IEEE Computational Intelligence Magazine - February 2022 - 36

II. Related Work
Our work contrasts with prior research that seeks to improve
the understanding of time series predictions using post-hoc
interpretations [8], [10]-[13]. These approaches provide feature
importance, or relevance, at a given time step. Although
listing the most discriminating features is insightful, the resulting
explanations are often regarded as misleading and unreliable
[14]. In contrast, PIP weaves interpretability into its
training process. While domain knowledge is needed to
understand post-hoc visualizations, PIP pictures are based on
the end-users' background. For example, post-hoc methods
can isolate portions of an accelerometer signal that cause the
learned model to predict human activity. However, the provided
information may not be useful to a nurse nor informative.
Alternatively, PIP provides a picture explanation of the
classification, offering patients and caregivers an intuitive
explanation of the finding.
Creating glass-box models for time series classifiers has also
previously been considered. IETNet [15] graphs a heatmap of
the class-influential channels during multivariant time series
classification. DPSN [16] tackles few-shot learning by employing
a prototypical network [17] to learn a prototype for each
class from a symbolic Fourier approximation transformation of
the data. Gee et al. [18] and ProSeNet [5] utilize prototype networks.
These models integrate a designed layer into the network
architecture to learn the prototypes.
The approach of Gee et al. [18] and the ProSeNet algorithm
[5] can be considered case-based approaches to interpretability.
These methods learn explainable prototypes and
classify new data based on similarity to the prototypes. This type
of network includes an encoder and a prototype layer. The
encoder can be any network structure that encodes input data,
such as a convolutional neural network (CNN) or a recursive
neural network (RNN). The prototype layer aims to learn a
prototype that is close to at least one of the encoded inputs. Gee
et al. [18] adapt an image prototype classifier introduced by Li
et al. [19] by coercing time series data to appear as graph images.
Inputs
Encoder
f(x)
z
Generator
g(z)
ypic
FIGURE 2 The PIP architecture. The model consists of four components: an encoder f(x) that transforms the input to the latent space z, a prototype
layer p with m prototypes, a fully connected layer FC with a softmax layer for multi-class classification, and a generator g(z) for converting
the latent space to a correct sketch representation of the data. PIP learns the prototypes p by minimizing the distance between embedded input
z and each prototype
p .i
Additionally, PIP minimizes the distance between each prototype pi
prototype to cluster around one class. The input to the fully connected layer FC is the normalized distance a between the encoded data z, and
each prototype p. PIP simultaneously utilizes the generator to transform the encoded data g(z) into their respective picture
types are based on the encoded data, the generator converts the learned prototypes into their corresponding pictures after training.
pic
pm
am
The image prototype classifier contains a decoder that transforms
the encoded data and prototypes into the original image. Likewise,
Gee et al. [18] employ a decoder to transform learned prototypes
into time series graphs. Since time series are not readily
interpretable as images, Gee et al. [18] feed training inputs into
the model post-training to determine what class each prototype
represents. ProSeNet [5] does not include a decoder in its
architecture and illustrates the prototypes as it learns.
ProSeNet's learned prototypes are intangible because they are
based on encoded inputs. Like Gee et al. [18], ProSeNet [5]
adds a post-training step to determine each prototype's class.
PIP represents a case-based prototype learner that offers distinct
contributions from these previous works. Unlike prior
approaches, PIP learns visual interpretations based on externallyprovided
sketches; thus, they do not need to be labeled after
training. Furthermore, PIP's prototypes do not require significant
expertise to understand. Because these methods optimize multiple
criteria (e.g., classification accuracy and interpretability), they
rely upon multi-term loss functions. While previous methods
employ a costly cross-validation step to tune the ratios between
these loss terms, PIP directly learns these ratios during training.
III. Methods
Complex models such as ensemble and deep neural networks
have demonstrated the ability to achieve high accuracy on time
series data [20], [21], but they are difficult to interpret. To
address this dichotomy between performance and interpretability
[7], PIP learns a set of prototypes by leaning on userdefined
explanations (hand-drawn sketches). Before training,
users sketch a picture for each of the C classes. This customization
allows PIP to adapt its prototypes to suit the needs of each
audience. For instance, a clinician may design different signs
than an engineer for a set of wearable sensor data.
A. PIP Architecture
Figure 2 illustrates the PIP architecture. Our model consists of
four main components: an encoder, a prototype layer, a generator,
Prototypes
p1
p2
FC
a1
a2
Output
Walking Upstairs
and the embedded data z, thus encouraging each
y . Since the proto36
IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2022
. . .
. . .
. . .

IEEE Computational Intelligence Magazine - February 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2022