Model merging and task arithmetic have emerged as promising scalable approaches to merge multiple single-task checkpoints to one multi-task model, but their applicability is reduced by significant performance loss. Previous works have linked these drops to interference in the weight space and erasure of important task-specific features. Instead, in this work we show that the information required to solve each task is still preserved after merging as different tasks mostly use non-overlapping sets of weights. We propose TALL-masks, a method to identify these task supports given a collection of task vectors and show that one can retrieve >99% of the single task accuracy by applying our masks to the multi-task vector, effectively compressing the individual checkpoints. We study the statistics of intersections among constructed masks and reveal the existence of selfish and catastrophic weights, i.e., parameters that are important exclusively to one task and irrelevant to all tasks but detrimental to multi-task fusion. For this reason, we propose Consensus Merging, an algorithm that eliminates such weights and improves the general performance of existing model merging approaches. Our experiments in vision and NLP benchmarks with up to 20 tasks, show that Consensus Merging consistently improves existing approaches. Furthermore, our proposed compression scheme reduces storage from 57Gb to 8.2Gb while retaining 99.7% of original performance.
相關內容
In coding theory, codes are usually designed with a certain level of randomness to facilitate analysis and accommodate different channel conditions. However, the resulting random code constructed can be suboptimal in practical implementations. Represented by a bipartite graph, the Batched Sparse Code (BATS Code) is a randomly constructed erasure code that utilizes network coding to achieve near-optimal performance in wireless multi-hop networks. In the performance analysis in the previous research, it is implicitly assumed that the coded batches in the BATS code are independent. This assumption holds only asymptotically when the number of input symbols is infinite, but it does not generally hold in a practical setting where the number of input symbols is finite, especially when the code is constructed randomly. We show that dependence among the batches significantly degrades the code's performance. In order to control the batch dependence through graphical design, we propose constructing the BATS code in a structured manner. A hardware-friendly structured BATS code called the Cyclic-Shift BATS (CS-BATS) code is proposed, which constructs the code from a small base graph using light-weight cyclic-shift operations. We demonstrate that when the base graph is properly designed, a higher decoding rate and a smaller complexity can be achieved compared with the random BATS code.
Recent work on discrete speech tokenization has paved the way for models that can seamlessly perform multiple tasks across modalities, e.g., speech recognition, text to speech, speech to speech translation. Moreover, large language models (LLMs) pretrained from vast text corpora contain rich linguistic information that can improve accuracy in a variety of tasks. In this paper, we present a decoder-only Discrete Multimodal Language Model (DMLM), which can be flexibly applied to multiple tasks (ASR, T2S, S2TT, etc.) and modalities (text, speech, vision). We explore several critical aspects of discrete multi-modal models, including the loss function, weight initialization, mixed training supervision, and codebook. Our results show that DMLM benefits significantly, across multiple tasks and datasets, from a combination of supervised and unsupervised training. Moreover, for ASR, it benefits from initializing DMLM from a pretrained LLM, and from a codebook derived from Whisper activations.
Applications from manipulation to autonomous vehicles rely on robust and general object tracking to safely perform tasks in dynamic environments. We propose the first certifiably optimal category-level approach for simultaneous shape estimation and pose tracking of an object of known category (e.g. a car). Our approach uses 3D semantic keypoint measurements extracted from an RGB-D image sequence, and phrases the estimation as a fixed-lag smoothing problem. Temporal constraints enforce the object's rigidity (fixed shape) and smooth motion according to a constant-twist motion model. The solutions to this problem are the estimates of the object's state (poses, velocities) and shape (paramaterized according to the active shape model) over the smoothing horizon. Our key contribution is to show that despite the non-convexity of the fixed-lag smoothing problem, we can solve it to certifiable optimality using a small-size semidefinite relaxation. We also present a fast outlier rejection scheme that filters out incorrect keypoint detections with shape and time compatibility tests, and wrap our certifiable solver in a graduated non-convexity scheme. We evaluate the proposed approach on synthetic and real data, showcasing its performance in a table-top manipulation scenario and a drone-based vehicle tracking application.
Training and inference with large machine learning models that far exceed the memory capacity of individual devices necessitates the design of distributed architectures, forcing one to contend with communication constraints. We present a framework for distributed computation over a quantum network in which data is encoded into specialized quantum states. We prove that for models within this framework, inference and training using gradient descent can be performed with exponentially less communication compared to their classical analogs, and with relatively modest overhead relative to standard gradient-based methods. We show that certain graph neural networks are particularly amenable to implementation within this framework, and moreover present empirical evidence that they perform well on standard benchmarks. To our knowledge, this is the first example of exponential quantum advantage for a generic class of machine learning problems that hold regardless of the data encoding cost. Moreover, we show that models in this class can encode highly nonlinear features of their inputs, and their expressivity increases exponentially with model depth. We also delineate the space of models for which exponential communication advantages hold by showing that they cannot hold for linear classification. Our results can be combined with natural privacy advantages in the communicated quantum states that limit the amount of information that can be extracted from them about the data and model parameters. Taken as a whole, these findings form a promising foundation for distributed machine learning over quantum networks.
Inspired by recent progress in dynamic programming approaches for weighted model counting, we investigate a dynamic-programming approach in the context of boolean realizability and synthesis, which takes a conjunctive-normal-form boolean formula over input and output variables, and aims at synthesizing witness functions for the output variables in terms of the inputs. We show how graded project-join trees, obtained via tree decomposition, can be used to compute a BDD representing the realizability set for the input formulas in a bottom-up order. We then show how the intermediate BDDs generated during realizability checking phase can be applied to synthesizing the witness functions in a top-down manner. An experimental evaluation of a solver -- DPSynth -- based on these ideas demonstrates that our approach for Boolean realizabilty and synthesis has superior time and space performance over a heuristics-based approach using same symbolic representations. We discuss the advantage on scalability of the new approach, and also investigate our findings on the performance of the DP framework.
This work formulates model selection as an infinite-armed bandit problem, namely, a problem in which a decision maker iteratively selects one of an infinite number of fixed choices (i.e., arms) when the properties of each choice are only partially known at the time of allocation and may become better understood over time, via the attainment of rewards.Here, the arms are machine learning models to train and selecting an arm corresponds to a partial training of the model (resource allocation).The reward is the accuracy of the selected model after its partial training.We aim to identify the best model at the end of a finite number of resource allocations and thus consider the best arm identification setup. We propose the algorithm Mutant-UCB that incorporates operators from evolutionary algorithms into the UCB-E (Upper Confidence Bound Exploration) bandit algorithm introduced by Audiber et al.Tests carried out on three open source image classification data sets attest to the relevance of this novel combining approach, which outperforms the state-of-the-art for a fixed budget.
We consider dimension reduction of multiview data, which are emerging in scientific studies. Formulating multiview data as multi-variate data with block structures corresponding to the different views, or views of data, we estimate top eigenvectors from multiview data that have two-fold sparsity, elementwise sparsity and blockwise sparsity. We propose a Fantope-based optimization criterion with multiple penalties to enforce the desired sparsity patterns and a denoising step is employed to handle potential presence of heteroskedastic noise across different data views. An alternating direction method of multipliers (ADMM) algorithm is used for optimization. We derive the l2 convergence of the estimated top eigenvectors and establish their sparsity and support recovery properties. Numerical studies are used to illustrate the proposed method.
Graph Neural Networks (GNNs) have received considerable attention on graph-structured data learning for a wide variety of tasks. The well-designed propagation mechanism which has been demonstrated effective is the most fundamental part of GNNs. Although most of GNNs basically follow a message passing manner, litter effort has been made to discover and analyze their essential relations. In this paper, we establish a surprising connection between different propagation mechanisms with a unified optimization problem, showing that despite the proliferation of various GNNs, in fact, their proposed propagation mechanisms are the optimal solution optimizing a feature fitting function over a wide class of graph kernels with a graph regularization term. Our proposed unified optimization framework, summarizing the commonalities between several of the most representative GNNs, not only provides a macroscopic view on surveying the relations between different GNNs, but also further opens up new opportunities for flexibly designing new GNNs. With the proposed framework, we discover that existing works usually utilize naive graph convolutional kernels for feature fitting function, and we further develop two novel objective functions considering adjustable graph kernels showing low-pass or high-pass filtering capabilities respectively. Moreover, we provide the convergence proofs and expressive power comparisons for the proposed models. Extensive experiments on benchmark datasets clearly show that the proposed GNNs not only outperform the state-of-the-art methods but also have good ability to alleviate over-smoothing, and further verify the feasibility for designing GNNs with our unified optimization framework.
Embedding entities and relations into a continuous multi-dimensional vector space have become the dominant method for knowledge graph embedding in representation learning. However, most existing models ignore to represent hierarchical knowledge, such as the similarities and dissimilarities of entities in one domain. We proposed to learn a Domain Representations over existing knowledge graph embedding models, such that entities that have similar attributes are organized into the same domain. Such hierarchical knowledge of domains can give further evidence in link prediction. Experimental results show that domain embeddings give a significant improvement over the most recent state-of-art baseline knowledge graph embedding models.
Dynamic programming (DP) solves a variety of structured combinatorial problems by iteratively breaking them down into smaller subproblems. In spite of their versatility, DP algorithms are usually non-differentiable, which hampers their use as a layer in neural networks trained by backpropagation. To address this issue, we propose to smooth the max operator in the dynamic programming recursion, using a strongly convex regularizer. This allows to relax both the optimal value and solution of the original combinatorial problem, and turns a broad class of DP algorithms into differentiable operators. Theoretically, we provide a new probabilistic perspective on backpropagating through these DP operators, and relate them to inference in graphical models. We derive two particular instantiations of our framework, a smoothed Viterbi algorithm for sequence prediction and a smoothed DTW algorithm for time-series alignment. We showcase these instantiations on two structured prediction tasks and on structured and sparse attention for neural machine translation.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191