IEEE Circuits and Systems Magazine - Q2 2020 - 42

HD Model

Expansion

Figure 13. Overview of SemiHD framework supporting self-training in HD space [53].

Encoded Data
Encoder
Labeled
Data

Unlabeled
Data

Encoder

Encoded Data

Label

Training

Update Training Data

Class k Vector

Class 2 Vector

Class 1 Vector

Encoded Data

Label

Label C
HD Model

Encoded Data

Encoding
Training Data

Class k Vector

Class 2 Vector

Class 1 Vector

Query Vector

S% High
Confidence

Similarity

Encoder
Test Data

Inference
Labelling Data/Model Expansion
Confidence
Model Generation

Encoded
Training Data
42

6) A Binary Framework for HD Computing
Generally speaking, HD classification using binary
hypervectors shows lower accuracy but higher
energy efficiency than non-binary ones. This is because the non-binary framework makes use of the
costly cosine similarity rather than the hardwarefriendly Hamming distance metric. In [35], BinHD
uses three main blocks, encoding, associative
search and counter modules, dealing with binary
hypervectors. Their evaluation shows that, over
four practical applications, the proposed BinHD
can reach 12.4# and 6.3# energy efficiency and
speedup in training process, while 13.8# and 9.9#
during inference, compared to the state-of-art HD
computing algorithm with a comparable classification accuracy.

IEEE CIRCUITS AND SYSTEMS MAGAZINE

7) HD Computing for Semi-Supervised Learning
In [53], SemiHD has been proposed as a self-training or self-learning approach for semi-supervised
learning, where the training data is composed of a
small portion of labeled data and a large portion of
unlabeled data.
The SemiHD framework is depicted in Fig. 13
and the flow is illustrated as follows. 1). Encode
all the data points, labeled and unlabeled, into HD
space with d = 10, 000 dimensions. 2). Start training from the labeled data to generate k hypervectors, each representing one class. 3). Predict the
label for unlabeled data points. Labeling is performed by checking the similarity of unlabeled
data with all the class hypervectors, and return
the label which shows the highest similarity. 4).
Select and add S% of unlabeled data with highest confidence to labeled data, where S is defined
as the expansion rate. In [53], typically S = 5. 5).
Redo the training task based on the expanded
labeled data. Such iterative process stops when
the accuracy does not change more than 0.1%. 6).
Once the model has already been trained, perform
the inference task by comparing the similarity of
each test data with the trained model, to return
the label with maximum similarity.
Their evaluation shows that the SemiHD can on
average improve the classification of supervised
HD by 10.2%. Additionally, compared to the best
CPU implementation, the FPGA -counterpart of
SemiHD offers 7.11# faster speed and 12.6# energy efficiency.
8) HD Computing for Unsupervised Learning
HD computing has also been used in several unsupervised applications. See [57]-[61].
SECOND QUARTER 2020

IEEE Circuits and Systems Magazine - Q2 2020

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