Scaling model capacity has been vital in the success of deep learning.

Modularity is a compelling solution to continual learning (CL), the problem of modeling sequences of related tasks. Learning and then composing modules to solve different tasks provides an abstraction to address the principal challenges of CL including catastrophic forgetting, backward and forward transfer across tasks, and sub-linear model growth. We introduce local module composition (LMC), an approach to modular CL where each module is provided a local structural component that estimates a module's relevance to the input. Dynamic module composition is performed layer-wise based on local relevance scores. We demonstrate that agnosticity to task identities (IDs) arises from (local) structural learning that is module-specific as opposed to the task-and/or model-specific as in previous works, making LMC applicable to more CL settings compared to previous works.

Scaling model capacity has been vital in the success of deep learning. For a typical network, necessary compute resources and training time grow dramatically with model size. Conditional computation is a promising way to increase the number of parameters with a relatively small increase in resources. We propose a training algorithm that flexibly chooses neural modules based on the data to be processed. Both the decomposition and modules are learned end-to-end. In contrast to existing approaches, training does not rely on regularization to enforce diversity in module use. We apply modular networks both to image recognition and language modeling tasks, where we achieve superior performance compared to several baselines. Introspection reveals that modules specialize in interpretable contexts.

Scaling model capacity has been vital in the success of deep learning.

Our approach enables efficient optimization and sharing across modules. R1:"motivate their work very well, is technical sound," R2:"idea seems to be new," R3:"very important problem, better We will address reviewers' comments as follows. Theoretical grounding: the paper is not well grounded in neural network theory. R2 also asks "Why a dot product for the weighting?" But weighting itself indicates multiplication. R2 has not provided an alternative way for weighting. Meta-learning is attracting, Comparison to state-of-the-art (e.g. R2 also has not provided a reference on multi-task RL for us to compare. How to adopt it in multi-task RL is an interesting direction to study, but it is out of the scope of our paper. While R2 complains about our writing, other reviewers all have positive feedback: "I liked to read the paper, For the routing network, the inputs are the same as the policy including both states and task embedding.

Modularity is a compelling solution to continual learning (CL), the problem of modeling sequences of related tasks. Learning and then composing modules to solve different tasks provides an abstraction to address the principal challenges of CL including catastrophic forgetting, backward and forward transfer across tasks, and sub-linear model growth. We introduce local module composition (LMC), an approach to modular CL where each module is provided a local structural component that estimates a module's relevance to the input. Dynamic module composition is performed layer-wise based on local relevance scores. We demonstrate that agnosticity to task identities (IDs) arises from (local) structural learning that is module-specific as opposed to the task- and/or model-specific as in previous works, making LMC applicable to more CL settings compared to previous works.

The paper is concerned with conditional computation, which is an interesting topic yet at early stages of research, and as such one that requires much research and investigation. The paper proposes a latent-variable approach to constructing modular networks, modeling the choice of processing modules in a layer as a discrete latent variable. A modular network is composed of L modular layers, each comprised of M modules and a controller. Each module is a function (standard layer) f_i(x; \theta_i). The controller accepts the input, chooses K of the M modules to process the input, and outputs the as the module output. Modular layers can be stacked, or placed anywhere inside a standard network.

Multivariate time series have many applications, from healthcare and meteorology to life science. Although deep learning models have shown excellent predictive performance for time series, they have been criticised for being "black-boxes" or non-interpretable. This paper proposes a novel modular neural network model for multivariate time series prediction that is interpretable by construction. A recurrent neural network learns the temporal dependencies in the data while an attention-based feature selection component selects the most relevant features and suppresses redundant features used in the learning of the temporal dependencies. A modular deep network is trained from the selected features independently to show the users how features influence outcomes, making the model interpretable. Experimental results show that this approach can outperform state-of-the-art interpretable Neural Additive Models (NAM) and variations thereof in both regression and classification of time series tasks, achieving a predictive performance that is comparable to the top non-interpretable methods for time series, LSTM and XGBoost.

Training robots with reinforcement learning (RL) typically involves heavy interactions with the environment, and the acquired skills are often sensitive to changes in task environments and robot kinematics. Transfer RL aims to leverage previous knowledge to accelerate learning of new tasks or new body configurations. However, existing methods struggle to generalize to novel robot-task combinations and scale to realistic tasks due to complex architecture design or strong regularization that limits the capacity of the learned policy. We propose Policy Stitching, a novel framework that facilitates robot transfer learning for novel combinations of robots and tasks. Our key idea is to apply modular policy design and align the latent representations between the modular interfaces. Our method allows direct stitching of the robot and task modules trained separately to form a new policy for fast adaptation. Our simulated and real-world experiments on various 3D manipulation tasks demonstrate the superior zero-shot and few-shot transfer learning performances of our method. Our project website is at: http://generalroboticslab.com/PolicyStitching/ .