Stability through plasticity: Finding robust memories through representational drift.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Memories are stored in synapses, with learning altering synaptic weights.
• Altered synaptic weights can interfere with existing memories, creating a plasticity-stability trade-off.
• Neural representations can change without learning, a phenomenon called representational drift.

Purpose of the Study:

• Investigate whether representational drift yields unique or advantageous neural representations.
• Determine if drift explores representations different from learned ones.
• Clarify the role of drift in memory stability and plasticity.

Main Methods:

• Defined the nonlinear solution space manifold of synaptic weights.
• Simulated representational drift as diffusion within this manifold.
• Introduced a novel allocation procedure to shift representations.

Main Results:

• Representational drift uncovers noise-robust representations with inactive/saturated neurons.
• These drift-explored solutions are entropically favored but difficult to learn due to lack of gradients.
• The allocation procedure shifts representations into a learning-conducive regime.

Conclusions:

• Representational drift discovers robust, albeit hard-to-learn, neural representations.
• Combining drift with allocation resolves the learnability-robustness trade-off.
• This approach offers a novel strategy for memory storage and continual learning.