Knowledge editing aims to update outdated or incorrect knowledge in large language models (LLMs). However, current knowledge editing methods have limited scalability for lifelong editing. This study explores the fundamental reason why knowledge editing fails in lifelong editing. We begin with the closed-form solution derived from linear associative memory, which underpins state-of-the-art knowledge editing methods. We extend the solution from single editing to lifelong editing, and through rigorous mathematical derivation, identify an interference term in the final solution, suggesting that editing knowledge may impact irrelevant knowledge. Further analysis of the interference term reveals a close relationship with superposition between knowledge representations. When knowledge superposition does not exist in language models, the interference term vanishes, allowing for lossless knowledge editing. Experiments across numerous language models reveal that knowledge superposition is universal, exhibiting high kurtosis, zero mean, and heavy-tailed distributions with clear scaling laws. Ultimately, by combining theory and experiments, we demonstrate that knowledge superposition is the fundamental reason for the failure of lifelong editing. Moreover, this is the first study to investigate knowledge editing from the perspective of superposition and provides a comprehensive observation of superposition across numerous real-world language models.

Generating graph-structured molecular data involves understanding complex graph distributions. This is crucial for de novo drug molecule design, especially when incorporating drug-drug interactions (DDIs). Existing graph generative methods often struggle to capture the graphs' permutation invariance or fail to model the collaborative dependencies among various molecular components, including atom-, bond-, and textual-level DDI information. To address these limitations, we propose DDI-Diff, a knowledge-driven dual diffusion model for DDI-based drug design. DDI-Diff employs a continuous time framework and introduces a collaborative graph diffusion process, leveraging a stochastic differential equations (SDEs) system to jointly model node and edge distribution. During pretraining, we use a large-scale, unconditional dataset, followed by conditional training on DrugBank. This enhances the model's ability to generate molecular structures that align with DDI-aware knowledge, effectively capturing the collaborative effects among drugs. Then, we validated our model using DrugBank, demonstrating that DDI-Diff improves accuracy by 4.35% more than current state-of-the-art methods across all labels and highlighting its potential in collaborative drug design.

Time series forecasting is one of the most important topics in data science. Almost every business needs to predict the future in order to make better decisions and allocate resources more effectively. Examples of time series forecasting use cases are financial forecasting, demand forecasting in logistics for operational planning of assets, demand forecasting for Azure resources, and energy demand forecasting for campus buildings and data centers. The goal of this tutorial is to demonstrate state-of-the-art forecasting approaches to problems in retail and introduce a new repository focusing on best-practices in forecasting domain, along with a library of forecasting utilities [1].

This study aims to develop an automatic classifier based on deep learning for exacerbation frequency in patients with chronic obstructive pulmonary disease (COPD). A three-layer deep belief network (DBN) with two hidden layers and one visible layer was employed to develop classification models and the models' robustness to exacerbation was analyzed. Subjects from the COPDGene cohort were labeled with exacerbation frequency, defined as the number of exacerbation events per year. A total of 10 300 subjects with 361 features each were included in the analysis. After feature selection and parameter optimization, the proposed classification method achieved an accuracy of 91.99%, using a ten-fold cross validation experiment. The analysis of DBN weights showed that there was a good visual spatial relationship between the underlying critical features of different layers. Our findings show that the most sensitive features obtained from the DBN weights are consistent with the consensus showed by clinical rules and standards for COPD diagnostics. We, thus, demonstrate that DBN is a competitive tool for exacerbation risk assessment for patients suffering from COPD.

Computerized healthcare has undergone rapid development thanks to the advances in medical imaging and machine learning technologies. Especially, recent progress on deep learning opens a new era for multimedia based clinical decision support. In this paper, we use deep learning with brain network and clinical relevant text information to make early diagnosis of Alzheimer's Disease (AD). The clinical relevant text information includes age, gender, and ApoE gene of the subject. The brain network is constructed by computing the functional connectivity of brain regions using resting-state functional magnetic resonance imaging (R-fMRI) data. A targeted autoencoder network is built to distinguish normal aging from mild cognitive impairment, an early stage of AD. The proposed method reveals discriminative brain network features effectively and provides a reliable classifier for AD detection. Compared to traditional classifiers based on R-fMRI time series data, about 31.21 percent improvement of the prediction accuracy is achieved by the proposed deep learning method, and the standard deviation reduces by 51.23 percent in the best case that means our prediction model is more stable and reliable compared to the traditional methods. Our work excavates deep learning's advantages of classifying high-dimensional multimedia data in medical services, and could help predict and prevent AD at an early stage.

Modern e-health systems have undergone rapid development thanks to the advances in communications, computing and machine learning technology. Especially, deep learning has great superiority in image analysis and disease prediction. In this paper, we use Alzheimer's Disease (AD) as an example to show advantages of deep learning in diagnosing brain diseases and providing clinical decision support. Firstly, we convert raw functional magnetic resonance imaging (fMRI) to a matrix to represent activity of 90 brain regions. Secondly, to represent the functional connectivity between different brain regions, a correlation matrix is obtained by calculating the correlation between each pair of brain regions. In the next, a targeted autoencoder network is built to classify the correlation matrix, which is sensitive to AD. Finally, the experiment results show that our proposed method for AD prediction achieves much better effects than traditional means. It finds the correlations between different brain regions efficiently, provides strong reference for AD prediction. Compared to Support Vector Machine (SVM), about 25% improvement is gained in prediction accuracy. The e-health field becomes more complete and effective owing to that. Our work helps predict AD at an early stage and take measures to slow down or even prevent the onset of it.

We present a novel regularization framework for demosaicking by viewing image as smooth signal on a weighted graph. The restoration problem is formed as a minimization of variation of the signal on graph. As an initial step, we build a weight matrix which measures the similarity between every pair of pixels, from an estimate of the full color image. After that, a two-stage optimization is carried out: first, we assume that graph Laplacian is signal dependent and solve a non-quadratic problem by gradient descent; then, we pose a variational problem on graph with a static Laplacian, under the constraint of consistency with the available samples in each color component. Performance evaluation shows that our approach can improve the previous demosaicking methods both quantitively and visually, by alleviating the artifical effect. Moreover, the mapping from image to signal on graph provides a general method for image processing.

It has been shown in a previous version of this paper that hierarchical cooperation achieves a linear throughput scaling for unicast traffic, which is due to the advantage of long-range concurrent transmissions and the technique of distributed multiple-input-multiple-output (MIMO). In this paper, we investigate the scaling law for multicast traffic with hierarchical cooperation, where each of thennodes communicates withkrandomly chosen destination nodes. Specifically, we propose a new class of scheduling policies for multicast traffic. By utilizing the hierarchical cooperative MIMO transmission, our new policies can obtain an aggregate throughput of Ω(( [(n)/(k)])1-ε) for any ε >; 0. This achieves a gain of nearly √{[(n)/(k)]} compared to the noncooperative scheme in Li 's work (Proc. ACM MobiCom, 2007, pp. 266-277). Among all four cooperative strategies proposed in our paper, one is superior in terms of the three performance metrics: throughput, delay, and energy consumption. Two factors contribute to the optimal performance: multihop MIMO transmission and converge-based scheduling. Compared to the single-hop MIMO transmission strategy, the multihop strategy achieves a throughput gain of ( [(n)/(k)])[(h-1)/(h(2h-1))] and meanwhile reduces the energy consumption byk[( α-2)/ 2]times approximately, whereh>; 1 is the number of the hierarchical layers, and α >; 2 is the path-loss exponent. Moreover, to schedule the traffic with the converge multicast instead of the pure multicast strategy, we can dramatically reduce the delay by a factor of about ( [(n)/(k)])[(h)/ 2]. Our optimal cooperative strategy achieves an approximate delay-throughput tradeoffD(n,k)/T(n,k)=Θ(k) whenh→ ∞. This tradeoff ratio is identical to that of noncooperative scheme, while the throughput is greatly improved.

In this paper, we define multicast for an ad hoc network through nodes' mobility as MotionCast and study the delay and capacity tradeoffs for it. Assuming nodes move according to an independently and identically distributed (i.i.d.) pattern and each desires to send packets to$k$distinctive destinations, we compare the delay and capacity in two transmission protocols: one uses 2-hop relay algorithm without redundancy; the other adopts the scheme of redundant packets transmissions to improve delay while at the expense of the capacity. In addition, we obtain the maximum capacity and the minimum delay under certain constraints. We find that the per-node delay and capacity for the 2-hop algorithm without redundancy are$\Theta(1/k)$and$\Theta(n\log k)$, respectively; for the 2-hop algorithm with redundancy, they are$\Omega(1/k\sqrt{n\log k})$and$\Theta(\sqrt{n\log k})$, respectively. The capacity of the 2-hop relay algorithm without redundancy is better than the multicast capacity of static networks developed by Li [IEEE/ACM Trans. Netw., vol. 17, no. 3, pp. 950–961, Jun. 2009] as long as$k$is strictly less than$n$in an order sense, while when$k=\Theta(n)$, mobility does not increase capacity anymore. The ratio between delay and capacity satisfies delay/rate$~~\geq O(nk\log k)$for these two protocols, which are both smaller than that of directly extending the fundamental tradeoff for unicast established by Neely and Modiano [IEEE Trans. Inf. Theory, vol. 51, no. 6, pp. 1917–1937, Jun. 2005] to multicast, i.e., delay/rate$~~\geq O(n k^2)$. More importantly, we have proved that the fundamental delay–capacity tradeoff ratio for multicast is delay/rate$~~\geq O(n\log k)$, which would guide us to design better routing schemes for multicast.

A new class of scheduling policies for multicast traffic are proposed in this paper. By utilizing hierarchical cooperative MIMO transmission, our new policies can obtain an aggregate throughput of $\Omega\big((\frac{n}{k})^{1-\epsilon}\big)$ for any $\epsilon>0$. This achieves a gain of nearly $\sqrt{\frac{n}{k}}$ compared with non-cooperative scheme in \cite{paper:MulticastCapacityXYLi}. Between the two cooperative strategies in our paper, the converge-based one is superior to the other on delay, while the throughput and energy consumption performances are nearly the same. Moreover, to schedule the traffic in a converge multicast manner instead of the simple multicast, we can dramatically reduce the delay by a factor nearly $(\frac{n}{k})^\frac{h}{2}$, where $h>1$ is the number of the hierarchical layers. Our optimal cooperative strategy achieves an approximate delay-throughput tradeoff $D(n,k)/T(n,k)=\Theta(k)$ when $h\rightarrow\infty$. This tradeoff ratio is identical to that of non-cooperative scheme, while the throughput performance is greatly improved. Besides, for certain $k$ and $h$, the tradeoff ratio is even better than that of unicast.

We study the effect of various geometries on the capacity of wireless networks via percolation, which was not considered much before. Percolation theory was first applied to derive an achievable rate 1/√n in [5] by constructing a highway system, in contrast with the previous result Θ(1/√n log n) in [6], where n is the number of the nodes. While a highway system that consists of both horizontal and vertical edge-disjoint paths exists in a square network, B. Liu et al. in [1] pointed out that the horizontal paths will disappear if the width of a strip network is increasing more slowly than \log n. In this paper, first we take a deeper look at the percolation in a strip network. We discover that when a highway system exists, the capacity is restricted by the maximum length of the sides. Moreover, a sub-highway system is still in presence when the highway system disappears. Secondly, we consider the situations in a triangle network. Conditions that percolation highway exists in it, and the achievable rate for a triangle network are discussed. We find that corner effect can be a bottleneck of the capacity. By combining the achievable rate of the former networks, we attribute the variance between them to their symmetry discrepancies. Finally, we turn to the capacity of three dimensional (3D) networks via percolation. The whole study shows that geometric symmetry plays a significant role in the percolation and the capacity, thereby shedding a light on the network design and the scheduling.

In this paper, we define multicast for ad hoc network through nodes' mobility as MotionCast, and study the capacity and delay tradeoffs for it. Assuming nodes move according to an independently and identically distributed (i.i.d.) pattern and each desires to send packets to k distinctive destinations, we compare the capacity and delay in two transmission protocols: one uses 2-hop relay algorithm without redundancy, the other adopts the scheme of redundant packets transmissions to improve delay while at the expense of the capacity. In addition, we obtain the maximum capacity and the minimum delay under certain constraints. We find that the per-node capacity and delay for 2-hop algorithm without redundancy are Θ(1/k) and Θ(nlog k), respectively; and for 2-hop algorithm with redundancy they are Ω(1/(k√nlog k)) and Θ(√nlog k), respectively. The capacity of the 2-hop relay algorithm without redundancy is better than the multicast capacity of static networks developed in [3] as long as k is strictly less than n in an order sense; while when k=Θ(n), mobility does not increase capacity anymore. The ratio between delay and capacity satisfies delay/rate ≥ O(nklog k) for these two protocols, which is smaller than that of directly extending the fundamental tradeoff for unicast established in [1] to multicast, i.e., O(nk2).

Extensive studies on supporting voice traffic over wireless 802.11 networks have been carried out in the literature. Most of them were focused only on one hop infrastructure mode. This paper addresses the issue of voice over multihop wireless networks. Through simulation considering a simple topology, we identify the voice traffic bottleneck in 802.11 MAC and thus propose a burst queue (BQ) scheme to cooperate with 802.11b and 802.11e MAC protocols. Extensive simulation results under variety of circumstances show that BQ scheme can support 50%~100% more voice calls and obtain 20%~50% lower delay at the cost of a little higher loss ratio which is tolerable to voice flows. Even for the grid and random topology, our BQ scheme can provide smaller packet loss ratio, and shorter end-to-end delay.