Complete reference for CMR model parameters

This page documents all parameters used in jaxcmr CMR models, their meanings, and typical value ranges.

Context Drift Parameters

These control how the temporal context representation evolves.

Parameter Description Typical Range
encoding_drift_rate How much context changes when studying each item. Higher values = faster drift, more temporal distinctiveness 0.3 - 0.9
start_drift_rate Context drift at the transition from study to recall. Often similar to encoding drift rate 0.3 - 0.9
recall_drift_rate Context drift after recalling each item. Controls how much recalled items reinstate their context 0.3 - 0.9

Intuition: With low drift rates, context changes slowly and distant items share similar contexts. With high drift rates, each item has a more distinct context, increasing temporal precision.

Primacy Parameters

These create the primacy effect (better memory for early list items).

Parameter Description Typical Range
primacy_scale Multiplier for the first item’s learning rate. Higher = stronger primacy 1.0 - 5.0
primacy_decay Exponential decay rate across serial positions. Higher = primacy effect decays faster 0.3 - 1.5

The learning rate for position \(i\) is:

\[\text{lr}_i = \text{primacy\_scale} \times e^{-\text{primacy\_decay} \times i}\]

Association Strength Parameters

These control the strength of item-context associations.

Parameter Description Typical Range
learning_rate Base strength for new MFC associations (item → context) 0.3 - 1.0
shared_support Pre-experimental MCF association strength (all context units → all items) 0.01 - 0.2
item_support Additional MCF association strength for matching context-item pairs 0.1 - 0.5

Intuition: shared_support creates a baseline retrieval probability for all items. item_support gives each item extra support from its corresponding context unit.

Retrieval Parameters

These control the retrieval process.

Parameter Description Typical Range
choice_sensitivity Softmax temperature for converting activations to probabilities. Higher = more deterministic (winner-take-all) 0.1 - 2.0

With choice_sensitivity = 0, retrieval is random among available items. As it increases, the most activated item becomes more likely to be retrieved.

Stopping Parameters

These control when recall terminates (using PositionalTermination).

Parameter Description Typical Range
stop_probability_scale Base probability of stopping at first recall position 0.01 - 0.2
stop_probability_growth Exponential growth rate of stopping probability 0.1 - 0.5

Stop probability at recall position \(i\) is:

\[P(\text{stop})_i = \text{scale} \times e^{\text{growth} \times i}\]

Boolean Parameters

Parameter Description Default
learn_after_context_update If true, associations use the updated (post-drift) context state Varies
allow_repeated_recalls If true, items can be recalled multiple times False

Example Parameter Set

A complete parameter dictionary for CMR:

Code 
params = {
    # Context drift
    "encoding_drift_rate": 0.5,
    "start_drift_rate": 0.5,
    "recall_drift_rate": 0.5,

    # Primacy
    "primacy_scale": 1.0,
    "primacy_decay": 0.6,

    # Association strength
    "learning_rate": 0.5,
    "shared_support": 0.05,
    "item_support": 0.25,

    # Retrieval
    "choice_sensitivity": 0.6,

    # Stopping
    "stop_probability_scale": 0.05,
    "stop_probability_growth": 0.2,

    # Boolean flags
    "learn_after_context_update": True,
    "allow_repeated_recalls": False,
}

Model Variants

Different CMR variants may require additional parameters:

Semantic CMR

For models with semantic similarity:

Parameter Description
semantic_scale Weight for semantic similarity in retrieval competition

Positional CMR

For models with positional encoding:

Parameter Description
position_encoding_weight Weight for positional vs. temporal context