How MINERVA works

Hintzman’s multiple-trace model, as an audio effect

MINERVA II (Hintzman, 1984, 1986, 1988) is a model of human memory with a striking premise: there are no stored abstractions, only traces of individual experiences. Every experience is laid down as its own trace. Remembering is not looking something up: a probe (the current experience) is compared with every trace at once, each trace responds in proportion to its similarity, and what comes back, the echo, is the sum of all those responses. Generalisation, prototypes and familiarity all emerge from that one operation.

MINERVA Space Echo takes the model literally, with audio.

Traces

The input is cut into segments (1 bar by default; from 10 ms to 20 s, tempo-synced or free). Each finished segment can be stored as a trace, which holds:

its audio, which is what the echo plays back;
its address, a feature vector that describes it, which is what MINERVA compares.

The address is a small spectrogram: 16 time slots × 24 frequency bands (60 Hz – 16 kHz), 384 numbers in all. Band energies are measured in decibels and then standardised within the trace (mean 0, standard deviation 1), so the address describes the shape of the sound, its rhythm and timbre, independent of how loud it was. Slots that received no audio are 0: in MINERVA, 0 means “not encoded”.

Optionally the address can be ternary (−1, 0, +1), as in Hintzman’s original simulations, with Ternary Threshold deciding how far from average a value must be to count.

This spectrogram is the Spectrum address. Every trace also stores four other addresses, describing its notes (Pitch Class, Pitch), its sound (Timbre) and its attacks (Rhythm); the Address parameter decides which one MINERVA compares, or a weighted mix of them. See Addresses.

Similarity

When a segment ends, it becomes the probe P and is compared with every stored trace T:

\[ S = \frac{1}{N}\sum_{j} P_j\,T_j \]

where N is the number of features that are non-zero in either vector (Hintzman similarity; Cosine is the alternative). S is 1 for identical addresses, near 0 for unrelated ones, and negative for opposites.

Activation

Each trace’s activation is its similarity raised to a power, keeping the sign:

\[ A = \operatorname{sign}(S)\,|S|^{p} \]

The Activation Power p is the model’s most important knob. Hintzman used p = 3 (the default). With p = 1 every trace contributes almost equally and the echo is a blur of everything in memory, a schema. With p = 9 only near-identical traces respond.

Activation is also scaled by each trace’s strength (which fades if forgetting is on) and, optionally, by Recency.

The echo

The echo is the activation-weighted sum of the traces’ audio:

\[ \text{echo}(t) = g \sum_i A_i\,\text{trace}_i(t) \]

The normalisation g comes from Echo Normalization: Sum divides by \(\sum|A_i|\) so the echo keeps a constant level; Max by the largest activation; Familiarity leaves weak echoes weak, so an unfamiliar cue gets a quiet answer. Negligible activations are dropped so only traces that matter are mixed, and Max Active Traces caps how many are mixed at once.

The echo plays during the next segment, in time with the grid, which is why the model behaves like a delay. Echo Level Tracking makes the echo’s level follow the cue’s, the way a tape repeat follows what was played, so feedback decays naturally.

Intensity and familiarity

MINERVA’s intensity is the sum of activations, \(I = \sum_i A_i\): how strongly memory as a whole resonates with the probe. Hintzman used it to model recognition (“have I heard this before?”). The plug-in shows it, together with the strongest single activation (familiarity), in the window, and can route familiarity to the echo’s tone and feedback.

The echo also has an address of its own: the activation-weighted mean of the answering traces’ addresses. The window shows it as Echo, next to what you played (Heard). An iterative head uses it to cue memory again: the echo of the echo.

Encoding and forgetting

MINERVA’s learning parameter L is the probability that each feature is stored correctly. Here, Encoding Failure zeroes each stored feature with that probability (the cue you heard stays intact), so recall from poorly encoded traces is vaguer. Over time, Forgetting / Segment keeps knocking features out and Fading / Segment reduces strength until a trace drops out of memory altogether.

Sequences

Jamieson and Mewhort (2009) extended MINERVA to sequence learning by storing each event together with the event before it, [n−1 | n], much like the context units of an Elman network. Cue with the event just seen against the traces’ n−1 halves, and the echo’s other half is an expectation of what comes next. MINERVA Space Echo can do the same with segments of audio, and can let each echo’s expectation cue the next echo; see Sequences.

The pipeline

flowchart TD
    IN([Input]) --> SEG[Cut into segments]
    SEG --> FEAT["Address: 16 slots × 24 bands"]
    FEAT -- store --> STORE[(Memory: traces)]
    FEAT -- probe --> SIM[Similarity to every trace]
    STORE --> SIM
    SIM --> ACT["Activation A = sign(S)·|S|^p"]
    ACT --> ECHO["Echo = Σ A · trace audio"]
    STORE -. audio .-> ECHO
    ECHO --> OUT([Output: dry + echo])
    ECHO -. feedback .-> SEG