Dendritic Computation

Motivation(s)

Many artificial neuronal networks assume an individual neuron sums up the synaptic input and decides whether to initiate an action potential via comparing the sum against a (nonlinear) threshold. This model ignores the recent findings on how the linear and nonlinear mechanisms in the dendritic tree could serve as computational building blocks.

Proposed Solution(s)

The authors review recent works showing neuronal dendrites may be essential for the computations performed by the neuron and the network.

Evaluation(s)

The authors list promising results such as

  • directional selectivity in visual neurons of flies,

  • coincidence detection in auditory neurons of chicks,

  • temporal integration over long timescales in oculomotor system of goldfishes,

  • image processing (e.g. convolution) in dendrites of flies,

  • lobula giant movement detector of locusts, and

  • forward masking in omega neurons of crickets.

Note that these are still just experiments and only serves as a guide for future research.

Future Direction(s)

  • Would duplicating the weights in a neural network but force the duplicate to decay in a manner similar to reinforcement learning be beneficial?

  • If the topology of a neural network undergoes remodelling (e.g. pruning, expansion, clustering, Hebbian theory) during back propagation, would the error improve?

Question(s)

  • What are some experiments that involve stimulating a living person’s brain?

Analysis

Deep learning with nonlinear activation functions and hidden layers is using the most basic approximation of neurons: each neuron receives hundreds to thousands of input spike trains, yet it has only one axon for signaling a single output. Although the experimental findings are promising, the subjects are still non-human or dead brain cell samples. Nevertheless, the results provide valuable insights into the brain such as the dendritic mechanism that backpropagates action potentials.

Notes

Computations in Passive Dendrites

  • Nonlinear interaction between excitation and shunting inhibition on small dendritic branches can implement logical operations.

  • A high-frequency fluctuating current injected in the dendritic pipette will evoke high-frequency and large-amplitude local voltage responses, but the response recorded by the somatic pipette will be attenuated and smoothed (low pass filtered).

  • Excitatory inputs to the same branch tend to sum sublinearly, whereas inputs on different branches sum linearly.

    • In neurons with active dendrites, clusters of inputs that are synchronously active on the same branch can evoke a local dendritic spike.

  • Coincidence detection in pyramidal neurons is tuned by their dendritic branching pattern.

References

LHausser05

Michael London and Michael Häusser. Dendritic computation. Annu. Rev. Neurosci., 28:503–532, 2005.