Vector Hash Memory Model

Original Paper: Vector Hash Memory Model

The presentation summarized a new model called VectorHASH (Vector Hippocampal-Scaffolded Heteroassociative Memory), how episodic, associative, and spatial memory can arise from a common hippocampal scaffold.

Summary of the Paper (VectorHASH Model)

Core Problem and Solution

• Traditional memory networks (e.g., Hopfield Network) suffer from the Memory Cliff: recall accuracy drops abruptly to zero once storage slightly exceeds neuron count.
• The VectorHASH model avoids catastrophic failure by providing a memory continuum—recall that gradually degrades near capacity, staying close to the theoretical limit.

Key Architectural Concepts

1. Separation of Scaffold and Content

   • Scaffold = stable structure (coordinates, grid cells).
   • Content = variable memory (sensory input).
   • Architecture grounded in biology: Sensory Input Layer + Hippocampus + Grid Cells in entorhinal cortex
     • EC → HC connections: fixed and random (scaffold).
     • HC → Sensory connections: plastic, experience-dependent.
   • Heteroassociation: sensory input linked to hippocampal scaffold.

2. Low-Dimensional Encoding for Sequences

   • Episodic memory stored via Velocity Shift Mechanism:
     • Only changes are stored in a low-dimensional velocity vector (often 2D).
     • Tracks grid phase shifts instead of high-dimensional states.

Performance and Capacity

• Robustness to Noise: Attractor recovery even with 25% noise in HC state.
• Exponential Capacity: Stable attractors grow exponentially with number of grid modules; HC neurons scale linearly.
• Gradual Degradation: Quality of recall decreases smoothly at high load (vs. collapse).
• Efficiency: Recall of sequences up to 14,000 steps, vs. ~53 for Hopfield.
• AI Comparison: Outperforms autoencoders in image recall fidelity → biological constraints as powerful inductive biases.

Spatial and Episodic Memory Functions

• Spatial Functions:

  • Place Cells & Grid Cells reproduced.
  • Zero-Shot Inference for new paths.
  • Supports remapping → distinct, non-interfering codes across environments.

• Episodic Functions:

  • Sequences maintained, but fidelity of sensory details decays with load.

• Memory Consolidation:

  • Repeated inputs strengthen HC→Sensory weights.
  • Leads to stronger recall and resilience to HC damage.

Connection to the Memory Palace

• Method of Loci aligns with VectorHASH:
  • Fixed palace path = stable scaffold.
  • Linking items to locations = heteroassociative learning.
• Explains vast storage capacity of mnemonic strategies.

Summary of the Discussion

Inductive Bias and AI

• Catastrophic Forgetting Mitigation: Fixed scaffold serves as inductive bias.
• Scaling vs. Bias:
  • Debate between brute-force scaling (vision transformers) vs. neuroscience-inspired inductive biases (convolutional neural networks)
  • Robotics and constrained domains may favor bias-based approaches.

Neuroscientific Extensions and Implications

• Beyond Hippocampus:

  • Extended to Frontal Cortex (FC) → improved performance in reward-seeking tasks with positional + evidential information.

• Sensory Input Modalities:

  • Different modalities (visual, auditory, motor) may wire uniquely.
  • Feynman anecdote illustrates individual modality dependence in cognition.

• Early Fixed Learning:

  • HC–EC scaffold connections set early in development, and do not change after that.
  • Raises questions about critical periods in learning.

• Pathology:

  • Model relevance to Alzheimer’s disease and hippocampal dysfunction.

Original Video