Efficient Lifelong Learning with A-GEM

2019 • Catastrophic Forgetting • Continual Learning • ICLR 2019 • Lifelong Learning • AI • CV • CL • ICLR

08 Jan 2019

Contributions

• A new (and more realistic) evaluation protocol for lifelong learning where each data point is observed just once and a disjoint set of tasks are used for training and validation.

• A new metric that focuses on the efficiency of the models - in terms of sample complexity and computational (and memory) costs.

• Modification of Gradient Episodic Memory ie GEM which reduces the computational overhead of GEM without compromising on the results.

• Empirical validation that using task descriptors help lifelong learning models and improve their few-shot learning capabilities.

• Link to the paper

• Link to the code

Learning Protocol

• Two group of datasets - one for training and evaluation (D^EV) and other for cross validation (D^CV).

• Data can be sampled multiple times for cross-validation dataset but only once from the training dataset.

• Each group of dataset (say D^EV or D^CV) is a list of task-specific datasets D_k (k is the task index).

• Each sample in D_k is of the form (x, t, y) where x is the data, t is the task descriptor and y is the output.

• D_k contains B^k minibatches of data.

Metrics

Accuracy

• a_k,i,j = accuracy on test task j after training on ith minibatch of training task k.

• A_k = mean over all j = 1 to k (a_{k, Bk, j}) ie train the model on data for task k and then test it on all the tasks.

Forgetting Measure

• f_j^k = forgetting on task j after training on all minibatches upto task k.

• f_j^k = max over all l = 1 to k-1 (a_{l, Blj} - a_{k, Bkj})

• Forgetting = F_k = mean over all j = 1 to k-1 (f_j^k)

LCA - Learning Curve Area

• Z_b = average b shot performance where b is the minibatch number.

• Z_b = mean over all k = 0 to T (a_{k, b, k})

• LCA_β = mean over all b = 0 to β (Z_b)

• One special case is LCA₀ which is the forward transfer performance or performance on the unseen task.

• In experiments, β is kept small as we want the model to learn from few examples.

Model

• GEM has been shown to be very effective in single epoch setting but introduces a very high computational overhead.

• Average GEM (AGEM) reduces this overhead by sampling (and using) only some examples from the episodic memory instead of using all the examples.

• While GEM provides better guarantees in terms of worst-case forgetting, AGEM provides better guarantees in terms of average accuracy.

Joint Embedding Model Using Compositional Task Descriptors

• Compositional Task Descriptors are used to speed training on the subsequent tasks.

• A matrix specifying the attribute value of objects (to be recognized in the task) are used.

• A joint-embedding space between image features and attribute embeddings is learned.

Experiments

Datasets

• Permuted MNIST

• Split CIFAR

• Split CUB

• Split AWA

Setup

• Integer task descriptors for MNIST and CIFAR and class attributes as descriptors for CUB and AWA

• Baselines include GEM, iCaRL, Elastic Weight Consolidation, Progressive Neural Networks etc.

Results

• AGEM outperforms other models on all the datasets expect MNIST where the Progressive Neural Networks lead. One reason could be that MNIST has a large number of training examples per task. But Progressive Neural Networks lead to bad utilization of capacity.

• While AGEM and GEM have similar performance, GEM has a much higher computational and memory overhead.

• Use of task descriptors improves the accuracy for all the models.

• It seems that AGEM offers a good tradeoff between average accuracy performance and efficiency - in terms of sample efficiency, memory requirements and computational costs.