🤖 machine learning

Modelling Emotional Memory in Children with Tensor Networks

This paper demonstrates that a classical tensor network model incorporating emotional valence significantly outperforms standard psychological models (achieving 77.98% accuracy) in capturing the order-dependent structure of children's emotional memory, thereby validating the utility of quantum-inspired methods for modeling such cognitive phenomena.

Original authors: Henry Groves, Lucia F. Jackson, Barbara-Anne Robertson, Jonte R. Hance

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Henry Groves, Lucia F. Jackson, Barbara-Anne Robertson, Jonte R. Hance

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Memory Isn't a Video Camera

Imagine your brain isn't like a video camera that records life in perfect, fixed snapshots. Instead, think of memory like a jazz improvisation. When you try to remember something, your brain doesn't just "play back" a file; it actively rebuilds the scene. It takes pieces of what you saw, felt, and thought, and puts them back together.

The problem is, this rebuilding process is messy. It's influenced by how you feel right now, what you believe, and the order in which things happened. The paper argues that standard ways of studying memory (like simple math) are too rigid to capture this messiness, especially when emotions are involved.

The Experiment: The Toy Box Game

The researchers wanted to see how children remember a sequence of events when emotions are mixed in.

The Setup: They showed 50 children (ages 4–11) a sequence of 5 toys.
The Twist: Some toys were "neutral" (plain wooden blocks), and some were "positive" (fun cars, action figures).
The Task: After watching the toys appear one by one, the children had to pick the toys out of a box and put them back in the exact order they saw them.

What they found with standard tools:
Using traditional psychology methods, they confirmed that kids remembered the fun toys better than the boring ones. However, these methods failed to explain why the order mattered. They couldn't see how remembering one toy changed the chances of remembering the next one. It was like trying to describe a complex dance by only counting the steps, ignoring the rhythm and the music.

The New Tool: The "Tensor Network"

To fix this, the researchers used a mathematical tool called a Tensor Network.

The Analogy: Imagine you are trying to predict the weather. A simple model might say, "If it rained yesterday, it will rain today." But a Tensor Network is like a giant, interconnected web of dominoes.
In this web, knocking over one domino (remembering Toy #1) doesn't just affect the next one; it changes the tension in the whole web, influencing how likely the other dominoes are to fall (remembering Toy #2, #3, etc.).
This tool is "quantum-inspired." It doesn't mean the brain is a quantum computer. It just means the math used to describe how particles interact in physics is surprisingly good at describing how human thoughts interact. It allows for superpositions (where a memory can be "remembered" and "forgotten" at the same time until you make a choice) and interference (where one memory changes the probability of another).

The Three Models Tested

The researchers built three different versions of this "web" to see which one worked best:

The Simple Model (The "On/Off" Switch):
- This model only asked: "Did they remember it? Yes or No?"
- Result: It was terrible (less than 50% accuracy). It was like trying to describe a color photo using only black and white pixels. It missed all the nuance.
The Better Model (The "Three-Option" Switch):
- This model added a middle option: "Remembered but in the wrong order."
- Result: It got worse (around 22% accuracy). Why? Because the data was too sparse. There were too many possible combinations for the limited number of children to learn from. It was like trying to guess a password with too many possible letters and not enough clues.
The Winning Model (The "Emotion-Aware" Web):
- This model combined the "Three-Option" idea with the Emotional Context. It forced the math to treat the type of toy (fun vs. boring) and the act of remembering as one single, inseparable event.
- Result: 77.98% accuracy. This was a massive jump.

Why Did the "Emotion-Aware" Model Win?

The paper suggests that emotions act like a structural glue for memory.

The "Fun Toy" Effect: When a child saw a fun toy, it didn't just make that specific toy easier to remember. It actually reshaped the entire web of how they remembered the toys before and after it.
The Analogy: Think of the sequence of toys as a string of beads. If you put a heavy, glowing gem (a fun toy) in the middle, it doesn't just shine; it pulls the string tight, changing how the other beads sit.
The winning model realized that you cannot predict what a child will remember without knowing the emotional flavor of the toys they just saw. The emotion wasn't just a side note; it was part of the memory's DNA.

The Takeaway

Memory is Contextual: You can't understand how a child remembers a sequence without understanding the emotional context of every item in that sequence.
Order Matters: What you remember first changes what you remember next, and emotions amplify this effect.
New Math for Old Problems: By using "quantum-inspired" math (Tensor Networks) that accounts for these complex, shifting relationships, the researchers could predict children's memories with much higher accuracy than traditional psychology models.

In short: Emotions don't just make memories brighter; they change the architecture of the memory itself. And to understand that architecture, we need a new kind of mathematical map.

Technical Summary: Modelling Emotional Memory in Children with Tensor Networks

1. Problem Statement

Human memory is not a static replay of events but an active, reconstructive process sensitive to context, emotional valence, and temporal order. While classical psychological models acknowledge that emotional experiences are often remembered more vividly than neutral ones (the emotional memory enhancement effect), they struggle to explain the complex, order-dependent dynamics of temporal memory. Specifically, standard models often assume:

Independence: Recall of one item is statistically independent of others.
Linearity: Emotional valence acts as a simple scalar modifier of recall probability.
Fixed Encoding: Memories are stored as fixed traces rather than evolving states.

These assumptions fail to capture how the emotional context of preceding and succeeding items influences the recall of a specific object. Furthermore, classical probabilistic models face the "curse of dimensionality" when attempting to model joint probability distributions of sequences with multiple outcomes, requiring an exponential number of parameters ( $d^n$ ) that become computationally intractable.

The paper posits that quantum-inspired frameworks, specifically Tensor Networks, offer a more suitable mathematical structure. Unlike classical models, these frameworks naturally handle non-commutativity (order dependence), interference, and context-sensitive state evolution without requiring exponential parameters, provided the underlying correlations can be compressed.

2. Methodology

A. Experimental Protocol

The authors conducted a study with 50 children (aged 4–11) to generate data on temporal order memory.

Stimuli: Five toys were presented sequentially in each trial. Toys were categorized as Positive (visually engaging, e.g., action figures) or Neutral (plain wooden blocks).
Procedure: Children viewed a sequence of 5 toys (10s each) and were later asked to select and arrange them in the original order from a box containing 7 items (5 targets + 2 decoys).
Design: A 2x2 within-subjects design manipulating Valence (Positive vs. Neutral) and Repetition (Initial presentation vs. Re-presentation after a delay).
Data Collection: 50 participants completed up to 10 trials. Data included object selection accuracy, order accuracy, and response times.

B. Classical Analysis

Initial analysis used Repeated-Measures ANOVAs and t-tests.

Findings: Positively valenced objects were recalled with higher accuracy and better temporal order than neutral objects. However, standard statistical tools failed to demonstrate the inherent order-dependence of the event, showing only mean differences with high noise. A classical linear model failed to adequately model the event, yielding low accuracy.

C. Tensor Network Modeling

The authors developed three progressive models using Matrix Product States (MPS), a type of tensor network, to factorize the joint probability distribution of recall outcomes.

Qubit Model (Binary):
- State Space: Each item $i$ is a binary variable: $|0\rangle$ (incorrect) or $|1\rangle$ (correct).
- Structure: A 5th-order tensor representing the joint probability of 5 items.
- Method: The tensor was constructed from experimental frequencies (with Laplace smoothing to handle sparsity) and factorized into an MPS using Sequential Singular Value Decomposition (SSVD).
- Prediction: Future items were predicted by conditioning on past observations and marginalizing future sites.
Qutrit Model (Three-State):
- State Space: Expanded to distinguish between:
  - $|0\rangle$ : Completely forgotten/incorrect item.
  - $|1\rangle$ : Correct item, incorrect position.
  - $|2\rangle$ : Correct item, correct position.
- Challenge: This increased the state space from $2^5$ (32) to $3^5$ (243), exacerbating data sparsity.
Valence-Primed Qutrit Model (Composite State):
- Innovation: This model explicitly encodes emotional valence into the physical index of the tensor network.
- State Space: Each site represents a composite state $|v_i, s_i\rangle$ , where $v_i \in \{0, 1\}$ (Neutral/Positive) and $s_i \in \{0, 1, 2\}$ (Recall outcome). This creates a 6-dimensional local space ( $2 \times 3$ ).
- Mechanism: The model treats the valence as a "priming" stage that conditions the recall state. During prediction, the known valence of the target item is used to project the tensor into the relevant 3-dimensional recall subspace before evaluating probabilities. This forces the network to learn the joint distribution of emotional context and recall outcome simultaneously.

3. Key Results

A. Model Performance

Qubit Model: Achieved 48.96% overall sequence accuracy, falling below the 50% binary chance baseline. It suffered from severe class imbalance, defaulting to predicting "incorrect" for most items.
Qutrit Model: Accuracy dropped to 22.71%. The model failed to predict "out-of-order" recalls (F1-score: 0.00), likely due to extreme data sparsity (243 possible sequences vs. 404 trials) and the inability to learn the conditional distribution of the minority class.
Valence-Primed Qutrit Model: Achieved 77.98% overall sequence accuracy.
- It successfully predicted the "out-of-order" class (F1-score: 0.618) and the majority "in-order" class.
- Accuracy was uniform across sequence positions (~70-78%), unlike the previous models which showed sharp drops after the first item.

B. Structural Analysis (Bond Dimensions)

Qubit/Qutrit: The singular value decay was rapid, and bond dimensions were low, suggesting the models were essentially memorizing the majority class rather than capturing complex dependencies.
Valence-Primed: The empirical bond dimensions rose significantly (e.g., [6, 22, 24, 6]), and singular values decayed more gradually. This indicates that the valence-conditioned model successfully captured rich, non-trivial correlations ("entanglement") between items.
Threshold Sensitivity: Unlike the simpler models, the valence-primed model's accuracy degraded when the SVD truncation threshold was increased beyond $10^{-2}$ . This confirms that the higher-order singular values contained genuine predictive information regarding the interplay of valence and order, rather than just noise.

C. Empirical Observations on Valence

The probability landscapes revealed that emotional valence is not a simple scalar modifier but a structural variable:

Positive-Neutral: A positive item followed by a neutral item often led to the failure to recall the neutral item, suggesting attentional resources were consumed by the positive item.
Neutral-Positive: A positive item following a neutral one was recalled well, consistent with emotional memory enhancement.
Positive-Positive: Consecutive positive items created competing encoding demands, reducing the recall of the second item.

4. Significance and Claims

The paper claims that the Valence-Primed Tensor Network demonstrates the efficacy of quantum-inspired methods for modeling order-dependent phenomena in human memory.

Superiority over Classical Models: The massive increase in accuracy (from ~23% to ~78%) by incorporating valence as a structural component, rather than an external variable, suggests that classical models assuming independence or linear encoding are insufficient for this task.
Context-Sensitivity: The results support the view that memory is a dynamic, state-dependent process where the emotional context of an event reshapes the probability structure of subsequent recall, rather than simply strengthening a fixed memory trace.
Quantum Cognition Framework: The study provides preliminary behavioral support for quantum cognition frameworks in developmental psychology. It suggests that children's temporal memory involves interference effects and non-commutative dependencies (where the order of evaluation matters) that are naturally captured by the Hilbert space structure of tensor networks.
Novel Tool: The experimental protocol introduced serves as a new tool for exploring emotional temporal memory in children, capable of being analyzed by both classical and quantum-like cognitive models.

The authors remain modest, noting that the study is limited by a single dataset, potential data sparsity issues, and the lack of a validated database for children's emotional toys. They do not claim to prove quantum physical processes in the brain, but rather that the mathematical structure of quantum mechanics (Hilbert spaces, superposition, interference) offers a coherent and more accurate framework for modeling the observed cognitive behaviors.