Declarative Memory Blending

• Like “weighted avg”
• store no of pulses
• activation decays in time
  • $A(t)=log(t-t_{creation}{)}^{-d}+\text{mismatchpenalty}$
  • Retrieval probability
    • Softmax
    • $P_i=\frac{e^{\frac{A_t}{t}}}{{\Sigma }_i{e}^{\frac{A_t}{t}}}$
  • Adds up to 1
  • $Result={\Sigma }_j{P}_j{V}_j$
  • t controls noise
    • if t is high : 1/no of competitors , more prob of retrieval
  • Looking for long interval (partial matching)
  • Penalty for short intervals
  • Apply $P_i$
  • Weighted avg $Result$
• Too short is positive. else negative, correct is 0
• no of pulses to wait : duration + feedback from memory

Fit

• Exp done on generated data as well
• Compares if same as when run on original
• Does well with unmodified mode
•