In this work, we reveal a strong implicit bias of stochastic gradient descent (SGD) that drives overly expressive networks to much simpler subnetworks, thereby dramatically reducing the number of independent parameters, and improving generalization. To reveal this bias, we identify invariant sets, or subsets of parameter space that remain unmodified by SGD. We focus on two classes of invariant sets that correspond to simpler (sparse or low-rank) subnetworks and commonly appear in modern architectures. Our analysis uncovers that SGD exhibits a property of stochastic attractivity towards these simpler invariant sets. We establish a sufficient condition for stochastic attractivity based on a competition between the loss landscape's curvature around the invariant set and the noise introduced by stochastic gradients. Remarkably, we find that an increased level of noise strengthens attractivity, leading to the emergence of attractive invariant sets associated with saddle-points or local maxima of the train loss. We observe empirically the existence of attractive invariant sets in trained deep neural networks, implying that SGD dynamics often collapses to simple subnetworks with either vanishing or redundant neurons. We further demonstrate how this simplifying process of stochastic collapse benefits generalization in a linear teacher-student framework. Finally, through this analysis, we mechanistically explain why early training with large learning rates for extended periods benefits subsequent generalization.
相關內容
In this study, we introduce CT-LLM, a 2B large language model (LLM) that illustrates a pivotal shift towards prioritizing the Chinese language in developing LLMs. Uniquely initiated from scratch, CT-LLM diverges from the conventional methodology by primarily incorporating Chinese textual data, utilizing an extensive corpus of 1,200 billion tokens, including 800 billion Chinese tokens, 300 billion English tokens, and 100 billion code tokens. This strategic composition facilitates the model's exceptional proficiency in understanding and processing Chinese, a capability further enhanced through alignment techniques. Demonstrating remarkable performance on the CHC-Bench, CT-LLM excels in Chinese language tasks, and showcases its adeptness in English through SFT. This research challenges the prevailing paradigm of training LLMs predominantly on English corpora and then adapting them to other languages, broadening the horizons for LLM training methodologies. By open-sourcing the full process of training a Chinese LLM, including a detailed data processing procedure with the obtained Massive Appropriate Pretraining Chinese Corpus (MAP-CC), a well-chosen multidisciplinary Chinese Hard Case Benchmark (CHC-Bench), and the 2B-size Chinese Tiny LLM (CT-LLM), we aim to foster further exploration and innovation in both academia and industry, paving the way for more inclusive and versatile language models.
In this work, we propose a method to improve the energy efficiency and fairness of simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) for mobile users, ensuring reduced power consumption while maintaining reliable communication. To achieve this, we introduce a new parameter known as the subsurface assignment variable, which determines the number of STAR-RIS elements allocated to each user. We then formulate a novel optimization problem by concurrently optimizing the phase shifts of the STAR-RIS and subsurface assignment variable. We leverage the deep reinforcement learning (DRL) technique to address this optimization problem. The DRL model predicts the phase shifts of the STAR-RIS and efficiently allocates elements of STAR-RIS to the users. Additionally, we incorporate a penalty term in the DRL model to facilitate intelligent deactivation of STAR-RIS elements when not in use to enhance energy efficiency. Through extensive experiments, we show that the proposed method can achieve fairly high and nearly equal data rates for all users in both the transmission and reflection spaces in an energy-efficient manner.
Vision Transformers (ViTs) have demonstrated remarkable performance in image classification tasks, particularly when equipped with local information via region attention or convolutions. While such architectures improve the feature aggregation from different granularities, they often fail to contribute to the robustness of the networks. Neural Cellular Automata (NCA) enables the modeling of global cell representations through local interactions, with its training strategies and architecture design conferring strong generalization ability and robustness against noisy inputs. In this paper, we propose Adaptor Neural Cellular Automata (AdaNCA) for Vision Transformer that uses NCA as plug-in-play adaptors between ViT layers, enhancing ViT's performance and robustness against adversarial samples as well as out-of-distribution inputs. To overcome the large computational overhead of standard NCAs, we propose Dynamic Interaction for more efficient interaction learning. Furthermore, we develop an algorithm for identifying the most effective insertion points for AdaNCA based on our analysis of AdaNCA placement and robustness improvement. With less than a 3% increase in parameters, AdaNCA contributes to more than 10% absolute improvement in accuracy under adversarial attacks on the ImageNet1K benchmark. Moreover, we demonstrate with extensive evaluations across 8 robustness benchmarks and 4 ViT architectures that AdaNCA, as a plug-in-play module, consistently improves the robustness of ViTs.
In this work, we address the challenge of encoding speech captured by a microphone array using deep learning techniques with the aim of preserving and accurately reconstructing crucial spatial cues embedded in multi-channel recordings. We propose a neural spatial audio coding framework that achieves a high compression ratio, leveraging single-channel neural sub-band codec and SpatialCodec. Our approach encompasses two phases: (i) a neural sub-band codec is designed to encode the reference channel with low bit rates, and (ii), a SpatialCodec captures relative spatial information for accurate multi-channel reconstruction at the decoder end. In addition, we also propose novel evaluation metrics to assess the spatial cue preservation: (i) spatial similarity, which calculates cosine similarity on a spatially intuitive beamspace, and (ii), beamformed audio quality. Our system shows superior spatial performance compared with high bitrate baselines and black-box neural architecture. Demos are available at //xzwy.github.io/SpatialCodecDemo. Codes and models are available at //github.com/XZWY/SpatialCodec.
In this work, we present a tutorial on how to account for the computational time complexity overhead of signal processing in the spectral efficiency (SE) analysis of wireless waveforms. Our methodology is particularly relevant in scenarios where achieving higher SE entails a penalty in complexity, a common trade-off present in 6G candidate waveforms. We consider that SE derives from the data rate, which is impacted by time-dependent overheads. Thus, neglecting the computational complexity overhead in the SE analysis grants an unfair advantage to more computationally complex waveforms, as they require larger computational resources to meet a signal processing runtime below the symbol period. We demonstrate our points with two case studies. In the first, we refer to IEEE 802.11a-compliant baseband processors from the literature to show that their runtime significantly impacts the SE perceived by upper layers. In the second case study, we show that waveforms considered less efficient in terms of SE can outperform their more computationally expensive counterparts if provided with equivalent high-performance computational resources. Based on these cases, we believe our tutorial can address the comparative SE analysis of waveforms that operate under different computational resource constraints.
We present GS (Guarded Successor), a novel decidable temporal logic with several unique distinctive features. Among those, it allows infinitely many data values that come not only with equality but with a somehow rich theory too: the first-order theory of atomless Boolean algebras. The language also distinguishes between inputs and outputs, and has a decision procedure for determining whether for all inputs exist outputs, at each point of time. Moreover, and maybe most surprisingly, the data values can be nothing but sentences in GS itself. We also present a non-temporal fragment called NSO (Nullary Second Order) that enjoys merely this last property. These results are crucial necessary ingredients in any meaningful design of safe AI. Finally, all those results are obtained from a novel treatment of the first-order theory of atomless Boolean algebras.
In the wake of rapid advancements in artificial intelligence (AI), we stand on the brink of a transformative leap in data systems. The imminent fusion of AI and DB (AIxDB) promises a new generation of data systems, which will relieve the burden on end-users across all industry sectors by featuring AI-enhanced functionalities, such as personalized and automated in-database AI-powered analytics, self-driving capabilities for improved system performance, etc. In this paper, we explore the evolution of data systems with a focus on deepening the fusion of AI and DB. We present NeurDB, an AI-powered autonomous data system designed to fully embrace AI design in each major system component and provide in-database AI-powered analytics. We outline the conceptual and architectural overview of NeurDB, discuss its design choices and key components, and report its current development and future plan.
We present CoDEx, a set of knowledge graph completion datasets extracted from Wikidata and Wikipedia that improve upon existing knowledge graph completion benchmarks in scope and level of difficulty. In terms of scope, CoDEx comprises three knowledge graphs varying in size and structure, multilingual descriptions of entities and relations, and tens of thousands of hard negative triples that are plausible but verified to be false. To characterize CoDEx, we contribute thorough empirical analyses and benchmarking experiments. First, we analyze each CoDEx dataset in terms of logical relation patterns. Next, we report baseline link prediction and triple classification results on CoDEx for five extensively tuned embedding models. Finally, we differentiate CoDEx from the popular FB15K-237 knowledge graph completion dataset by showing that CoDEx covers more diverse and interpretable content, and is a more difficult link prediction benchmark. Data, code, and pretrained models are available at //bit.ly/2EPbrJs.
Graph convolutional networks (GCNs) have recently become one of the most powerful tools for graph analytics tasks in numerous applications, ranging from social networks and natural language processing to bioinformatics and chemoinformatics, thanks to their ability to capture the complex relationships between concepts. At present, the vast majority of GCNs use a neighborhood aggregation framework to learn a continuous and compact vector, then performing a pooling operation to generalize graph embedding for the classification task. These approaches have two disadvantages in the graph classification task: (1)when only the largest sub-graph structure ($k$-hop neighbor) is used for neighborhood aggregation, a large amount of early-stage information is lost during the graph convolution step; (2) simple average/sum pooling or max pooling utilized, which loses the characteristics of each node and the topology between nodes. In this paper, we propose a novel framework called, dual attention graph convolutional networks (DAGCN) to address these problems. DAGCN automatically learns the importance of neighbors at different hops using a novel attention graph convolution layer, and then employs a second attention component, a self-attention pooling layer, to generalize the graph representation from the various aspects of a matrix graph embedding. The dual attention network is trained in an end-to-end manner for the graph classification task. We compare our model with state-of-the-art graph kernels and other deep learning methods. The experimental results show that our framework not only outperforms other baselines but also achieves a better rate of convergence.
Convolutional Neural Networks (CNNs) have gained significant traction in the field of machine learning, particularly due to their high accuracy in visual recognition. Recent works have pushed the performance of GPU implementations of CNNs to significantly improve their classification and training times. With these improvements, many frameworks have become available for implementing CNNs on both CPUs and GPUs, with no support for FPGA implementations. In this work we present a modified version of the popular CNN framework Caffe, with FPGA support. This allows for classification using CNN models and specialized FPGA implementations with the flexibility of reprogramming the device when necessary, seamless memory transactions between host and device, simple-to-use test benches, and the ability to create pipelined layer implementations. To validate the framework, we use the Xilinx SDAccel environment to implement an FPGA-based Winograd convolution engine and show that the FPGA layer can be used alongside other layers running on a host processor to run several popular CNNs (AlexNet, GoogleNet, VGG A, Overfeat). The results show that our framework achieves 50 GFLOPS across 3x3 convolutions in the benchmarks. This is achieved within a practical framework, which will aid in future development of FPGA-based CNNs.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191