Canada

The rapid progress of Transformers in artificial intelligence has come at the cost of increased resource consumption and greenhouse gas emissions due to growing model sizes. Prior work suggests using pretrained small models to improve training efficiency, but this approach may not be suitable for new model structures. On the other hand, training from scratch can be slow, and progressively stacking layers often fails to achieve significant acceleration. To address these challenges, we propose a novel method called Apollo, which prepares lessons for expanding operations by learning high-layer functionality during training of low layers. Our approach involves low-value-prioritized sampling (LVPS) to train different depths and weight sharing to facilitate efficient expansion. We also introduce an interpolation method for stable model depth extension. Experiments demonstrate that Apollo achieves state-of-the-art acceleration ratios, even rivaling methods using pretrained models, making it a universal and efficient solution for training deep models while reducing time, financial, and environmental costs.

AAAI 2024

Preparing Lessons for Progressive Training on Language Models

learning & optimization for snlp

text classification

language models

technical paper

We are pleased to announce the Thirty-Eighth AAAI Conference on Artificial Intelligence (AAAI-24), which will be held in Vancouver, British Columbia at the Vancouver Convention Centre – West Building from 20-27 February, 2024.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-24 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.

We expect for AAAI-24 to be an in-person conference – one author of all accepted papers will be expected to present work in person unless there are exceptional circumstances that prevent this.<br><br><br><br>

In order to access the AAAI-24 event page you need to register [here](https://aaai.org/aaai-conference/registration/)

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. 

Game theory, as an analytical tool, is frequently utilized to analyze human behavior in social science research. With the high alignment between the behavior of Large Language Models (LLMs) and humans, a promising research direction is to employ LLMs as substitutes for humans in game experiments, enabling social science research. However, despite numerous empirical researches on the combination of LLMs and game theory, the capability boundaries of LLMs in game theory remain unclear. In this research, we endeavor to systematically analyze LLMs in the context of game theory. Specifically, rationality, as the fundamental principle of game theory, serves as the metric for evaluating players' behavior --- building a clear desire, refining belief about uncertainty, and taking optimal actions. Accordingly, we select three classical games (dictator game, Rock-Paper-Scissors, and ring-network game) to analyze to what extent LLMs can achieve rationality in these three aspects. The experimental results indicate that even the current state-of-the-art LLM (GPT-4) exhibits substantial disparities compared to humans in game theory. For instance, LLMs struggle to build desires based on uncommon preferences, fail to refine belief from many simple patterns, and may overlook or modify refined belief when taking actions. Therefore, we consider that introducing LLMs into game experiments in the field of social science should be approached with greater caution.

Can Large Language Models Serve as Rational Players in Game Theory? A Systematic Analysis | VIDEO

This paper proposes SAT-based techniques to calculate a specific normal form of a given finite mathematical structure (model). The normal form is obtained by permuting the domain elements so that the representation of the structure is lexicographically smallest possible. Such a normal form is of interest to mathematicians as it enables easy cataloging of algebraic structures. In particular, two structures are isomorphic precisely when their normal forms are the same. This form is also natural to inspect as mathematicians have been using it routinely for many decades.

We develop a novel approach where a SAT solver is used in a black-box fashion to compute the smallest representative. The approach constructs the representative gradually and searches the space of possible isomorphisms, requiring a small number of variables. However, the approach may lead to a large number of SAT calls and therefore we devise propagation techniques to reduce this number. The paper focuses on finite structures with a single binary operation (encompassing groups, semigroups, etc.). However, the approach is generalizable to arbitrary finite structures. We provide an implementation of the proposed algorithm and evaluate it on a variety of algebraic structures.

SAT-Based Techniques for Lexicographically Smallest Finite Models | VIDEO

Graph neural networks (GNNs) have found widespread application in modeling graph data across diverse domains. While GNNs excel in scenarios where the testing data shares the distribution of their training counterparts (in distribution, ID), they often exhibit incorrect predictions when confronted with samples from an unfamiliar distribution (out-of-distribution, OOD). To identify and reject OOD samples with GNNs, recent studies have explored graph OOD detection, often focusing on training a specific model or modifying the data on top of a well-trained GNN. Despite their effectiveness, these methods come with heavy training resources and costs, as they need to optimize the GNN-based models on training data. Moreover, their reliance on modifying the original GNNs and accessing training data further restricts their universality. To this end, this paper introduces a method to detect \textbf{G}raph \textbf{O}ut-\textbf{o}f-\textbf{D}istribution \textbf{A}t \textbf{T}est-time (namely GOODAT), a \textit{data-centric}, \textit{unsupervised}, and \textit{plug-and-play} solution that operates independently of training data and modifications of GNN architecture. With a lightweight graph masker, GOODAT can learn informative subgraphs from test samples, enabling the capture of distinct graph patterns between OOD and ID samples. To optimize the graph masker, we meticulously design three unsupervised objective functions based on the graph information bottleneck principle, motivating the masker to capture compact yet informative subgraphs for OOD detection. Comprehensive evaluations confirm that our GOODAT method outperforms state-of-the-art benchmarks across a variety of real-world datasets.

GOODAT: Towards Test-Time Graph Out-of-Distribution Detection | VIDEO

Large Language Models (LLM) exhibit zero-shot mathematical reasoning capacity as a behavior emergent with scale, commonly manifesting as chain-of-thoughts (CoT) reasoning. However, multiple empirical findings suggest that this prowess is exclusive to LLMs that have exorbitant sizes (beyond 50 billion parameters). Meanwhile, educational neuroscientists suggest that symbolic algebraic manipulation be introduced around the same time as arithmetic word problems so as to modularize language-to-formulation, symbolic manipulation of the formulation, and endgame arithmetic.
In this paper, we start with the hypothesis that much smaller LMs, which are weak at multi-step reasoning, can achieve reasonable arithmetic reasoning if arithmetic word problems are posed as a formalize-then-solve task.
In our architecture, which we call SyReLM, the LM serves the role of a translator to map natural language arithmetic questions into a formal language (FL) description. A symbolic solver then evaluates the FL expression to obtain the answer.
A small frozen LM, equipped with an efficient low-rank adapter, is capable of generating FL expressions that incorporate natural language descriptions of the arithmetic problem (e.g., variable names and their purposes, formal expressions combining variables, etc.).
We adopt policy-gradient reinforcement learning to train the adapted LM, informed by the non-differentiable symbolic solver. This marks a sharp departure from the recent development in tool-augmented LLMs, in which the external tools (e.g., calculator, Web search, etc.) are essentially detached from the learning phase of the LM. SyReLM shows massive improvements (e.g., 31.6 absolute point improvement in accuracy on the SVAMP dataset using GPT-J 6B model)  over base LMs, while keeping our testbed easy to diagnose and interpret, and within the reach of most researchers.

Frugal LMs Trained to Invoke Symbolic Solvers Achieve Parameter-Efficient Arithmetic Reasoning

In Environment Design, one interested party seeks to affect another agent's decisions by applying changes to the environment. Most research on planning environment (re)design assumes the interested party's objective is to facilitate the recognition of goals and plans, and search over the space of environment modifications to find the minimal set of changes that simplify those tasks and optimise a particular metric. This search space is usually intractable, so existing approaches devise metric-dependent pruning techniques for performing search more efficiently. This results in approaches that are not able to generalise across different objectives and/or metrics. In this paper, we argue that the interested party could have objectives and metrics that are not necessarily related to recognising agents' goals or plans. Thus, to generalise the task of Planning Environment Redesign, we develop a general environment redesign approach that is metric-agnostic and leverages recent research on top-quality planning to efficiently redesign planning environments according to any interested party's objective and metric. Experiments over a set of environment redesign benchmarks show that our general approach outperforms existing approaches when using well-known metrics, such as facilitating the recognition of goals, as well as its effectiveness when solving environment redesign tasks that optimise a novel set of different metrics.

Generalising Planning Environment Redesign

For many combinatorial optimization problems, the translation to SAT introduces a substantial-size blow-up, constraining the feasible instance size. To address this limitation, algorithmic frameworks like SLIM have emerged, leveraging multiple local SAT calls to enhance global heuristic solutions. The performance of such algorithmic frameworks heavily relies on critical parameters, including the size of selected local instances and the time allocated per SAT call.

This paper presents a comprehensive study of configuration techniques for the TW-SLIM approach, which tackles the problem of computing small-width tree decompositions for large graphs using the SLIM paradigm. We explore various approaches for configuring TW-SLIM, including offline learning on training data and real-time configurations. These methods can notably improve over the default TW-SLIM configuration when suitably adapted to the specific multi-SAT calls setting with dynamically changing problem instances.

Building upon these insights, we propose the iterative cascading policy—a novel hybrid technique uniquely combining offline and real-time methods. The iterative cascading policy employs a pool of 30 configurations obtained through clustering-based offline methods, deploying them in dynamic cascades across multiple rounds. In each round, the 30 configurations are tested according to the cascading ordering, and the best tree decomposition is retained for further improvement, with the option to adjust the following ordering of cascades. This iterative approach significantly enhances the performance of TW-SLIM beyond baseline results, even within varying global timeouts. This highlights the effectiveness of the proposed iterative cascading policy in enhancing the efficiency and efficacy of complex algorithmic frameworks like TW-SLIM.

SAT-Based Tree Decomposition with Iterative Cascading Policy Selection

A fundamental challenge of bipartite graph representation learning is how to extract informative node embeddings. Self-Supervised Learning (SSL) is a promising paradigm to address this challenge. Most recent bipartite graph SSL methods are based on contrastive learning which learns embeddings by discriminating positive and negative node pairs. Contrastive learning usually requires a large number of negative node pairs, which could lead to computational burden and semantic errors. In this paper, we introduce a novel synergistic representation learning model (STERLING) to learn node embeddings without negative node pairs. STERLING preserves the unique local and global synergies in bipartite graphs. The local synergies are captured by maximizing the similarity of the inter-type and intra-type positive node pairs, and the global synergies are captured by maximizing the mutual information of co-clusters. Theoretical analysis demonstrates that STERLING could improve the connectivity between different node types in the embedding space. Extensive empirical evaluation on various benchmark datasets and tasks demonstrates the effectiveness of STERLING for extracting node embeddings.

Sterling: Synergistic Representation Learning on Bipartite Graphs

With the rapid advancement of AI applications, the growing needs for data privacy and model robustness have highlighted the importance of machine unlearning, especially in thriving graph-based scenarios. However, most existing graph unlearning strategies primarily rely on well-designed architectures or manual process, rendering them less user-friendly and posing challenges in terms of deployment efficiency. Furthermore, striking a balance between unlearning performance and framework generalization is also a pivotal concern. To address the above issues, we propose Mutual Evolution Graph Unlearning (MEGU), a new mutual evolution paradigm that simultaneously evolves the predictive and unlearning capacities of graph unlearning. By incorporating aforementioned two components, MEGU ensures complementary optimization in a unified training framework that aligns with the prediction and unlearning requirements. Extensive experiments on 9 graph benchmark datasets demonstrate the superior performance of MEGU in addressing unlearning requirements at the feature, node, and edge levels. Specifically, MEGU achieves average performance improvements of 2.7%, 2.5%, and 3.2% across these three levels of unlearning tasks when compared to state-of-the-art baselines. Furthermore, MEGU exhibits satisfactory training efficiency, reducing time and space overhead by an average of 159.8x and 9.6x, respectively, in comparison to retraining GNN from scratch.

Towards Effective and General Graph Unlearning via Mutual Evolution | VIDEO

One approach to probabilistic inference involves counting the number of models of a given Boolean formula. Here, we are interested in inferences involving higher-order objects, i.e., functions. We study the following task: Given a Boolean specification between a set of inputs and outputs, count the number of functions of inputs such that the specification is met. Such functions are called Skolem functions.

We are motivated by the recent development of scalable approaches to Boolean function synthesis. This stands in relation to our problem analogously to the relationship between Boolean satisfiability and the model counting problem. Yet, counting Skolem functions poses considerable new challenges. From the complexity-theoretic standpoint, counting Skolem functions is not only #P-hard; it is quite unlikely to have an FPRAS (Fully Polynomial Randomized Approximation Scheme) as the problem of synthesizing a Skolem function remains challenging, even given access to an NP oracle.

The primary contribution of this work is the first algorithm, SkolemFC, that computes the number of Skolem functions. SkolemFC relies on technical connections between counting functions and propositional model counting: our algorithm makes a linear number of calls to an approximate model counter and computes an estimate of the number of Skolem functions with theoretical guarantees. Our prototype displays impressive scalability, handling benchmarks comparably to state-of-the-art Skolem function synthesis engines, even though counting all such functions ostensibly poses a greater challenge than synthesizing a single function.

An Approximate Skolem Function Counter

The metareasoning framework aims to enable autonomous agents to factor in planning costs when making decisions. In this work, we develop the first non-myopic metareasoning algorithm for planning with Markov decision processes. Our method learns the behaviour of anytime probabilistic planning algorithms from performance data. Specifically, we propose a novel model for metareasoning, based on contextual performance profiles that predict the value of the planner's current solution given the time spent planning, the state of the planning algorithm's internal parameters, and the difficulty of the planning problem being solved. This model removes the need to assume that the current solution quality is always known, broadening the class of metareasoning problems that can be addressed. We then employ deep reinforcement learning to learn a policy that decides, at each timestep, whether to continue planning or start executing the current plan, and how to set hyperparameters of the planner to enhance its performance. We demonstrate our algorithm's ability to perform effective metareasoning in two domains.

Preparing Lessons for Progressive Training on Language Models

Downloads

Next from AAAI 2024

Can Large Language Models Serve as Rational Players in Game Theory? A Systematic Analysis | VIDEO

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES