Large-scale pre-trained language models (LLMs) have demonstrated exceptional performance in various natural language processing (NLP) tasks. However, the massive size of these models poses huge challenges for their deployment in real-world applications. While numerous model compression techniques have been proposed, most of them are not well-suited for achieving extreme model compression when there is a significant gap in model scale. In this paper, we introduce a novel compression paradigm called Retrieval-based Knowledge Transfer (RetriKT), which effectively transfers the knowledge of LLMs to extremely small-scale models (e.g., 1%). In particular, our approach extracts knowledge from LLMs to construct a knowledge store, from which the small-scale model can retrieve relevant information and leverage it for effective inference. To improve the quality of the model, soft prompt tuning and Proximal Policy Optimization (PPO) reinforcement learning techniques are employed. Extensive experiments are conducted on low-resource tasks from SuperGLUE and GLUE benchmarks. The results demonstrate that the proposed approach significantly enhances the performance of small-scale models by leveraging the knowledge from LLMs.
The Option Keyboard (OK) was recently proposed as a method for transferring behavioral knowledge across tasks. OK transfers knowledge by adaptively combining subsets of known behaviors using Successor Features (SFs) and Generalized Policy Improvement (GPI). However, it relies on hand-designed state-features and task encodings which are cumbersome to design for every new environment. In this work, we propose the "Successor Features Keyboard" (SFK), which enables transfer with discovered state-features and task encodings. To enable discovery, we propose the "Categorical Successor Feature Approximator" (CSFA), a novel learning algorithm for estimating SFs while jointly discovering state-features and task encodings. With SFK and CSFA, we achieve the first demonstration of transfer with SFs in a challenging 3D environment where all the necessary representations are discovered. We first compare CSFA against other methods for approximating SFs and show that only CSFA discovers representations compatible with SF&GPI at this scale. We then compare SFK against transfer learning baselines and show that it transfers most quickly to long-horizon tasks.
In recent years, transfer learning has garnered significant attention in the machine learning community. Its ability to leverage knowledge from related studies to improve generalization performance in a target study has made it highly appealing. This paper focuses on investigating the transfer learning problem within the context of nonparametric regression over a reproducing kernel Hilbert space. The aim is to bridge the gap between practical effectiveness and theoretical guarantees. We specifically consider two scenarios: one where the transferable sources are known and another where they are unknown. For the known transferable source case, we propose a two-step kernel-based estimator by solely using kernel ridge regression. For the unknown case, we develop a novel method based on an efficient aggregation algorithm, which can automatically detect and alleviate the effects of negative sources. This paper provides the statistical properties of the desired estimators and establishes the minimax optimal rate. Through extensive numerical experiments on synthetic data and real examples, we validate our theoretical findings and demonstrate the effectiveness of our proposed method.
Sampling biases can cause distribution shifts between train and test datasets for supervised learning tasks, obscuring our ability to understand the generalization capacity of a model. This is especially important considering the wide adoption of pre-trained foundational neural networks -- whose behavior remains poorly understood -- for transfer learning (TL) tasks. We present a case study for TL on the Sentiment140 dataset and show that many pre-trained foundation models encode different representations of Sentiment140's manually curated test set $M$ from the automatically labeled training set $P$, confirming that a distribution shift has occurred. We argue training on $P$ and measuring performance on $M$ is a biased measure of generalization. Experiments on pre-trained GPT-2 show that the features learnable from $P$ do not improve (and in fact hamper) performance on $M$. Linear probes on pre-trained GPT-2's representations are robust and may even outperform overall fine-tuning, implying a fundamental importance for discerning distribution shift in train/test splits for model interpretation.
Fine-tuning a pre-trained deep neural network has become a successful paradigm in various machine learning tasks. However, such a paradigm becomes particularly challenging with tabular data when there are discrepancies between the feature sets of pre-trained models and the target tasks. In this paper, we propose TabToken, a method aims at enhancing the quality of feature tokens (i.e., embeddings of tabular features). TabToken allows for the utilization of pre-trained models when the upstream and downstream tasks share overlapping features, facilitating model fine-tuning even with limited training examples. Specifically, we introduce a contrastive objective that regularizes the tokens, capturing the semantics within and across features. During the pre-training stage, the tokens are learned jointly with top-layer deep models such as transformer. In the downstream task, tokens of the shared features are kept fixed while TabToken efficiently fine-tunes the remaining parts of the model. TabToken not only enables knowledge transfer from a pre-trained model to tasks with heterogeneous features, but also enhances the discriminative ability of deep tabular models in standard classification and regression tasks.
A major concern of deep learning models is the large amount of data that is required to build and train them, much of which is reliant on sensitive and personally identifiable information that is vulnerable to access by third parties. Ideas of using the quantum internet to address this issue have been previously proposed, which would enable fast and completely secure online communications. Previous work has yielded a hybrid quantum-classical transfer learning scheme for classical data and communication with a hub-spoke topology. While quantum communication is secure from eavesdrop attacks and no measurements from quantum to classical translation, due to no cloning theorem, hub-spoke topology is not ideal for quantum communication without quantum memory. Here we seek to improve this model by implementing a decentralized ring topology for the federated learning scheme, where each client is given a portion of the entire dataset and only performs training on that set. We also demonstrate the first successful use of quantum weights for quantum federated learning, which allows us to perform our training entirely in quantum.
In many settings, such as scientific inference, optimization, and transfer learning, the learner has a well-defined objective, which can be treated as estimation of a target parameter, and no intrinsic interest in characterizing the entire data-generating process. Usually, the learner must also contend with additional sources of uncertainty or variables -- with nuisance parameters. Bayesian active learning, or sequential optimal experimental design, can straightforwardly accommodate the presence of nuisance parameters, and so is a natural active learning framework for such problems. However, the introduction of nuisance parameters can lead to bias in the Bayesian learner's estimate of the target parameters, a phenomenon we refer to as negative interference. We characterize the threat of negative interference and how it fundamentally changes the nature of the Bayesian active learner's task. We show that the extent of negative interference can be extremely large, and that accurate estimation of the nuisance parameters is critical to reducing it. The Bayesian active learner is confronted with a dilemma: whether to spend a finite acquisition budget in pursuit of estimation of the target or of the nuisance parameters. Our setting encompasses Bayesian transfer learning as a special case, and our results shed light on the phenomenon of negative transfer between learning environments.
Dataset distillation plays a crucial role in creating compact datasets with similar training performance compared with original large-scale ones. This is essential for addressing the challenges of data storage and training costs. Prevalent methods facilitate knowledge transfer by matching the gradients, embedding distributions, or training trajectories of synthetic images with those of the sampled original images. Although there are various matching objectives, currently the strategy for selecting original images is limited to naive random sampling. We argue that random sampling overlooks the evenness of the selected sample distribution, which may result in noisy or biased matching targets. Besides, the sample diversity is also not constrained by random sampling. Additionally, current methods predominantly focus on single-dimensional matching, where information is not fully utilized. To address these challenges, we propose a novel matching strategy called Dataset Distillation by Bidirectional REpresentAtive Matching (DREAM+), which selects representative original images for bidirectional matching. DREAM+ is applicable to a variety of mainstream dataset distillation frameworks and significantly reduces the number of distillation iterations by more than 15 times without affecting performance. Given sufficient training time, DREAM+ can further improve the performance and achieve state-of-the-art results. We have released the code at github.com/NUS-HPC-AI-Lab/DREAM+.
The ubiquitous availability of smartphones and smartwatches with integrated inertial measurement units (IMUs) enables straightforward capturing of human activities. For specific applications of sensor based human activity recognition (HAR), however, logistical challenges and burgeoning costs render especially the ground truth annotation of such data a difficult endeavor, resulting in limited scale and diversity of datasets. Transfer learning, i.e., leveraging publicly available labeled datasets to first learn useful representations that can then be fine-tuned using limited amounts of labeled data from a target domain, can alleviate some of the performance issues of contemporary HAR systems. Yet they can fail when the differences between source and target conditions are too large and/ or only few samples from a target application domain are available, each of which are typical challenges in real-world human activity recognition scenarios. In this paper, we present an approach for economic use of publicly available labeled HAR datasets for effective transfer learning. We introduce a novel transfer learning framework, Cross-Domain HAR, which follows the teacher-student self-training paradigm to more effectively recognize activities with very limited label information. It bridges conceptual gaps between source and target domains, including sensor locations and type of activities. Through our extensive experimental evaluation on a range of benchmark datasets, we demonstrate the effectiveness of our approach for practically relevant few shot activity recognition scenarios. We also present a detailed analysis into how the individual components of our framework affect downstream performance.
Fault diagnosis of rotating machinery plays a important role for the safety and stability of modern industrial systems. However, there is a distribution discrepancy between training data and data of real-world operation scenarios, which causing the decrease of performance of existing systems. This paper proposed a transfer learning based method utilizing acoustic and vibration signal to address this distribution discrepancy. We designed the acoustic and vibration feature fusion MAVgram to offer richer and more reliable information of faults, coordinating with a DNN-based classifier to obtain more effective diagnosis representation. The backbone was pre-trained and then fine-tuned to obtained excellent performance of the target task. Experimental results demonstrate the effectiveness of the proposed method, and achieved improved performance compared to STgram-MFN.
Traffic prediction represents one of the crucial tasks for smartly optimizing the mobile network. Recently, Artificial Intelligence (AI) has attracted attention to solve this problem thanks to its ability in cognizing the state of the mobile network and make intelligent decisions. Research on this topic has concentrated on making predictions in a centralized fashion, i.e., by collecting data from the different network elements and process them in a cloud center. This translates into inefficiencies due to the large amount of data transmissions and computations required, leading to high energy consumption. In this work, we investigate a fully decentralized AI solution for mobile traffic prediction that allows data to be kept locally, reducing energy consumption through collaboration among the base station sites. To do so, we propose a novel prediction framework based on edge computing and Deep Transfer Learning (DTL) techniques, using datasets obtained at the edge through a large measurement campaign. Two main Deep Learning architectures are designed based on Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) and tested under different training conditions. Simulation results show that the CNN architectures outperform the RNNs in accuracy and consume less energy. In both scenarios, DTL contributes to an accuracy enhancement in 85% of the examined cases compared to their stand-alone counterparts. Additionally, DTL significantly reduces computational complexity and energy consumption during training, resulting in a reduction of the energy footprint by 60% for CNNs and 90% for RNNs. Finally, two cutting-edge eXplainable Artificial Intelligence techniques are employed to interpret the derived learning models.
There has been a significant surge of interest in visually grounded grammar induction in recent times. While a variety of models have been developed for the task and have demonstrated impressive performance, they have not been evaluated on text domains that are different from the training domain, so it is unclear if the improvements brought by visual groundings are transferable. Our study aims to fill this gap and assess the degree of transferability. We start by extending VC-PCFG (short for Visually-grounded Compound PCFG~\citep{zhao-titov-2020-visually}) in such a way that it can transfer across text domains. We consider a zero-shot transfer learning setting where a model is trained on the source domain and is directly applied to target domains, without any further training. Our experimental results suggest that: the benefits from using visual groundings transfer to text in a domain similar to the training domain but fail to transfer to remote domains. Further, we conduct data and result analysis; we find that the lexicon overlap between the source domain and the target domain is the most important factor in the transferability of VC-PCFG.
This paper proposes an approach of Ladder Bottom-up Convolutional Bidirectional Variational Autoencoder (LCBVAE) architecture for the encoder and decoder, which is trained on the image translation of the dotted Arabic expiration dates by reconstructing the Arabic dotted expiration dates into filled-in expiration dates. We employed a customized and adapted version of Convolutional Recurrent Neural Network CRNN model to meet our specific requirements and enhance its performance in our context, and then trained the custom CRNN model with the filled-in images from the year of 2019 to 2027 to extract the expiration dates and assess the model performance of LCBVAE on the expiration date recognition. The pipeline of (LCBVAE+CRNN) can be then integrated into an automated sorting systems for extracting the expiry dates and sorting the products accordingly during the manufacture stage. Additionally, it can overcome the manual entry of expiration dates that can be time-consuming and inefficient at the merchants. Due to the lack of the availability of the dotted Arabic expiration date images, we created an Arabic dot-matrix True Type Font (TTF) for the generation of the synthetic images. We trained the model with unrealistic synthetic dates of 60,000 images and performed the testing on a realistic synthetic date of 3000 images from the year of 2019 to 2027, represented as yyyy/mm/dd. In our study, we demonstrated the significance of latent bottleneck layer with improving the generalization when the size is increased up to 1024 in downstream transfer learning tasks as for image translation. The proposed approach achieved an accuracy of 97% on the image translation with using the LCBVAE architecture that can be generalized for any downstream learning tasks as for image translation and reconstruction.
Multilingual Language Models offer a way to incorporate multiple languages in one model and utilize cross-language transfer learning to improve performance for different Natural Language Processing (NLP) tasks. Despite progress in multilingual models, not all languages are supported as well, particularly in low-resource settings. In this work, we investigate the linguistic representation of different languages in multilingual models. We start by asking the question which languages are supported in popular multilingual models and which languages are left behind. Then, for included languages, we look at models' learned representations based on language family and dialect and try to understand how models' learned representations for~(1) seen and~(2) unseen languages vary across different language groups. In addition, we test and analyze performance on downstream tasks such as text generation and Named Entity Recognition. We observe from our experiments that community-centered models -- models that focus on languages of a given family or geographical location and are built by communities who speak them -- perform better at distinguishing between languages in the same family for low-resource languages. Our paper contributes to the literature in understanding multilingual models and their shortcomings and offers insights on potential ways to improve them.
Convolutional neural network (CNN), as an important model in artificial intelligence, has been widely used and studied in different disciplines. The computational mechanisms of CNNs are still not fully revealed due to the their complex nature. In this study, we focused on 4 extensively studied CNNs (AlexNet, VGG11, VGG13, and VGG16) which has been analyzed as human-like models by neuroscientists with ample evidence. We trained these CNNs to emotion valence classification task by transfer learning. Comparing their performance with human data, the data unveiled that these CNNs would partly perform as human does. We then update the object-based AlexNet using self-attention mechanism based on neuroscience and behavioral data. The updated FE-AlexNet outperformed all the other tested CNNs and closely resembles human perception. The results further unveil the computational mechanisms of these CNNs. Moreover, this study offers a new paradigm to better understand and improve CNN performance via human data.
Remote medical diagnosis has emerged as a critical and indispensable technique in practical medical systems, where medical data are required to be efficiently compressed and transmitted for diagnosis by either professional doctors or intelligent diagnosis devices. In this process, a large amount of redundant content irrelevant to the diagnosis is subjected to high-fidelity coding, leading to unnecessary transmission costs. To mitigate this, we propose diagnosis-oriented medical image compression, a special semantic compression task designed for medical scenarios, targeting to reduce the compression cost without compromising the diagnosis accuracy. However, collecting sufficient medical data to optimize such a compression system is significantly expensive and challenging due to privacy issues and the lack of professional annotation. In this study, we propose DMIC, the first efficient transfer learning-based codec, for diagnosis-oriented medical image compression, which can be effectively optimized with only few-shot annotated medical examples, by reusing the knowledge in the existing reinforcement learning-based task-driven semantic coding framework, i.e., HRLVSC [1]. Concretely, we focus on tuning only the partial parameters of the policy network for bit allocation within HRLVSC, which enables it to adapt to the medical images. In this work, we validate our DMIC with the typical medical task, Coronary Artery Segmentation. Extensive experiments have demonstrated that our DMIC can achieve 47.594%BD-Rate savings compared to the HEVC anchor, by tuning only the A2C module (2.7% parameters) of the policy network with only 1 medical sample.
Brain tumors are increasingly prevalent, characterized by the uncontrolled spread of aberrant tissues in the brain, with almost 700,000 new cases diagnosed globally each year. Magnetic Resonance Imaging (MRI) is commonly used for the diagnosis of brain tumors and accurate classification is a critical clinical procedure. In this study, we propose an efficient solution for classifying brain tumors from MRI images using custom transfer learning networks. While several researchers have employed various pre-trained architectures such as RESNET-50, ALEXNET, VGG-16, and VGG-19, these methods often suffer from high computational complexity. To address this issue, we present a custom and lightweight model using a Convolutional Neural Network-based pre-trained architecture with reduced complexity. Specifically, we employ the VGG-19 architecture with additional hidden layers, which reduces the complexity of the base architecture but improves computational efficiency. The objective is to achieve high classification accuracy using a novel approach. Finally, the result demonstrates a classification accuracy of 96.42%.
Few-shot learning or meta-learning leverages the data scarcity problem in machine learning. Traditionally, training data requires a multitude of samples and labeling for supervised learning. To address this issue, we propose a one-shot unsupervised meta-learning to learn the latent representation of the training samples. We use augmented samples as the query set during the training phase of the unsupervised meta-learning. A temperature-scaled cross-entropy loss is used in the inner loop of meta-learning to prevent overfitting during unsupervised learning. The learned parameters from this step are applied to the targeted supervised meta-learning in a transfer-learning fashion for initialization and fast adaptation with improved accuracy. The proposed method is model agnostic and can aid any meta-learning model to improve accuracy. We use model agnostic meta-learning (MAML) and relation network (RN) on Omniglot and mini-Imagenet datasets to demonstrate the performance of the proposed method. Furthermore, a meta-learning model with the proposed initialization can achieve satisfactory accuracy with significantly fewer training samples.
Continual pre-training has been urgent for adapting a pre-trained model to a multitude of domains and tasks in the fast-evolving world. In practice, a continually pre-trained model is expected to demonstrate not only greater capacity when fine-tuned on pre-trained domains but also a non-decreasing performance on unseen ones. In this work, we first investigate such anytime fine-tuning effectiveness of existing continual pre-training approaches, concluding with unanimously decreased performance on unseen domains. To this end, we propose a prompt-guided continual pre-training method, where we train a hypernetwork to generate domain-specific prompts by both agreement and disagreement losses. The agreement loss maximally preserves the generalization of a pre-trained model to new domains, and the disagreement one guards the exclusiveness of the generated hidden states for each domain. Remarkably, prompts by the hypernetwork alleviate the domain identity when fine-tuning and promote knowledge transfer across domains. Our method achieved improvements of 3.57% and 3.4% on two real-world datasets (including domain shift and temporal shift), respectively, demonstrating its efficacy.
Biological and artificial information processing systems form representations of the world that they can use to categorize, reason, plan, navigate, and make decisions. How can we measure the similarity between the representations formed by these diverse systems? Do similarities in representations then translate into similar behavior? If so, then how can a system's representations be modified to better match those of another system? These questions pertaining to the study of representational alignment are at the heart of some of the most promising research areas in contemporary cognitive science, neuroscience, and machine learning. In this Perspective, we survey the exciting recent developments in representational alignment research in the fields of cognitive science, neuroscience, and machine learning. Despite their overlapping interests, there is limited knowledge transfer between these fields, so work in one field ends up duplicated in another, and useful innovations are not shared effectively. To improve communication, we propose a unifying framework that can serve as a common language for research on representational alignment, and map several streams of existing work across fields within our framework. We also lay out open problems in representational alignment where progress can benefit all three of these fields. We hope that this paper will catalyze cross-disciplinary collaboration and accelerate progress for all communities studying and developing information processing systems.