IEEE Robotics & Automation Magazine - June 2023 - 14

And a gradient projection memory (GPM)
method that takes advantage of orthogonal
gradient descent is proved to introduce
little interference with past tasks [18].
In this article, we employ the RGO [19]
"
WHILE LIFELONG
LEARNING METHODS
method, which belongs to the regularization-based
method. It leverages an iteratively
updated gradient optimizer and a virtual
feature encoding layer (FEL) to reduce the
catastrophic forgetting problem. Compared
with the replay method and parameter isolation
method, it is more computationally
efficient and does not need to maintain a
dynamically changing model architecture.
Recently, we found a direction constrained
optimization (DCO) [20] method, which
also belongs to the regularization-based
method. It studies the principal directions
of the trajectory of the optimizer. A linear
autoencoder is introduced to approximate
corresponding top forbidden principal
directions, and they are incorporated into
the loss function as a regularization term. However, the DCO
method needs to save the top forbidden principal directions after
the convergence of each task, and parameters from all previous
tasks are required to be retrained in the first step of the DCO
algorithm. In our RGO method [19], only the model parameters
of the last task are required when a new task arrives.
ARE HELPFUL TO HANDLE
SITUATIONS WHERE
DIFFERENT INDUSTRIAL
SCENARIOS ARRIVE
SEQUENTIALLY, THERE
IS STILL A GREAT
CHALLENGE IN LIFELONG
LEARNING CALLED
THE CATASTROPHIC
FORGETTING PROBLEM.
„
min LL 0
i
j
k
PROBLEM FORMULATION
We consider the problem of learning multiple defect classification
tasks sequentially. Assume that there are K
defect classification tasks |{ ,, ,} .
" Tk ! g12 K , For each
k
defect classification task T ,k it can be represented as Tk =
" xy i ! g k
(, ), ,, ,,, where x ,ki denotes the image,
y ,ki is the corresponding label denoting whether there is defect
ki ki
,,
" 12 n
/
=
-
1
1
k
=
in the image, and nk is the total number
of training samples for the kth task. It is
noted that there is only one kind of defect
in each task and there is no overlapping
between tasks.
The K defect classification tasks arrive
is the learned
sequentially and (; )fxi
defect classifier, where i is the weight of
the classifier, and ()l $
is the loss function
associating with the classifier and new
samples. Then the loss function of task Tk
can be expressed as:
L ()
k
ii=
=
1
the classifier (; )fxi
1 / ^^ hh (1)
nk i
nk
lf xy
;, .
ki ki
,,
The purpose is to continuously update
to generalize to all
the tasks. Concretely speaking, we would
like the updated classifier after task Tk to
still yield satisfying classification results
in previous task
T ,j where jk .1
The
accuracy of the current model on all previous tasks can be converted
into the summarization of the loss function of the current
model on all previous tasks, so the optimization problem can be
expressed as:
j (),().iidsubject to
(2)
(2) is meant to find a local optimum point of the current
task, which is able to minimize the total loss of previous tasks
so as to overcome the catastrophic forgetting problem. Figure
3 gives an intuitive demonstration of (2). Assume that the
background is an optimization surface obtained after learning
previous tasks, in which the lighter the color, the better
the classification performance is expected. Then the current
task Tk arrives and the curve illustrates the loss function for
task
T .k Due to the complexity of the neural network, there
could be several local optimum points. Although the yellow
point in Figure 3 is the best optimum point for the task
T ,k ,
Tk
the optimization goal is actually to find an optimum point that
performs well on both task Tk and previous tasks. So, the optimum
point eventually moves to the red point in Figure 3 where
the background is in a lighter color, and a local optimum point
is obtained as the result after task
T .k
FIGURE 3. An intuitive demonstration of RGO algorithm. Assume
that the background is an optimization surface obtained after
learning the previous tasks, and the curve illustrates the loss
function for task Tk. Although the yellow point is the optimum point
for task Tk, the RGO algorithm actually tries to find an optimum
point that performs both good on task Tk and the previous tasks.
So, the optimum point eventually moves to the red point and a
local optimum point is obtained as the result after task Tk.
14 IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023
PROPOSED FRAMEWORK
As shown in Figure 4, an RGO defect classification framework
is proposed, in which the trained defect classification
model is able to adapt continuously to the new task while
preserving its performance on previous tasks. For each task,
the defect classification model is expected to distinguish
whether the image contains the specific defect, and our
recently proposed the RGO method [19] is employed for
lifelong learning.
IEEE Robotics & Automation Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2023
