text
stringlengths
0
820
we start in grid cell j. For the very first query, we can de-
fine a dummy initial grid cell d, so that cost function c(d, k)
captures the initial query cost. Let qtdenote a query per-
formed in step t. Our ultimate goal is to solve the following
optimization problem:
max
{qt}U(x;{qt})≡∑
ty(qt)
s.t.∶∑
t≥0c(qt−1, qt)≤C,(1)
where c(q−1, q0)=c(d, q0).
In order to succeed, we need to use the labels from previ-
ously queried cells to decide which cell to query next. This
is a conventional setup in active search, where an impor-
tant means to accomplish such learning is by introducing
a model fto predict a mapping between (in our instance)
a grid cell and the associated label (whether it contains a
target object) [10, 15, 13]. However, in many domains of
policy networkfeature extractiongrid prediction
query/search top cellgrid state (o)updatereducesamplesearch budget (B)Figure 2: An overview of the VAS framework.
interest, such as most visual domains of the kind we con-
sider, the query budget Cand the number of grid cells N
are very small compared to the dimension of the input x,
far too small to learn a meaningful prediction f. Instead,
we suppose that we have a dataset of tasks (aerial images)
for which we have labeled whether each of the grid cells
contains the target object. Let this dataset be denoted by
D={(xi, yi)}, with each xi=(x(1)
i, x(2)
i, . . . , x(N)
i)the
task image and yi=(y(1)
i, y(2)
i, . . . , y(N)
i)its correspond-
ing grid cell labels. Then, at decision (or inference) time,
we observe the task aerial image x, including its partition
into the grid cells, and choose queries {qt}sequentially to
maximize U(x;{qt}).
We consider two variations of the model above. In the
first, each instance (xi, yi)in the training data D, as well
as(x, y)at decision time (when yis unobserved before
queries) are generated i.i.d. from the same distribution. In
the second variation, while instances (x, y)are still i.i.d. at
decision time, their distribution can be different from that
of the training data D. The latter variation falls within the
broader category of test-time adaptation (TTA) settings, but
with the special structure pertinent to our model above.
4. Solution Approach
Visual active search over the area defined by xand its
constituent grid cells is a dynamic decision problem. As
such, we model it as a budget-constrained episodic Markov
decision process (MDP), where the search budget Cis de-
fined for each instance xat decision time . In this MDP,
the actions are simply choices over which grid cell to query
next; we denote the set of grids by A={1, . . . , N}. Since
in our model there is never any value to query a grid cell
more than once, we restrict actions available at each step
to be only grids that have not yet been queried (in princi-
ple, this restriction can also be learned). Policy network
inputs include: 1) the overall input x, which is crucial in
providing the broad perspective on each search problem, 2)
outcomes of past search queries o(we detail our representa-
tion of this presently), and 3) remaining budget B≤C. State
transition simply updates the remaining budget and adds the
outcome of the latest search query to state. Finally, an im-
mediate reward for query a grid cell jisR(x, o, j)=y(j).
We represent outcomes of search query history oas fol-
3
feature extraction networkgrid prediction networkinput image (x)
tilegrid state[o(1), o(2),…, o(N)]ResNet 34 (frozen)tileNx14x14Dx14x141x1convflatten1x1conv1x14x14
updateremaining query budget (B)
…step 1grid probabilitiesMLPstep 2step KFigure 3: Our V AS policy network architecture, showing the grid probabilities at three different steps.
lows. Each element of ocorresponds to a grid cell j, so that
o=(o(1), . . . , o(N)).o(j)=0ifjhas not been previously
queried. If grid cell jhas been previously queried,
o(j)← /leftr⫯g⊸tl⫯ne{1,ify(j)=1
−1,ify(j)=0.(2)
Armed with this MDP problem representation, we next
describe our proposed deep reinforcement learning ap-
proach for learning a search policy that makes use of a
dataset Dof past search tasks. Specifically, we use the RE-
INFORCE policy gradient algorithm [28] to directly learn
a search policy ψ(x, o, B ;θ)where θare the parameters of
the policy that we learn. Specifically, we maximize the fol-
lowing objective function:
∇J(θ)=M
i=1Ti
t=11∑t≥0c(qt−1,qt)≤C∇logψθ(ai
t∣xi, oi
t, Bi
t)Ri
t
(3)
Where Mis the number of example search task seen during
training and Rtis the discounted cumulative reward defined
asRt=∑T
k=tγk−tRkwith a discount factor γ∈[0,1].