IEEE Circuits and Systems Magazine - Q2 2020 - 43

1) HD Classification for Character Recognition
HD classification has been used for character recognition
in [62] and later in [56]. As shown in Fig. 14, the input
image is composed of 7 # 5 = 35 pixels. Each pixel has
two possible values, that is 0 or 1, representing black
or white. 1). Encode each pixel to a binary hypervector
(indexHV). Totally 35 orthogonal indexHVs are stored in
the item memory. 2). Based on HoloGN encoding-an encoding method proposed in [62] to address image data
using HD computing-the indexHV is shifted depending
on the pixel value. Accumulate all 35 indexHVs and perform a majority rule by a thresholding block to generate
a holoHV for one input image. 3). The supervised controller will only be activated when this HD system conducts
supervised learning. Otherwise, the system conducts the
one-shot learning. The supervised controller accumulates
the holoHVs for the same class and employs the thresholding block and generates the letterHV to be stored in
the associative memory. The total number of letterHVs
is 26. 4). During the test phase, the query hypervector is
generated following the same module with test data. Then
the similarity of each query hypervector is computed for
all trained letterHVs to find the most similar class.
Results in [56] show that HD computing performs
well for character recognition. Further optimization for
HD computing may be conducted by reducing the dimensionality and increasing the input image size. The
results also show that HD computing offers great robustness against noise. The system of 4,000-bit hypervectors
achieves comparable average accuracy to its 12,000-bit
counterpart at 0% distortion, and achieves an average
accuracy of 89.94% with 14.29% distortion.
D. Summary
As mentioned above, HD computing shows great potential in dealing with data in the form of signals [28, 43, 65,

Input Image

R1C1 IndexHV

Shifter

A LetterHV

R1C2 IndexHV

Shifter

B LetterHV

Distance

R7C5 IndexHV

Shifter

Item Memory

Bundling

C1 C2 C3 C4 C5
1 1 0 1 1
1 0 1 0 1
0 1 1 1 0
0 0 0 0 0
0 1 1 1 0
0 1 1 1 0
0 1 1 1 0

Majority

R1
R2
R3
R4
R5
R6
R7

66, 75], letters [30, 64], and images [9, 56, 62], as long as
these can be transformed into the HD space. Such preprocessing may include feature extraction and encoding.
Evaluation shows that HD computing achieves good results for seizure detection [43, 66]. In addition, HD computing can also be combined with quantization technique
to binarize HD model with minimal accuracy loss [76].
Table 3 offers more details about improvement strategies
adopted in HD computing for accuracy and efficiency. As
can been seen from Table 4, HD computing offers an acceptable accuracy, but with quite high efficiency. In some
applications like DNA sequencing [31], HD computing outperforms other machine learning methods.
There still exist some interesting papers not discussed in detail in this review paper. Interested readers can refer to the following references, which include
but are not limited to: 1). Considering the security issue
when IoT devices release the offload computation to the
cloud, [77] illustrates how the proposed SecureHD accelerates efficiency with high security. 2). To balance
the tradeoff between efficiency and accuracy, QubitHD
[76] is proposed as a stochastic binarization algorithm
to achieve comparable accuracy to the non-binarized
counterparts. SparseHD [37] takes advantage of the sparsity of the trained HD model for acceleration.
HD computing is still in its infancy. Future directions
may include but is not limited to:
■■ More cognitive tasks: Inspired by [32], apart from
the engineering aspect of HD computing, which is
to solve classification tasks, more "cognition" aspects of HD computing should be explored. Such
tasks i-nclude but are not limited to analogical
reasoning, semantic generalization and relational
representation.
■ ■ Feature exaction and encoding method: Since
HD computing cannot directly address data like
signals and images, feature exaction is vital to
representation of information. For example, [75]

Z LetterHV
Distance

Bundling

Encoder

Distance

Supervised
Controller

Least Comparator

C. Images

Associative Memory

Figure 14. Block diagram of the HD character recognition system [56].

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