Harvard University COVID-19 updates

Department News

Comparative Computation Across Biological and Artificial Neural Networks [Engert- and Schier Labs]

Comparative Computation Across Biological and Artificial Neural Networks [Engert- and Schier Labs]

In a recent paper in Neuron (PDF) the Engert and Schier labs uncover striking similarities in stimulus representation and computation across biological and artificial neural networks performing temperature gradient navigation.

During evolution adaptive pressure shapes an animal’s behavior and morphology. This can be prominently seen when comparing specialized appendages across divergent species facing the same task such as fore-limbs adapted to digging through earth in moles and mole crickets. At the same time the adaptive pressure of attracting mates across a long distance leads to mate-calling behavior in diverse animal phyla such as mammals, amphibians and insects. The generation of adaptive behaviors is a major task of the brains of animals. It is therefore conceivable that similar to homologies observed in specialized appendages there are computational and circuit homologies across brains controlling similar behaviors.

To accomplish a behavioral goal, brains need to extract relevant information from the environment and represent this information in the form of neuronal activity. Since the nature of the behavioral task likely determines which information is most relevant it might be that as brains evolve to perform a certain task, they also evolve a specific way of representing environmental stimuli. Through brain wide functional imaging we recently identified how larval zebrafish represent temperature stimuli in order to create adaptive behavioral output underlying heat gradient navigation (Haesemeyer et al., 2018). Specifically, cells in the zebrafish hindbrain encode both the current temperature and the direction and rate of change of a heat stimulus. Knowing how a heat stimulus changes allows for a simple form of prediction which might be useful in the context of gradient navigation. We therefore wondered if this stimulus representation can be seen as a stereotypical solution that evolution arrives at because of the goal of navigating a temperature gradient much like the selection of specific limb morphologies that are useful for digging through soil.

One approach to answer the question to what extent a behavioral task dictates how the brain represents the environment, would be to compare brain wide computation and stimulus representation across divergent species. However, unlike studying external limb morphology this is a rather complex task, since obtaining large-scale functional neural data is difficult in many species. Inspired by studies over the past two decades that used artificial neural networks to understand withdrawal reflexes in leech (Lockery et al., 1989), visual processing in cortex (Yamins and DiCarlo, 2016) and other cognitive processes (Rumelhart and McClelland, 1982), we opted for a different strategy. Instead of comparing representations across species we trained an artificial neural network to navigate a temperature gradient using a larval zebrafish behavioral repertoire. A similarity in representation and computation between the two systems would argue that biological evolution and artificial network training arrive at a stereotypical solution within the large space of possible solutions.

After training a multi-layered convolutional network to navigate a naturalistic heat gradient using a zebrafish behavioral repertoire we indeed found clear similarities in computation and temperature representation between the network and larval zebrafish. Furthermore, performing virtual ablations in the artificial network revealed that it critically depends on those units that encode temperature in a fish-like manner, while other units were dispensable for gradient navigation. Interestingly, we found that integration times were also matched between larval zebrafish and the artificial networks, even though no explicit constraints on timing were given during training. This convergence in representational and computational features argues that larval zebrafish through evolution and the artificial neural network through training arrived at a common, stereotypical solution of temperature representation and computation. To test if and how the available motor repertoire influences how a network represents stimuli we trained a network variant that had the same architecture as the zebrafish network but controlled a C. elegans motor repertoire. This change tuned representations and strikingly changed the importance of individual response types for gradient navigation. Not unexpectedly, this argues that computations performed on environmental stimuli depend on the actions that are available to an animal when interacting with the environment.

The uncovered similarities between the biological and artificial neural networks suggest that artificial neural networks with their greater amenability to analysis could serve to generate testable hypotheses about biological circuits. And the network indeed makes strong predictions about the importance of individual cell types in gradient navigation and allowed us to identify a novel heat responsive cell type in the zebrafish brain.

Overall these results underline the utility of artificial networks as both comparative tools to biology as well as hypothesis generators for biological neural networks.

by Martin Haesemeyer


Florian Engert faculty profile, Engert lab website

(l to r) Alex Schier, Martin Haesemeyer, and Florian Engert

(l to r) Alex Schier, Martin Haesemeyer, and Florian Engert