these neurons express a single type of olfactory receptor [4]. Neurons expressing the same receptor converge in the antennal lobe, an olfactory relay center analogous to the vertebrate olfactory bulb, where they innervate a single glomerulus. Because most odorant molecules bind to multiple receptors, each odor is represented in the antennal lobe of both honeybees and Drosophila by a combination of active glomeruli. Olfactory information is transferred from the antennal lobe to higher brain centers, including the mushroom body, through projection neurons, each of which carries information from a single glomerulus. In Drosophila, it has been shown that individual mushroom body neurons, the Kenyon cells, receive on average seven inputs from an apparently random set of projection neurons [5]. Such randomization of sensory input is an important feature of the mushroom body, allowing it to construct an informative representation of olfactory stimuli useful for extracting associations.

The Kenyon cells connect to a small number of output neurons that are thought to mediate different learned behaviors, such as attraction and aversion [6]. Studies in locusts and Drosophila show that, when learning to associate a particular odor with reward or punishment, the connections between the Kenyon cells activated by that odor and the output neurons mediating an appropriate behavioral response are modified [7–9]. This experience-dependent modification requires the action of dopaminergic neurons that innervate the compartments where connections between the Kenyon cells and output neurons are made [6]. Anatomical and functional evidence suggests that there might be additional sites where learning takes place, at least in the honeybee mushroom body: the VUMmx1 neuron, an octopaminergic neuron activated by sugar that could mediate plasticity, innervates the input region of the mushroom body where projection neuron to Kenyon cell synapses are located [10]. In their study, Peng and Chittka [3] examine the effect of this additional site of plasticity on the learning performance of the mushroom body.

The model of the mushroom body developed by Peng and Chittka [3] contains 100 projection neurons that connect to 4000 Kenyon cells — numbers smaller than those for the real honeybee mushroom body but large enough to duplicate some of its functions. The model’s Kenyon cells connect to two output neurons, one mediating attraction and the other aversion, again a simplification compared to the multiple attractive and aversive outputs in the actual mushroom body. In the model, each odor activates 50% of the projection neurons and 5% of the Kenyon cells. The sparser activity in the mushroom body is the result of both low convergence of input (each Kenyon cell receives, on average, ten inputs from a random set of projection neurons) and global feedback inhibition. The behavior generated by the model in response to a panel of odors was determined by the preference index, a measure proportional to the difference in the activities of the appetitive and aversive output neurons.

Prior to training, odors activate the two output neurons approximately equally. Thus, the preference index is near zero, and odors trigger neither attraction nor aversion. Associations between a particular odor and a reward are learned through modification of the synapses between the Kenyon cells activated by that odor and the appetitive output neuron. Conversely, if the odor is associated with punishment, synapses to the output neuron mediating aversion are modified. This feature of the model is based on several recent studies that provide evidence for synaptic plasticity between Kenyon cells and output neurons [7–9]. This form of plasticity allows the model to exhibit simple associative learning, but Peng and Chittka [3] were interested in more complex relationships between stimulus and reward or punishment.

One task that Peng and Chittka [3] considered, which they call ‘patterning’, is similar to the exclusive-or problem famous in machine learning. Two odors, A and B, are associated with one valence (either attraction or aversion), while the mixture A–B is associated with the opposite valence. The variety of inputs that the different Kenyon cells receive through their random inputs produce a representation that is well suited to support such a task. Indeed, the model can perform this task solely on the basis of plasticity at Kenyon cell to output neuron synapses.

In another task, odor A is associated with reward and odor B with punishment. When a range of different odors is then tested, the most appetitive odor is A, meaning that the aversion to B does not affect the behavioral appetitive responses to odors similar to A. This result, which is due to the relatively narrow tuning of the trained aversive output neuron, disagrees with results observed in honeybees. Honeybees show a phenomenon called ‘peak shift’ in which the most appetitive odor is not A but an odor similar to A that is more different from B (Figure 1) [11–14]. In other words, the aversion to B affects the attraction to A in the honeybee but not in the model.

Figure 1. Peak-shift in the honeybee mushroom body.

The model of the honeybee mushroom body devised by Peng and Chittka [3] was trained with different learning paradigms. During an absolute learning task, the model learns to associate stimulus A (depicted as a red and blue flower) with a sugar reward. After training, the attraction of the modeled mushroom body peaks for the stimulus it was trained with (in this case, the red and blue flower). During a differential learning task, the model learns the same association as well as to associate stimulus B (depicted as a red flower) with a punishment. After training, the attraction of the modeled mushroom body peaks for a stimulus it was trained with (a blue flower). This phenomenon is called ‘peak-shift’ and is a form of learning that enables the brain to form inferences based on different past experiences.

To fix this discrepancy, Peng and Chittka [3] extended plasticity in the model to the synapses between the input projection neurons and the Kenyon cells. In this version of the model, reward strengthens the connections between active projection neurons and Kenyon cells, whereas punishment weakens them. This plasticity broadens the tuning of the output neurons so that the aversion to B now modifies the preference index in response to A, producing a peak shift. This improvement in the model comes with a cost, however. Performance on the patterning task is worse in the model with extended plasticity than in the original model, suggesting a decrease in discrimination that appears to disagree with the extremely good discrimination ability of bees [15].

The honeybee mushroom body contains two types of Kenyon cell: class I, also called ‘spiny Kenyon cells’, that have wide-field dendritic arbors, and class II or clawed Kenyon cells that have more localized arbors [16]. On the basis of results from the locust [17], Peng and Chittka [3] assumed that class I Kenyon cells receive input from a large number of projection neurons, in contrast to class II, the only Kenyon cells in the original model, which receive few inputs (ten on average). The expanded model indicates that class II Kenyon cells perform better in the patterning task, whereas class I Kenyon cells show a stronger peak-shift and thus show broader ‘generalization’. If type I Kenyon cells are, indeed, extensively connected to projection neurons, this observation provides insight into why two types of neuron are found in the honeybee mushroom body: to strike a balance between generalization (the peak shift) and discrimination (the patterning task). Together, these results imply that plasticity should be present at synapses from projection neuron to class I, but not class II, Kenyon cells [18]. Interestingly, the Drosophila mushroom body contains only sparsely connected Kenyon cells. This suggests that fruit flies may be more limited in their ability to generalize across appetitive and aversive odors.

Over the years, the mushroom body field has generated a large body of experimental data, but the function and significance of certain mushroom-body features have remained unclear. For example, spiny and clawed Kenyon cells were described more than two decades ago [16], but their role in learning is not immediately obvious from their morphology. The study by Peng and Chittka [3] suggests that these two types of neuron carry out different functions: spiny Kenyon cells appear better at generalizing, whereas clawed Kenyon cells may be better at discrimination. Although it remains unclear why not all mushroom bodies contain two types of Kenyon cell with different extremes of connectivity, this may be related to whether or not modulator-regulated plasticity developed at the projection neuron to Kenyon cell synapse. Two-stage learning with extensive input connectivity in type I Kenyon cells may significantly enhance the ability of the mushroom body in bees to support complex inference.