IEEE Circuits and Systems Magazine - Q2 2020 - 39

In dealing with signals, HD computing usually makes use of
floating point models to improve the classification
accuracy at the cost of high computation cost.
2) Seizure Detection Using HD Computing
The Laelap algorithm, which utilizes local binary pattern (LBP) codes to conduct the feature extraction from
iEEG signals, has been proposed in [43] for seizure prediction. Here HD computing is -applied to capture the
statistics of the time-varying LBP codes for all the
electrodes. Fig. 11 illustrates the complete processing
chain. 1). Since the down-sampling frequency is 512 Hz,
thus every one second (1 s) data contains 512 samples.
Among these samples, the sampled iEEG signals are encoded to 6-bit LBP codes. This completes the feature
extraction part. 2). It utilizes record-based encoding,
where two types of hypervectors are randomly generated. Specifically, each LBP code is transformed to a
d-dimensional hypervector C i, while the hypervectors
E i are used to represent the corresponding electrode
name. For every new sample, the hypervectors E i and
C i are bound together to form a composite hypervector S = [C 1 5 E 1 + g C n 5 E n], where n is the number of
electrodes for a specific patient. Then the histogram of
LBP codes H is computed for a moving window of 1 s
with 0.5 s overlap. Therefore the composite hypervector H = [S 1 + S 2 + g + S 512] is updated every 0.5 s. 3).
For learning, two prototype hypervectors P 1 and P 2
should be trained. For interictal prototype vector P 1,
all H computed over 30 s should be accumulated and
normalized to be stored in the associative memory. Depending on the seizure's duration, the ictal prototype
vector P 2 is generated using all H over an ictal state,
which may last 10 s to 30 s. 4). For classification, com-

Retraining

C3
Training
Dataset

C26

NN Model
Associative Memory

26
HD
Encoder

C1
C2
C3

Training
Dataset

Testing
Dataset

C26

Distance Similarity

HD
Encoder

3) Quantization in HD Computing
In dealing with signals, HD computing usually makes
use of floating point models to improve the classification accuracy at the cost of high computation cost. In
[25], QuantHD is proposed as a quantization of HD model, which projects the trained non-binary hypervectors
to a binary or ternary model, with elements in {0, 1} or
{-1, 0, +1}, to represent class hypervectors. To compensate the accuracy degradation caused by quantization, a
retraining approach is used where an iteration number
of 30 is pre-defined. The similarity check is no longer
cosine metric (non-binary model), but Hamming distance (binary model) or dot product (ternary model).
Compared to the existing binarized HD computing, such
QuantHD improves on average 17.2% accuracy with a
similar computation cost.

Testing

Training
Associative Memory

paring Pk with a query H, the label is updated every
0.5 s with the shortest Hamming distance Ham(H, Pk),
where k = 1, 2. 5). The algorithm also generates the seizure alarm. In postprocessing, if the last 10 labels all
indicate P 2 (t c = 10) and the distance score T 2 t r , then
the seizure alarm is generated.
The evaluation shows the Laelaps algorithm outperforms other machine learning methods, such as SVM, in
terms of energy efficiency. It is worth noting that many
simpler seizure detection and prediction algorithms
have been proposed in the literature [45]-[49]. A fair
comparison of classifier accuracy between HD and traditional classification needs to be explored in future.

Testing
Dataset

HD Model

Figure 10. VoiceHD+NN flow for training and testing [28].

SECOND QUARTER 2020

IEEE CIRCUITS AND SYSTEMS MAGAZINE

IEEE Circuits and Systems Magazine - Q2 2020

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q2 2020