FFSplit: Split Feed-Forward Network For Optimizing Accuracy-Efficiency Trade-off in Language Model Inference

The large number of parameters in Pretrained Language Models enhance their performance, but also make them resource-intensive, making it challenging to deploy them on commodity hardware like a single GPU. Due to the memory and power limitations of these devices, model compression techniques are often used to decrease both the model's size and its inference latency. This usually results in a trade-off between model accuracy and efficiency. Therefore, optimizing this balance is essential for effectively deploying LLMs on commodity hardware. A significant portion of the efficiency challenge is the Feed-forward network (FFN) component, which accounts for roughly $\frac{2}{3}$ total parameters and inference latency. In this paper, we first observe that only a few neurons of FFN module have large output norm for any input tokens, a.k.a. heavy hitters, while the others are sparsely triggered by different tokens. Based on this observation, we explicitly split the FFN into two parts according to the heavy hitters. We improve the efficiency-accuracy trade-off of existing compression methods by allocating more resource to FFN parts with heavy hitters. In practice, our method can reduce model size by 43.1\% and bring $1.25\sim1.56\times$ wall clock time speedup on different hardware with negligible accuracy drop.

Extension of TSVM to Multi-Class and Hierarchical Text Classification Problems With General Losses

Transductive SVM (TSVM) is a well known semi-supervised large margin learning method for binary text classification. In this paper we extend this method to multi-class and hierarchical classification problems. We point out that the determination of labels of unlabeled examples with fixed classifier weights is a linear programming problem. We devise an efficient technique for solving it. The method is applicable to general loss functions. We demonstrate the value of the new method using large margin loss on a number of multi-class and hierarchical classification datasets. For maxent loss we show empirically that our method is better than expectation regularization/constraint and posterior regularization methods, and competitive with the version of entropy regularization method which uses label constraints.

Predictive Approaches For Gaussian Process Classifier Model Selection

In this paper we consider the problem of Gaussian process classifier (GPC) model selection with different Leave-One-Out (LOO) Cross Validation (CV) based optimization criteria and provide a practical algorithm using LOO predictive distributions with such criteria to select hyperparameters. Apart from the standard average negative logarithm of predictive probability (NLP), we also consider smoothed versions of criteria such as F-measure and Weighted Error Rate (WER), which are useful for handling imbalanced data. Unlike the regression case, LOO predictive distributions for the classifier case are intractable. We use approximate LOO predictive distributions arrived from Expectation Propagation (EP) approximation. We conduct experiments on several real world benchmark datasets. When the NLP criterion is used for optimizing the hyperparameters, the predictive approaches show better or comparable NLP generalization performance with existing GPC approaches. On the other hand, when the F-measure criterion is used, the F-measure generalization performance improves significantly on several datasets. Overall, the EP-based predictive algorithm comes out as an excellent choice for GP classifier model selection with different optimization criteria.

Transductive Classification Methods for Mixed Graphs

In this paper we provide a principled approach to solve a transductive classification problem involving a similar graph (edges tend to connect nodes with same labels) and a dissimilar graph (edges tend to connect nodes with opposing labels). Most of the existing methods, e.g., Information Regularization (IR), Weighted vote Relational Neighbor classifier (WvRN) etc, assume that the given graph is only a similar graph. We extend the IR and WvRN methods to deal with mixed graphs. We evaluate the proposed extensions on several benchmark datasets as well as two real world datasets and demonstrate the usefulness of our ideas.

Graph Based Classification Methods Using Inaccurate External Classifier Information

In this paper we consider the problem of collectively classifying entities where relational information is available across the entities. In practice inaccurate class distribution for each entity is often available from another (external) classifier. For example this distribution could come from a classifier built using content features or a simple dictionary. Given the relational and inaccurate external classifier information, we consider two graph based settings in which the problem of collective classification can be solved. In the first setting the class distribution is used to fix labels to a subset of nodes and the labels for the remaining nodes are obtained like in a transductive setting. In the other setting the class distributions of all nodes are used to define the fitting function part of a graph regularized objective function. We define a generalized objective function that handles both the settings. Methods like harmonic Gaussian field and local-global consistency (LGC) reported in the literature can be seen as special cases. We extend the LGC and weighted vote relational neighbor classification (WvRN) methods to support usage of external classifier information. We also propose an efficient least squares regularization (LSR) based method and relate it to information regularization methods. All the methods are evaluated on several benchmark and real world datasets. Considering together speed, robustness and accuracy, experimental results indicate that the LSR and WvRN-extension methods perform better than other methods.