EmoPsy (CMACD): Brain-Inspired Emotion Engine

• EmoPsy (CMACD) is a hardware-optimized emotion recognition engine based on Hyperdimensional Computing that integrates combinatorial channel encoding and cellular automaton-based hypervector synthesis.
• It employs early sensor fusion and temporal encoding to robustly classify high versus low valence and arousal from high-channel-count physiological datasets such as AMIGOS and DEAP.
• The innovative design reduces memory and computational demands by over 98%, achieving competitive two-class accuracy while enabling always-on, low-power emotion tracking.

EmoPsy (CMACD) is a hardware-optimized, brain-inspired emotion recognition engine based on Hyperdimensional Computing (HDC), integrating combinatorial channel encoding, cellular automaton–driven on-the-fly hypervector generation, and early sensor fusion to classify high versus low valence and arousal using high-channel-count physiological datasets, such as AMIGOS and DEAP. The system achieves strong two-class accuracy while reducing memory and computational demands by over an order of magnitude, rendering it appropriate for always-on, low-power emotion-aware devices (Menon et al., 2021).

1. Hyperdimensional Computing Principles

EmoPsy employs hyperdimensional representations in a high-dimensional space ($D = 10\,000$), encoding data as random bipolar ($\pm1$) or binary ({0,1}) hypervectors (HVs). Binding—elementwise XOR for binary, multiplication for bipolar—and bundling—elementwise majority vote—implement algebraic memory and association operations. Quasi-orthogonality ensures that randomly drawn HVs are nearly disjoint: $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$ This algebra supports efficient, massively parallelized computation with hardware-friendly vector operations (Menon et al., 2021).

2. Combinatorial Channel Encoding

Conventional HDC approaches require three HVs per channel ($n>200$): item memory (iM) and two feature projections (PFP/NFP), resulting in substantial storage (e.g., 642 HVs for AMIGOS). EmoPsy's combinatorial binding scheme assigns each channel a triple $$\{V^{(c)}_{\mathrm{iM}}, V^{(c)}_{\mathrm{PFP}}, V^{(c)}_{\mathrm{NFP}}\}$$, with spatial encoding per channel-feature sample $f_{c,j}$,

$$\mathrm{SE}_{c,j} = V^{(c)}_{\mathrm{iM}}\,\,\oplus\,\, V^{(c)}_{\mathrm{sign}(f_{c,j})},$$

where $\mathrm{sign}(f)$ selects the feature sign. By drawing $v$ seed HVs and forming combinatorial triples, the channel encoding dimension scales as $\mathrm{TFC}(v) \approx \binom{v}{3}$. For instance, $\pm1$0 yields $\pm1$1 unique triples, reducing storage from 642 to 31 HVs (95% reduction on AMIGOS) and 714 to 32 HVs on DEAP (Menon et al., 2021).

3. Cellular Automaton-Based Hypervector Generation

Further memory efficiency derives from using a Rule-90 elementary cellular automaton (CA) to generate HVs on-the-fly: $\pm1$2 where $\pm1$3 denote cyclic bit shifts. Rule-90 CA offers near-maximal entropy, preserving orthogonality and Hamming weight ($\pm1$4). Only a small bank of HV seeds needs to be stored; the rest are synthesized at runtime with minimal logic (two shifts, one XOR per HV), slashing storage by an additional factor of 4–5×. For AMIGOS, moving from combinatorial encoding to CA-based generation reduces HV storage down to only 7 vectors ($\pm1$5 reduction) (Menon et al., 2021).

4. Early Sensor Fusion and Temporal Encoding

The processing pipeline includes:

1. Raw Feature Mapping: Each sample $\pm1$6 is mapped into HDC space using iM or the CA HV generator.
2. Modality-Specific Spatial Encoding:

$\pm1$7

where $\pm1$8 denotes channels for modality $\pm1$9.

1. Early Fusion:

$$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$0

replacing $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$1 temporal encoders with a single instance.

2. Temporal Encoding: An $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$2-gram of length $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$3 encodes sequential dependencies:

$$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$4

Training bundles temporal encodings by class, while inference assigns class labels via normalized Hamming distance: $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$5

5. Memory, Computation, and Hardware Implications

This architecture produces substantial resource gains:

• For AMIGOS ($$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$6): storage shrinks from 642 HVs to 31 HVs with combinatorial encoding, then to 7 HVs with Rule-90 CA (98.9% reduction); for DEAP ($$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$7): from 714 down to 11 HVs (98.5% reduction).
• Vector request rates drop ≥5×, as HVs are synthesized rather than fetched.
• On ASIC prototypes, the memory array area and SRAM-read energy plummet commensurately; dynamic power for CA-based HV generation is negligible (<1 µW per inference).
• Early fusion eliminates redundancy by removing the need for separate temporal encoders per modality, saving $$\mathbb{P}(\langle A, B\rangle \approx 0\,) \approx 1,\, D \gg 1.$$8 permutation hardware.
• Total area reduction reaches 87.3% over prior HDC implementations (Menon et al., 2021).

Table: Storage Requirements for HVs

Dataset	Baseline HVs	Combinatorial HVs	Rule-90 CA HVs
AMIGOS	642	31	7
DEAP	714	32	11

6. Accuracy and Comparative Performance

EmoPsy (CMACD) attains two-class accuracy consistently above high-performance benchmarks:

• AMIGOS: 87.1% for valence, 80.5% for arousal (versus SVM/XGB/ELM at 68–84%/66–83%; late-fusion HDC baseline 83.2%/70.1%).
• DEAP: 76.7% (valence), 74.2% (arousal) (compared with DBN/RBM+SVM/MESAE at 51–76%/64–77%).

Class sensitivity and specificity are balanced, as shown by confusion matrices and ROC curves (not shown in the primary source) (Menon et al., 2021).

7. Significance and Application Contexts

EmoPsy (CMACD) demonstrates that HDC with combinatorial and automaton-based encoding allows real-time, on-device emotion recognition using >200 physiological channels and multiple modalities, with extreme memory and area efficiency. This enables continuous, low-power emotion tracking in wearables and mobile systems unsuited to conventional deep learning frameworks. The methodological innovations are extensible to other multi-channel time-series tasks requiring minimal model footprint without sacrificing accuracy (Menon et al., 2021).