In this paper, we propose a highly parameter-efficient approach to scaling pre-trained language models (PLMs) to a deeper model depth. Unlike prior work that shares all parameters or uses extra blocks, we design a more capable parameter-sharing architecture based on matrix product operator (MPO). MPO decomposition can reorganize and factorize the information of a parameter matrix into two parts: the major part that contains the major information (central tensor) and the supplementary part that only has a small proportion of parameters (auxiliary tensors). Based on such a decomposition, our architecture shares the central tensor across all layers for reducing the model size and meanwhile keeps layer-specific auxiliary tensors (also using adapters) for enhancing the adaptation flexibility. To improve the model training, we further propose a stable initialization algorithm tailored for the MPO-based architecture. Extensive experiments have demonstrated the effectiveness of our proposed model in reducing the model size and achieving highly competitive performance.

Unsupervised lifelong learning refers to the ability to learn over time while memorizing previous patterns without supervision. Although great progress has been made in this direction, existing work often assumes strong prior knowledge about the incoming data (e.g., knowing the class boundaries), which can be impossible to obtain in complex and unpredictable environments. In this paper, motivated by real-world scenarios, we propose a more practical problem setting called online self-supervised lifelong learning without prior knowledge. The proposed setting is challenging due to the non-iid and single-pass data, the absence of external supervision, and no prior knowledge. To address the challenges, we propose Self-Supervised ContrAstive Lifelong LEarning without Prior Knowledge (SCALE) which can extract and memorize representations on the fly purely from the data continuum. SCALE is designed around three major components: a pseudo-supervised contrastive loss, a self-supervised forgetting loss, and an online memory update for uniform subset selection. All three components are designed to work collaboratively to maximize learning performance. We perform comprehensive experiments of SCALE under iid and four non-iid data streams. The results show that SCALE outperforms the state-of-the-art algorithm in all settings with improvements up to 3.83%, 2.77% and 5.86% in terms of kNN accuracy on CIFAR-10, CIFAR-100, and TinyImageNet datasets.

Online class-incremental continual learning is a specific task of continual learning. It aims to continuously learn new classes from data stream and the samples of data stream are seen only once, which suffers from the catastrophic forgetting issue, i.e., forgetting historical knowledge of old classes. Existing replay-based methods effectively alleviate this issue by saving and replaying part of old data in a proxy-based or contrastive-based replay manner. Although these two replay manners are effective, the former would incline to new classes due to class imbalance issues, and the latter is unstable and hard to converge because of the limited number of samples. In this paper, we conduct a comprehensive analysis of these two replay manners and find that they can be complementary. Inspired by this finding, we propose a novel replay-based method called proxy-based contrastive replay (PCR). The key operation is to replace the contrastive samples of anchors with corresponding proxies in the contrastive-based way. It alleviates the phenomenon of catastrophic forgetting by effectively addressing the imbalance issue, as well as keeps a faster convergence of the model. We conduct extensive experiments on three real-world benchmark datasets, and empirical results consistently demonstrate the superiority of PCR over various state-of-the-art methods.

With the rising emergence of decentralized and opportunistic approaches to machine learning, end devices are increasingly tasked with training deep learning models on-devices using crowd-sourced data that they collect themselves. These approaches are desirable from a resource consumption perspective and also from a privacy preservation perspective. When the devices benefit directly from the trained models, the incentives are implicit - contributing devices' resources are incentivized by the availability of the higher-accuracy model that results from collaboration. However, explicit incentive mechanisms must be provided when end-user devices are asked to contribute their resources (e.g., computation, communication, and data) to a task performed primarily for the benefit of others, e.g., training a model for a task that a neighbor device needs but the device owner is uninterested in. In this project, we propose a novel blockchain-based incentive mechanism for completely decentralized and opportunistic learning architectures. We leverage a smart contract not only for providing explicit incentives to end devices to participate in decentralized learning but also to create a fully decentralized mechanism to inspect and reflect on the behavior of the learning architecture.

Knowledge distillation is a method of transferring the knowledge from a complex deep neural network (DNN) to a smaller and faster DNN, while preserving its accuracy. Recent variants of knowledge distillation include teaching assistant distillation, curriculum distillation, mask distillation, and decoupling distillation, which aim to improve the performance of knowledge distillation by introducing additional components or by changing the learning process. Teaching assistant distillation involves an intermediate model called the teaching assistant, while curriculum distillation follows a curriculum similar to human education. Mask distillation focuses on transferring the attention mechanism learned by the teacher, and decoupling distillation decouples the distillation loss from the task loss. Overall, these variants of knowledge distillation have shown promising results in improving the performance of knowledge distillation.

Artificial neural networks are promising for general function approximation but challenging to train on non-independent or non-identically distributed data due to catastrophic forgetting. The experience replay buffer, a standard component in deep reinforcement learning, is often used to reduce forgetting and improve sample efficiency by storing experiences in a large buffer and using them for training later. However, a large replay buffer results in a heavy memory burden, especially for onboard and edge devices with limited memory capacities. We propose memory-efficient reinforcement learning algorithms based on the deep Q-network algorithm to alleviate this problem. Our algorithms reduce forgetting and maintain high sample efficiency by consolidating knowledge from the target Q-network to the current Q-network. Compared to baseline methods, our algorithms achieve comparable or better performance in both feature-based and image-based tasks while easing the burden of large experience replay buffers.

Deep neural networks (DNNs) have substantial computational and memory requirements, and the compilation of its computational graphs has a great impact on the performance of resource-constrained (e.g., computation, I/O, and memory-bound) edge computing systems. While efficient execution of their computational graph requires an effective scheduling algorithm, generating the optimal scheduling solution is a challenging NP-hard problem. Furthermore, the complexity of scheduling DNN computational graphs will further increase on pipelined multi-core systems considering memory communication cost, as well as the increasing size of DNNs. Using the synthetic graph for the training dataset, this work presents a reinforcement learning (RL) based scheduling framework RESPECT, which learns the behaviors of optimal optimization algorithms and generates near-optimal scheduling results with short solving runtime overhead. Our framework has demonstrated up to $\sim2.5\times$ real-world on-chip inference runtime speedups over the commercial compiler with ten popular ImageNet models deployed on the physical Coral Edge TPUs system. Moreover, compared to the exact optimization methods, the proposed RL scheduling improves the scheduling optimization runtime by up to 683$\times$ speedups compared to the commercial compiler and matches the exact optimal solutions with up to 930$\times$ speedups. Finally, we perform a comprehensive generalizability test, which demonstrates RESPECT successfully imitates optimal solving behaviors from small synthetic graphs to large real-world DNNs computational graphs.

Despite the potential benefits of machine learning (ML) in high-risk decision-making domains, the deployment of ML is not accessible to practitioners, and there is a risk of discrimination. To establish trust and acceptance of ML in such domains, democratizing ML tools and fairness consideration are crucial. In this paper, we introduce FairPilot, an interactive system designed to promote the responsible development of ML models by exploring a combination of various models, different hyperparameters, and a wide range of fairness definitions. We emphasize the challenge of selecting the ``best" ML model and demonstrate how FairPilot allows users to select a set of evaluation criteria and then displays the Pareto frontier of models and hyperparameters as an interactive map. FairPilot is the first system to combine these features, offering a unique opportunity for users to responsibly choose their model.

Training generalist agents is difficult across several axes, requiring us to deal with high-dimensional inputs (space), long horizons (time), and multiple and new tasks. Recent advances with architectures have allowed for improved scaling along one or two of these dimensions, but are still prohibitive computationally. In this paper, we propose to address all three axes by leveraging Language to Control Diffusion models as a hierarchical planner conditioned on language (LCD). We effectively and efficiently scale diffusion models for planning in extended temporal, state, and task dimensions to tackle long horizon control problems conditioned on natural language instructions. We compare LCD with other state-of-the-art models on the CALVIN language robotics benchmark and find that LCD outperforms other SOTA methods in multi task success rates while dramatically improving computational efficiency with a single task success rate (SR) of 88.7% against the previous best of 82.6%. We show that LCD can successfully leverage the unique strength of diffusion models to produce coherent long range plans while addressing their weakness at generating low-level details and control. We release our code and models at https://github.com/ezhang7423/language-control-diffusion.

Artificial intelligence (AI) has received significant attention in recent years, primarily due to breakthroughs in game playing, computer vision, and natural language processing that captured the imagination of the scientific community and the public at large. Many businesses, industries, and academic disciplines are now contemplating the application of AI to their own challenges. The federal government in the US and other countries have also invested significantly in advancing AI research and created funding initiatives and programs to promote greater collaboration across multiple communities. Some of the investment examples in the US include the establishment of the National AI Initiative Office, the launch of the National AI Research Resource Task Force, and more recently, the establishment of the National AI Advisory Committee. In 2021 INFORMS and ACM SIGAI joined together with the Computing Community Consortium (CCC) to organize a series of three workshops. The objective for this workshop series is to explore ways to exploit the synergies of the AI and Operations Research (OR) communities to transform decision making.