220.

During a first phase for each of a plurality of learning iterations for the L2TL architecture 200, the L2TL architecture 200 applies gradient decent-based optimization to learn the encoder weights Ω 210, the source classifier layer weights ζ_S202, and the target classifier layer weights ζ_T204 that minimize the loss function 201, and determines the coefficients 206 by sampling actions of the policy model 209. For instance, applying gradient descent-based optimization to learn the adaptable weights Ω 210, ζ_S202, and ζ_Tthat minimizes the loss function 201 may be represented by the following equation.

{circumflex over (Ω)},{circumflex over (ζ)}_S,{circumflex over (ζ)}_T=argmin_Ω,ζ_S_,ζ_T [Image Omitted] ({circumflex over (Φ)};Ω,ζ_S,ζ_T) (2)

During the first phase, the policy model 209 remains fixed. Although most batches would contain relevant samples from the source dataset 104, the loss function 201 might be skewed if most of source dataset samples in a batch are irrelevant (and would ideally get lower weights). To ease this problem, the first phase may use larger batch sizes at each iteration and dynamically select the most relevant examples therefrom. At each iteration, the L2TL architecture samples a training batch of size M_S·B_S, and uses the top B_Sof those with the highest weights for training updates during that iteration. This approach also yields computational benefits as the gradients would not be computed for most source dataset samples until convergence.

FIG. 2 shows the L2TL architecture 200 having a training objective during the first phase of the learning iteration for each of the source and target datasets 104, 106 that includes minimizing gradients between respective inputs x of the source and target datasets 104, 106 and their corresponding labels/outputs y. Here, the encoding network 152 encodes input features (e.g., images) from the source dataset 104 and the source classifier layer 154 (also referred to as “source decision layer”) uses the encoded input features to output a confidence score, whereby the training objective determines a source dataset classification loss (which corresponds to the first term L_S(f_S(x_j;Ω,ζ_S),y_j) in Eq. (1) of the loss function 201) between the confidence score for the input features and the corresponding labels/outputs y. The source dataset classification loss is multiplied by the importance weights 212, which correspond to λ(x_j,y_j;Φ) in Eq. (1) of the loss function 201, before a summing function 203 aggregates a target dataset classification loss (which corresponds to the second term L_T(f_T(x′_k;Ω,ζ_T),y′_k) in Eq. 1 of the loss function 201) with the product of the source dataset classification loss and the importance weights 212 to determine aggregate gradients for the loss function 201. Based on the aggregated gradients of the loss function of each iteration, the L2TL architecture 200 determines/updates each of the encoder weights Ω 210, source classifier layer weights ζ_S202, and the target classifier layer weights ζ_T204 until convergence of the source dataset 103.

The encoding network 152 also encodes input features (e.g., images) from the target dataset 106 and the target classifier layer 156 (also referred to as “source decision layer”) uses the encoded features to output a confidence score, whereby the training objective determines determines the target dataset classification loss, L_T(f_T(x′_k;Ω,ζ_T),y′_k), between the confidence score for the input features and the corresponding labels/outputs y for the target dataset. As mentioned above, the summing function 203 aggregates the target dataset classification loss with the product of the source dataset classification loss and the importance weights 208 to determine the aggregate gradients for the loss function 201.

During the second phase for each of a plurality of learning iterations for the L2TL architecture 200, the L2TL architecture 200 determines the policy weight Φ 208 that maximizes an evaluation metric, R, 220 for the loss function 201. Here, the L2TL architecture 200 determines the policy weight Φ 208 that maximizes the evaluation metric, R, 220 using the encoder weights Ω 210 and the target classifier layer weights ζ_T204 learned during the first phase. As used herein, the evaluation metric (also referred to as ‘evaluation metric’) 220 for the loss function quantifies performance of the deep learning model 150 (i.e., the encoder network 152 and the target classification layer 154) on the target evaluation dataset 106′. In some examples, the target evaluation dataset 106′ includes a subset of data samples (e.g., images) in the target dataset 106 that were not previously seen by the deep learning model 150 during the first phase of training. Accordingly, determining the policy weight Φ 208 includes optimizing the policy weight 208 to maximize the evaluation metric R_D′T220 on the target evaluation dataset 106′ may be represented as follows.

max
Φ

R

D
T
′

(

Ω
^

,

ζ
S

^

,

ζ
T

^

;
Φ

)

.

(
3
)

Where D′_Tis the unseen target evaluation dataset 106′ that may include a subset of data samples (input features x, output labels y) in the target dataset 106 that were not previously seen by the deep learning model 150 during the first phase. While the policy model 209 is fixed during the first phase, during the second phase, the L2TL architecture 200 trains the policy model 209 using policy gradient on the target evaluation dataset 106′ to compute a reward that maximizes the evaluation metric 220 using Eq. 3. Accordingly, L2TL architecture 200 determines/updates the policy weight Φ