Prospective Coding by Spiking Neurons.

Area of Science:

• Computational Neuroscience
• Learning and Memory
• Synaptic Plasticity

Background:

• Animals demonstrate predictive learning, but the underlying neuronal mechanisms and plasticity rules remain unclear.
• Understanding how neurons form associations and predict future events is fundamental to neuroscience.

Purpose of the Study:

• To propose a biologically plausible synaptic plasticity rule for learning predictions at the single-neuron level.
• To investigate how neurons can match their firing rate to expected future discounted firing rates.
• To explore the role of this rule in prospective coding and classical conditioning.

Main Methods:

• Developed a novel learning rule for a spiking two-compartment neuron model.
• Utilized spike-timing-dependent plasticity (STDP) with a narrow plasticity window (20 milliseconds).
• Demonstrated the rule's efficacy in predicting time-varying inputs, paired-associate tasks, and sequence reproduction with recurrent networks.

Main Results:

• The proposed plasticity rule enables neurons to learn associations and make predictions on timescales of seconds.
• The neuron learns to predict future discounted reward, closely mirroring temporal difference learning (TD(λ)) in reward prediction tasks.
• Prospective coding was successfully illustrated through various prediction tasks.

Conclusions:

• A single-neuron plasticity rule can support learning predictive associations over extended timescales.
• This mechanism offers a potential explanation for prospective coding and classical trace conditioning.
• The findings link synaptic plasticity to reinforcement learning principles.