Slowness: an objective for spike-timing-dependent plasticity?

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neurobiology

Background:

• The brain achieves invariant object recognition despite changing conditions.
• This ability is hypothesized to arise from learning temporal stability, where object identities change slower than context.
• Implementing this slowness principle with biologically realistic learning rules is an open question.

Purpose of the Study:

• To investigate how temporal stability or slowness can be implemented using biologically realistic spike-based learning rules.
• To explore the link between algorithms like Slow Feature Analysis and STDP.
• To derive specific STDP learning windows for slowness learning.

Main Methods:

• Implemented Slow Feature Analysis in linear continuous model neurons using a modified Hebbian learning rule.
• Analytically derived conditions for slowness learning in linear Poisson neurons using spike-timing-dependent plasticity (STDP).
• Analyzed STDP learning dynamics, considering the convolution of the learning window with the postsynaptic potential.

Main Results:

• Demonstrated that Slow Feature Analysis can be implemented via a modified Hebbian rule, linking it to the trace rule.
• Showed that STDP with a specific learning window can achieve slowness learning in linear Poisson neurons.
• Derived STDP learning windows compatible with physiological data and identified components sensitive to reversible and irreversible input statistics.

Conclusions:

• Slowness learning can be implemented by STDP with specific learning windows, offering a biologically plausible mechanism for invariant recognition.
• The functional interpretation of STDP depends not only on the learning window but also its convolution with the postsynaptic potential.
• Irreversible input statistics may lead to oscillatory weight dynamics rather than stable distributions, offering new insights into STDP's role.