IEEE Consumer Electronics Magazine - September/October 2022 - 103

However, the GAN is computationally complex
and unable to balance generated and new samples.
Similarly, an auditory scene analysis (ASA)
algorithm is utilized8 for incrementally learning
audio sounds in real-time environments. The ASA
algorithm consists of different operations for
audio features extractions. The real-time ASA algorithm
continuously learns audio scenes. The ASA
algorithm utilizes the hidden Markov model
(HMM) to detect old and new tasks. However,
HMM model has a large number of unstructured
parameters and is computationally complex.
Furthermore, a sequential self-teaching network
is utilized for training weakly labeled audio
data sequentially.9 The previous network supervises
the current network through cosupervision.
The sequential self-teaching framework
utilizes the teacher-student strategy. The knowledge
is transferred between two networks
through knowledge distillation. Similarly, the fast
incremental classifier and representation learning
(FastICARL) framework is utilized in the paper
by Kwon et al.,10 which allows effective incremental
learning in mobile sensing applications. The
FastICARL algorithm effectively reduces the
incremental learning time and reduces the memory
requirements. These research approaches
play a vital role in the incremental learning of
audio classification. However, large-scale incremental
learning still remains a challenging and
open problem.
Motivated by the above analysis, we develop
a single network framework with static memory
for incremental learning of audio classification in
this work. The proposed method is supposed to
learn new tasks incrementally while remembering
the old tasks by utilizing static memory. We
utilized the DenseNet-121 and DenseNet-201 as
the backbone models for audio classification.
The proposed method utilizes limited memory
and reduces the computational complexity of
the model.
PROPOSED METHODOLOGY
Audio classification plays a vital role in
human interaction with the environment. It is
important to build an intelligent system that can
recognize and understand audios incrementally.
In this era, task incremental learning is a
September/October 2022
challenging problem in audio classification. We
investigated this problem through a static memory
borrowing idea from the memory-based
method implemented in the paper by Lopez-Paz
and Ranzato.11 The static memory saves a fixed
number of samples from the old tasks and utilizes
it during the execution of the new tasks,
thereby mitigating catastrophic interference, as
shown in Figure 2.
Models
We utilize DenseNet-121 and DenseNet-201 as
backbone classification models for our proposed
method. DenseNets use concatenating terminology
to obtain collective knowledge from all preceding
layers. Each previous layer output is
connected with the future layer. It consists of
two important blocks: dense blocks and transition
layers. DenseNets manipulate the potential
of the network through feature reuse. DenseNets
architecture consists of a composition layer followed
by a batch normalization layer, a nonlinear
activation layer, and a convolution layer
with output features of n-channels. DenseNet121
and DenseNet-201 belong to the DenseNets
group of models designed for image classification.
The numbers 121 and 201 correspond to
the number of deep layers.
Incremental Learning With Static Memory
Suppose that we have tasks Tfrom streams of
i ;tk
i ;yk
i g, where Dk is the dataset
k set and validation set Dv
Pk
k. For kth task,
data Dk¼fxk
for kth task. Each datum Dk is split into training
Dtr
each dataset Dk contains a list of triplets
Dk ¼
i¼1ð xi;yi;tiÞ, where yi is the ith label of
the corresponding input spectrogram xi and ti
represents a task descriptor, which determines
the origin of the data coming from which task.
The task descriptor keeps observing the samples
coming from the training dataset Dk. During
incremental learning, the target is expected to
learn a predictor fu, where u 2 Rp, projecting the
test pair ðx; tÞ to the target y. The training dataset
Dk for kth task contains minibatches Bk gradient
for each task. The minibatches are
considered from the static memory to keep
model performance fast. We can calculate the
average accuracy for each jth set of test tasks
after training the model for ith minibatch of task
103
IEEE Consumer Electronics Magazine - September/October 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - September/October 2022
