Table A1.

The pseudocode of the model algorithm.

0	Requie: a Stepsize
1	$Require: {\beta }_1,{\beta }_2$ Exponential decay rates for the moment estimates
2	$Require: {\theta }_0$ : Initial parameter vector
3	$m_0$ ← 0 (Initial 1st moment vector)
4	$v_0$ ← 0 (Initial 2nd moment vector)
5	t ←0 : (Initial time step)
6	$While{\theta }_t$ not converged do
7	t ← t+1
8	Feature_vector ← BackboneNetwork(x) (Extract feature vector using CNN )
9	DropoutFeatureVector ← Dropout(Feature_vector) (Dropout layer)
10	${\widehat{y}}_{coarse}$← fullyConnectedNetwork1(DropoutFeatureVector) (Predict Coarse Classification)
11	${\widehat{y}}_{fine}$← fullyConnectedNetwork2(DropoutFeatureVector) (Predict Fine Classificatoin)
12	$Loss \leftarrow C{E}_{coarse}\left(y_{coarse},{\widehat{y}}_{coarse}\right)+C{E}_{fine}\left(y_{fine},{\widehat{y}}_{fine}\right)$ (Compute Loss)
13	$g_t$ ←${\nabla }_{\theta }Loss$ (Get gradients w.r.t. stochastic objective at timestep t)
14	$m_t$ ←${\beta }_1\ast \backslash m_{t-1}+\left(1-{\beta }_1\right)\ast g_t$ (Update biased first moment estimate)
15	$v_t$ ←${\beta }_2\ast v_{t-1}+\left(1-{\beta }_2\right)\ast g_t^2$ (Update biased second raw moment estimate)
16	$\widehat{v_t}$←$v_t/\left(1-{\beta }_1^2\right)$ (Compute bias-correced first moment estimate)
17	${\theta }_t$←${\theta }_{t-1}-\alpha \ast \widehat{m_t}/\left(\sqrt{ \widehat{v_t}}+\epsilon \right)$(Update parameters)
18	end while
19	$return{\theta }_t$ (Resulting parameters)