Tim Klinger

Junkyu Lee, Michael Katz, Don Joven Agravante et al. · ibm-research

Two common approaches to sequential decision-making are AI planning (AIP) and reinforcement learning (RL). Each has strengths and weaknesses. AIP is interpretable, easy to integrate with symbolic knowledge, and often efficient, but requires an up-front logical domain specification and is sensitive to noise; RL only requires specification of rewards and is robust to noise but is sample inefficient and not easily supplied with external knowledge. We propose an integrative approach that combines high-level planning with RL, retaining interpretability, transfer, and efficiency, while allowing for robust learning of the lower-level plan actions. Our approach defines options in hierarchical reinforcement learning (HRL) from AIP operators by establishing a correspondence between the state transition model of AI planning problem and the abstract state transition system of a Markov Decision Process (MDP). Options are learned by adding intrinsic rewards to encourage consistency between the MDP and AIP transition models. We demonstrate the benefit of our integrated approach by comparing the performance of RL and HRL algorithms in both MiniGrid and N-rooms environments, showing the advantage of our method over the existing ones.

Geraud Nangue Tasse, Matthew Riemer, Benjamin Rosman et al.

Recent success in developing increasingly general purpose agents based on sequence models has led to increased focus on the problem of deploying computationally limited agents within the vastly more complex real-world. A key challenge experienced in these more realistic domains is highly non-Markovian dependencies with respect to the agent's observations, which are less common in small controlled domains. The predominant approach for dealing with this in the literature is to stack together a window of the most recent observations (Frame Stacking), but this window size must grow with the degree of non-Markovian dependencies, which results in prohibitive computational and memory requirements for both action inference and learning. In this paper, we are motivated by the insight that in many environments that are highly non-Markovian with respect to time, the environment only causally depends on a relatively small number of observations over that time-scale. A natural direction would then be to consider meta-algorithms that maintain relatively small adaptive stacks of memories such that it is possible to express highly non-Markovian dependencies with respect to time while considering fewer observations at each step and thus experience substantial savings in both compute and memory requirements. Hence, we propose a meta-algorithm (Adaptive Stacking) for achieving exactly that with convergence guarantees and quantify the reduced computation and memory constraints for MLP, LSTM, and Transformer-based agents. Our experiments utilize popular memory tasks, which give us control over the degree of non-Markovian dependencies. This allows us to demonstrate that an appropriate meta-algorithm can learn the removal of memories not predictive of future rewards without excessive removal of important experiences. Code: https://github.com/geraudnt/adaptive-stacking

Tailia Malloy, Miao Liu, Matthew D. Riemer et al.

Humans learn quickly even in tasks that contain complex visual information. This is due in part to the efficient formation of compressed representations of visual information, allowing for better generalization and robustness. However, compressed representations alone are insufficient for explaining the high speed of human learning. Reinforcement learning (RL) models that seek to replicate this impressive efficiency may do so through the use of factored representations of tasks. These informationally simplistic representations of tasks are similarly motivated as the use of compressed representations of visual information. Recent studies have connected biological visual perception to disentangled and compressed representations. This raises the question of how humans learn to efficiently represent visual information in a manner useful for learning tasks. In this paper we present a model of human factored representation learning based on an altered form of a $β$-Variational Auto-encoder used in a visual learning task. Modelling results demonstrate a trade-off in the informational complexity of model latent dimension spaces, between the speed of learning and the accuracy of reconstructions.

Tim Klinger, Luke Liu, Soham Dan et al.

Compositional generalization is a key ability of humans that enables us to learn new concepts from only a handful examples. Neural machine learning models, including the now ubiquitous Transformers, struggle to generalize in this way, and typically require thousands of examples of a concept during training in order to generalize meaningfully. This difference in ability between humans and artificial neural architectures, motivates this study on a neuro-symbolic architecture called the Compositional Program Generator (CPG). CPG has three key features: \textit{modularity}, \textit{composition}, and \textit{abstraction}, in the form of grammar rules, that enable it to generalize both systematically to new concepts in a few-shot manner, as well as productively by length on various sequence-to-sequence language tasks. For each input, CPG uses a grammar of the input language and a parser to generate a parse in which each grammar rule is assigned its own unique semantic module, a probabilistic copy or substitution program. Instances with the same parse are always processed with the same composed modules, while those with different parses may be processed with different modules. CPG learns parameters for the modules and is able to learn the semantics for new rules and types incrementally, without forgetting or retraining on rules it's already seen. It achieves perfect generalization on both the SCAN and COGS benchmarks using just 14 examples for SCAN and 22 examples for COGS -- state-of-the-art accuracy with a 1000x improvement in sample efficiency.

Parikshit Ram, Tim Klinger, Alexander G. Gray

Compositionality is thought to be a key component of language, and various compositional benchmarks have been developed to empirically probe the compositional generalization of existing sequence processing models. These benchmarks often highlight failures of existing models, but it is not clear why these models fail in this way. In this paper, we seek to theoretically understand the role the compositional structure of the models plays in these failures and how this structure relates to their expressivity and sample complexity. We propose a general neuro-symbolic definition of compositional functions and their compositional complexity. We then show how various existing general and special purpose sequence processing models (such as recurrent, convolution and attention-based ones) fit this definition and use it to analyze their compositional complexity. Finally, we provide theoretical guarantees for the expressivity and systematic generalization of compositional models that explicitly depend on our proposed definition and highlighting factors which drive poor empirical performance.

Kinjal Basu, Keerthiram Murugesan, Subhajit Chaudhury et al. · ibm-research

Text-based games (TBGs) have emerged as an important collection of NLP tasks, requiring reinforcement learning (RL) agents to combine natural language understanding with reasoning. A key challenge for agents attempting to solve such tasks is to generalize across multiple games and demonstrate good performance on both seen and unseen objects. Purely deep-RL-based approaches may perform well on seen objects; however, they fail to showcase the same performance on unseen objects. Commonsense-infused deep-RL agents may work better on unseen data; unfortunately, their policies are often not interpretable or easily transferable. To tackle these issues, in this paper, we present EXPLORER which is an exploration-guided reasoning agent for textual reinforcement learning. EXPLORER is neurosymbolic in nature, as it relies on a neural module for exploration and a symbolic module for exploitation. It can also learn generalized symbolic policies and perform well over unseen data. Our experiments show that EXPLORER outperforms the baseline agents on Text-World cooking (TW-Cooking) and Text-World Commonsense (TWC) games.

Takuya Ito, Murray Campbell, Lior Horesh et al.

The rapid development of artificial intelligence (AI) systems has created an urgent need for their scientific quantification. While their fluency across a variety of domains is impressive, AI systems fall short on tests requiring algorithmic reasoning -- a glaring limitation given the necessity for interpretable and reliable technology. Despite a surge of reasoning benchmarks emerging from the academic community, no theoretical framework exists to quantify algorithmic reasoning in AI systems. Here, we adopt a framework from computational complexity theory to quantify algorithmic generalization using algebraic expressions: algebraic circuit complexity. Algebraic circuit complexity theory -- the study of algebraic expressions as circuit models -- is a natural framework to study the complexity of algorithmic computation. Algebraic circuit complexity enables the study of generalization by defining benchmarks in terms of the computational requirements to solve a problem. Moreover, algebraic circuits are generic mathematical objects; an arbitrarily large number of samples can be generated for a specified circuit, making it an ideal experimental sandbox for the data-hungry models that are used today. In this Perspective, we adopt tools from algebraic circuit complexity, apply them to formalize a science of algorithmic generalization, and address key challenges for its successful application to AI science.

Parikshit Ram, Kenneth L. Clarkson, Tim Klinger et al.

Various forms of sparse attention have been explored to mitigate the quadratic computational and memory cost of the attention mechanism in transformers. We study sparse transformers not through a lens of efficiency but rather in terms of learnability and generalization. Empirically studying a range of attention mechanisms, we find that input-dependent sparse attention models appear to converge faster and generalize better than standard attention models, while input-agnostic sparse attention models show no such benefits -- a phenomenon that is robust across architectural and optimization hyperparameter choices. This can be interpreted as demonstrating that concentrating a model's "semantic focus" with respect to the tokens currently being considered (in the form of input-dependent sparse attention) accelerates learning. We develop a theoretical characterization of the conditions that explain this behavior. We establish a connection between the stability of the standard softmax and the loss function's Lipschitz properties, then show how sparsity affects the stability of the softmax and the subsequent convergence and generalization guarantees resulting from the attention mechanism. This allows us to theoretically establish that input-agnostic sparse attention does not provide any benefits. We also characterize conditions when semantic focus (input-dependent sparse attention) can provide improved guarantees, and we validate that these conditions are in fact met in our empirical evaluations.

Takuya Ito, Luca Cocchi, Tim Klinger et al.

In transformers, the positional encoding (PE) provides essential information that distinguishes the position and order amongst tokens in a sequence. Most prior investigations of PE effects on generalization were tailored to 1D input sequences, such as those presented in natural language, where adjacent tokens (e.g., words) are highly related. In contrast, many real world tasks involve datasets with highly non-trivial positional arrangements, such as datasets organized in multiple spatial dimensions, or datasets for which ground truth positions are not known. Here we find that the choice of initialization of a learnable PE greatly influences its ability to learn interpretable PEs that lead to enhanced generalization. We empirically demonstrate our findings in three experiments: 1) A 2D relational reasoning task; 2) A nonlinear stochastic network simulation; 3) A real world 3D neuroscience dataset, applying interpretability analyses to verify the learning of accurate PEs. Overall, we find that a learned PE initialized from a small-norm distribution can 1) uncover interpretable PEs that mirror ground truth positions in multiple dimensions, and 2) lead to improved generalization. These results illustrate the feasibility of learning identifiable and interpretable PEs for enhanced generalization.

Maxwell Crouse, Pavan Kapanipathi, Subhajit Chaudhury et al.

Nearly all general-purpose neural semantic parsers generate logical forms in a strictly top-down autoregressive fashion. Though such systems have achieved impressive results across a variety of datasets and domains, recent works have called into question whether they are ultimately limited in their ability to compositionally generalize. In this work, we approach semantic parsing from, quite literally, the opposite direction; that is, we introduce a neural semantic parsing generation method that constructs logical forms from the bottom up, beginning from the logical form's leaves. The system we introduce is lazy in that it incrementally builds up a set of potential semantic parses, but only expands and processes the most promising candidate parses at each generation step. Such a parsimonious expansion scheme allows the system to maintain an arbitrarily large set of parse hypotheses that are never realized and thus incur minimal computational overhead. We evaluate our approach on compositional generalization; specifically, on the challenging CFQ dataset and three Text-to-SQL datasets where we show that our novel, bottom-up semantic parsing technique outperforms general-purpose semantic parsers while also being competitive with comparable neural parsers that have been designed for each task.

Tailia Malloy, Tim Klinger, Miao Liu et al.

This paper introduces an information-theoretic constraint on learned policy complexity in the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) reinforcement learning algorithm. Previous research with a related approach in continuous control experiments suggests that this method favors learning policies that are more robust to changing environment dynamics. The multi-agent game setting naturally requires this type of robustness, as other agents' policies change throughout learning, introducing a nonstationary environment. For this reason, recent methods in continual learning are compared to our approach, termed Capacity-Limited MADDPG. Results from experimentation in multi-agent cooperative and competitive tasks demonstrate that the capacity-limited approach is a good candidate for improving learning performance in these environments.

Tailia Malloy, Chris R. Sims, Tim Klinger et al.

Biological agents learn and act intelligently in spite of a highly limited capacity to process and store information. Many real-world problems involve continuous control, which represents a difficult task for artificial intelligence agents. In this paper we explore the potential learning advantages a natural constraint on information flow might confer onto artificial agents in continuous control tasks. We focus on the model-free reinforcement learning (RL) setting and formalize our approach in terms of an information-theoretic constraint on the complexity of learned policies. We show that our approach emerges in a principled fashion from the application of rate-distortion theory. We implement a novel Capacity-Limited Actor-Critic (CLAC) algorithm and situate it within a broader family of RL algorithms such as the Soft Actor Critic (SAC) and Mutual Information Reinforcement Learning (MIRL) algorithm. Our experiments using continuous control tasks show that compared to alternative approaches, CLAC offers improvements in generalization between training and modified test environments. This is achieved in the CLAC model while displaying the high sample efficiency of similar methods.

Tim Klinger, Dhaval Adjodah, Vincent Marois et al.

Compositional and relational learning is a hallmark of human intelligence, but one which presents challenges for neural models. One difficulty in the development of such models is the lack of benchmarks with clear compositional and relational task structure on which to systematically evaluate them. In this paper, we introduce an environment called ConceptWorld, which enables the generation of images from compositional and relational concepts, defined using a logical domain specific language. We use it to generate images for a variety of compositional structures: 2x2 squares, pentominoes, sequences, scenes involving these objects, and other more complex concepts. We perform experiments to test the ability of standard neural architectures to generalize on relations with compositional arguments as the compositional depth of those arguments increases and under substitution. We compare standard neural networks such as MLP, CNN and ResNet, as well as state-of-the-art relational networks including WReN and PrediNet in a multi-class image classification setting. For simple problems, all models generalize well to close concepts but struggle with longer compositional chains. For more complex tests involving substitutivity, all models struggle, even with short chains. In highlighting these difficulties and providing an environment for further experimentation, we hope to encourage the development of models which are able to generalize effectively in compositional, relational domains.

Cameron Allen, Michael Katz, Tim Klinger et al.

The difficulty of deterministic planning increases exponentially with search-tree depth. Black-box planning presents an even greater challenge, since planners must operate without an explicit model of the domain. Heuristics can make search more efficient, but goal-aware heuristics for black-box planning usually rely on goal counting, which is often quite uninformative. In this work, we show how to overcome this limitation by discovering macro-actions that make the goal-count heuristic more accurate. Our approach searches for macro-actions with focused effects (i.e. macros that modify only a small number of state variables), which align well with the assumptions made by the goal-count heuristic. Focused macros dramatically improve black-box planning efficiency across a wide range of planning domains, sometimes beating even state-of-the-art planners with access to a full domain model.

Clemens Rosenbaum, Ignacio Cases, Matthew Riemer et al.

Compositionality is a key strategy for addressing combinatorial complexity and the curse of dimensionality. Recent work has shown that compositional solutions can be learned and offer substantial gains across a variety of domains, including multi-task learning, language modeling, visual question answering, machine comprehension, and others. However, such models present unique challenges during training when both the module parameters and their composition must be learned jointly. In this paper, we identify several of these issues and analyze their underlying causes. Our discussion focuses on routing networks, a general approach to this problem, and examines empirically the interplay of these challenges and a variety of design decisions. In particular, we consider the effect of how the algorithm decides on module composition, how the algorithm updates the modules, and if the algorithm uses regularization.

Andres Campero, Aldo Pareja, Tim Klinger et al.

A hallmark of human cognition is the ability to continually acquire and distill observations of the world into meaningful, predictive theories. In this paper we present a new mechanism for logical theory acquisition which takes a set of observed facts and learns to extract from them a set of logical rules and a small set of core facts which together entail the observations. Our approach is neuro-symbolic in the sense that the rule pred- icates and core facts are given dense vector representations. The rules are applied to the core facts using a soft unification procedure to infer additional facts. After k steps of forward inference, the consequences are compared to the initial observations and the rules and core facts are then encouraged towards representations that more faithfully generate the observations through inference. Our approach is based on a novel neural forward-chaining differentiable rule induction network. The rules are interpretable and learned compositionally from their predicates, which may be invented. We demonstrate the efficacy of our approach on a variety of ILP rule induction and domain theory learning datasets.

Matthew Riemer, Tim Klinger, Djallel Bouneffouf et al.

Given the recent success of Deep Learning applied to a variety of single tasks, it is natural to consider more human-realistic settings. Perhaps the most difficult of these settings is that of continual lifelong learning, where the model must learn online over a continuous stream of non-stationary data. A successful continual lifelong learning system must have three key capabilities: it must learn and adapt over time, it must not forget what it has learned, and it must be efficient in both training time and memory. Recent techniques have focused their efforts primarily on the first two capabilities while questions of efficiency remain largely unexplored. In this paper, we consider the problem of efficient and effective storage of experiences over very large time-frames. In particular we consider the case where typical experiences are O(n) bits and memories are limited to O(k) bits for k << n. We present a novel scalable architecture and training algorithm in this challenging domain and provide an extensive evaluation of its performance. Our results show that we can achieve considerable gains on top of state-of-the-art methods such as GEM.

Shuohang Wang, Mo Yu, Jing Jiang et al.

A popular recent approach to answering open-domain questions is to first search for question-related passages and then apply reading comprehension models to extract answers. Existing methods usually extract answers from single passages independently. But some questions require a combination of evidence from across different sources to answer correctly. In this paper, we propose two models which make use of multiple passages to generate their answers. Both use an answer-reranking approach which reorders the answer candidates generated by an existing state-of-the-art QA model. We propose two methods, namely, strength-based re-ranking and coverage-based re-ranking, to make use of the aggregated evidence from different passages to better determine the answer. Our models have achieved state-of-the-art results on three public open-domain QA datasets: Quasar-T, SearchQA and the open-domain version of TriviaQA, with about 8 percentage points of improvement over the former two datasets.

Clemens Rosenbaum, Tim Klinger, Matthew Riemer

Multi-task learning (MTL) with neural networks leverages commonalities in tasks to improve performance, but often suffers from task interference which reduces the benefits of transfer. To address this issue we introduce the routing network paradigm, a novel neural network and training algorithm. A routing network is a kind of self-organizing neural network consisting of two components: a router and a set of one or more function blocks. A function block may be any neural network - for example a fully-connected or a convolutional layer. Given an input the router makes a routing decision, choosing a function block to apply and passing the output back to the router recursively, terminating when a fixed recursion depth is reached. In this way the routing network dynamically composes different function blocks for each input. We employ a collaborative multi-agent reinforcement learning (MARL) approach to jointly train the router and function blocks. We evaluate our model against cross-stitch networks and shared-layer baselines on multi-task settings of the MNIST, mini-imagenet, and CIFAR-100 datasets. Our experiments demonstrate a significant improvement in accuracy, with sharper convergence. In addition, routing networks have nearly constant per-task training cost while cross-stitch networks scale linearly with the number of tasks. On CIFAR-100 (20 tasks) we obtain cross-stitch performance levels with an 85% reduction in training time.

Shuohang Wang, Mo Yu, Xiaoxiao Guo et al.

In recent years researchers have achieved considerable success applying neural network methods to question answering (QA). These approaches have achieved state of the art results in simplified closed-domain settings such as the SQuAD (Rajpurkar et al., 2016) dataset, which provides a pre-selected passage, from which the answer to a given question may be extracted. More recently, researchers have begun to tackle open-domain QA, in which the model is given a question and access to a large corpus (e.g., wikipedia) instead of a pre-selected passage (Chen et al., 2017a). This setting is more complex as it requires large-scale search for relevant passages by an information retrieval component, combined with a reading comprehension model that "reads" the passages to generate an answer to the question. Performance in this setting lags considerably behind closed-domain performance. In this paper, we present a novel open-domain QA system called Reinforced Ranker-Reader $(R^3)$, based on two algorithmic innovations. First, we propose a new pipeline for open-domain QA with a Ranker component, which learns to rank retrieved passages in terms of likelihood of generating the ground-truth answer to a given question. Second, we propose a novel method that jointly trains the Ranker along with an answer-generation Reader model, based on reinforcement learning. We report extensive experimental results showing that our method significantly improves on the state of the art for multiple open-domain QA datasets.

Clemens Rosenbaum, Tian Gao, Tim Klinger

In this paper we present a new dataset and user simulator e-QRAQ (explainable Query, Reason, and Answer Question) which tests an Agent's ability to read an ambiguous text; ask questions until it can answer a challenge question; and explain the reasoning behind its questions and answer. The User simulator provides the Agent with a short, ambiguous story and a challenge question about the story. The story is ambiguous because some of the entities have been replaced by variables. At each turn the Agent may ask for the value of a variable or try to answer the challenge question. In response the User simulator provides a natural language explanation of why the Agent's query or answer was useful in narrowing down the set of possible answers, or not. To demonstrate one potential application of the e-QRAQ dataset, we train a new neural architecture based on End-to-End Memory Networks to successfully generate both predictions and partial explanations of its current understanding of the problem. We observe a strong correlation between the quality of the prediction and explanation.

Iulian Vlad Serban, Tim Klinger, Gerald Tesauro et al.

We introduce the multiresolution recurrent neural network, which extends the sequence-to-sequence framework to model natural language generation as two parallel discrete stochastic processes: a sequence of high-level coarse tokens, and a sequence of natural language tokens. There are many ways to estimate or learn the high-level coarse tokens, but we argue that a simple extraction procedure is sufficient to capture a wealth of high-level discourse semantics. Such procedure allows training the multiresolution recurrent neural network by maximizing the exact joint log-likelihood over both sequences. In contrast to the standard log- likelihood objective w.r.t. natural language tokens (word perplexity), optimizing the joint log-likelihood biases the model towards modeling high-level abstractions. We apply the proposed model to the task of dialogue response generation in two challenging domains: the Ubuntu technical support domain, and Twitter conversations. On Ubuntu, the model outperforms competing approaches by a substantial margin, achieving state-of-the-art results according to both automatic evaluation metrics and a human evaluation study. On Twitter, the model appears to generate more relevant and on-topic responses according to automatic evaluation metrics. Finally, our experiments demonstrate that the proposed model is more adept at overcoming the sparsity of natural language and is better able to capture long-term structure.