Alona Fyshe

Han Wang, Archit Sakhadeo, Adam White et al. · deepmind

The performance of reinforcement learning (RL) agents is sensitive to the choice of hyperparameters. In real-world settings like robotics or industrial control systems, however, testing different hyperparameter configurations directly on the environment can be financially prohibitive, dangerous, or time consuming. We propose a new approach to tune hyperparameters from offline logs of data, to fully specify the hyperparameters for an RL agent that learns online in the real world. The approach is conceptually simple: we first learn a model of the environment from the offline data, which we call a calibration model, and then simulate learning in the calibration model to identify promising hyperparameters. We identify several criteria to make this strategy effective, and develop an approach that satisfies these criteria. We empirically investigate the method in a variety of settings to identify when it is effective and when it fails.

Saeed Najafi, Alona Fyshe

Zero-Shot Relation Extraction (ZRE) is the task of Relation Extraction where the training and test sets have no shared relation types. This very challenging domain is a good test of a model's ability to generalize. Previous approaches to ZRE reframed relation extraction as Question Answering (QA), allowing for the use of pre-trained QA models. However, this method required manually creating gold question templates for each new relation. Here, we do away with these gold templates and instead learn a model that can generate questions for unseen relations. Our technique can successfully translate relation descriptions into relevant questions, which are then leveraged to generate the correct tail entity. On tail entity extraction, we outperform the previous state-of-the-art by more than 16 F1 points without using gold question templates. On the RE-QA dataset where no previous baseline for relation extraction exists, our proposed algorithm comes within 0.7 F1 points of a system that uses gold question templates. Our model also outperforms the state-of-the-art ZRE baselines on the FewRel and WikiZSL datasets, showing that QA models no longer need template questions to match the performance of models specifically tailored to the ZRE task. Our implementation is available at https://github.com/fyshelab/QA-ZRE.

Daniela Teodorescu, Alona Fyshe, Saif M. Mohammad

Emerging psychopathology studies are showing that patterns of changes in emotional state -- emotion dynamics -- are associated with overall well-being and mental health. More recently, there has been some work in tracking emotion dynamics through one's utterances, allowing for data to be collected on a larger scale across time and people. However, several questions about how emotion dynamics change with age, especially in children, and when determined through children's writing, remain unanswered. In this work, we use both a lexicon and a machine learning based approach to quantify characteristics of emotion dynamics determined from poems written by children of various ages. We show that both approaches point to similar trends: consistent increasing intensities for some emotions (e.g., anger, fear, joy, sadness, arousal, and dominance) with age and a consistent decreasing valence with age. We also find increasing emotional variability, rise rates (i.e., emotional reactivity), and recovery rates (i.e., emotional regulation) with age. These results act as a useful baselines for further research in how patterns of emotions expressed by children change with age, and their association with mental health.

Bilal Ghanem, Lauren Lutz Coleman, Julia Rivard Dexter et al.

Reading is integral to everyday life, and yet learning to read is a struggle for many young learners. During lessons, teachers can use comprehension questions to increase engagement, test reading skills, and improve retention. Historically such questions were written by skilled teachers, but recently language models have been used to generate comprehension questions. However, many existing Question Generation (QG) systems focus on generating literal questions from the text, and have no way to control the type of the generated question. In this paper, we study QG for reading comprehension where inferential questions are critical and extractive techniques cannot be used. We propose a two-step model (HTA-WTA) that takes advantage of previous datasets, and can generate questions for a specific targeted comprehension skill. We propose a new reading comprehension dataset that contains questions annotated with story-based reading comprehension skills (SBRCS), allowing for a more complete reader assessment. Across several experiments, our results show that HTA-WTA outperforms multiple strong baselines on this new dataset. We show that the HTA-WTA model tests for strong SCRS by asking deep inferential questions.

Daniela Teodorescu, Tiffany Cheng, Alona Fyshe et al.

Research in psychopathology has shown that, at an aggregate level, the patterns of emotional change over time -- emotion dynamics -- are indicators of one's mental health. One's patterns of emotion change have traditionally been determined through self-reports of emotions; however, there are known issues with accuracy, bias, and ease of data collection. Recent approaches to determining emotion dynamics from one's everyday utterances addresses many of these concerns, but it is not yet known whether these measures of utterance emotion dynamics (UED) correlate with mental health diagnoses. Here, for the first time, we study the relationship between tweet emotion dynamics and mental health disorders. We find that each of the UED metrics studied varied by the user's self-disclosed diagnosis. For example: average valence was significantly higher (i.e., more positive text) in the control group compared to users with ADHD, MDD, and PTSD. Valence variability was significantly lower in the control group compared to ADHD, depression, bipolar disorder, MDD, PTSD, and OCD but not PPD. Rise and recovery rates of valence also exhibited significant differences from the control. This work provides important early evidence for how linguistic cues pertaining to emotion dynamics can play a crucial role as biosocial markers for mental illnesses and aid in the understanding, diagnosis, and management of mental health disorders.

Cassidy Pirlot, Richard C. Gerum, Cory Efird et al.

As convolutional neural networks (CNNs) become more accurate at object recognition, their representations become more similar to the primate visual system. This finding has inspired us and other researchers to ask if the implication also runs the other way: If CNN representations become more brain-like, does the network become more accurate? Previous attempts to address this question showed very modest gains in accuracy, owing in part to limitations of the regularization method. To overcome these limitations, we developed a new neural data regularizer for CNNs that uses Deep Canonical Correlation Analysis (DCCA) to optimize the resemblance of the CNN's image representations to that of the monkey visual cortex. Using this new neural data regularizer, we see much larger performance gains in both classification accuracy and within-super-class accuracy, as compared to the previous state-of-the-art neural data regularizers. These networks are also more robust to adversarial attacks than their unregularized counterparts. Together, these results confirm that neural data regularization can push CNN performance higher, and introduces a new method that obtains a larger performance boost.

Liam Peet-Pare, Nidhi Hegde, Alona Fyshe

Fairness researchers in machine learning (ML) have coalesced around several fairness criteria which provide formal definitions of what it means for an ML model to be fair. However, these criteria have some serious limitations. We identify four key shortcomings of these formal fairness criteria, and aim to help to address them by extending performative prediction to include a distributionally robust objective.

Kirby Banman, Liam Peet-Pare, Nidhi Hegde et al.

Most convergence guarantees for stochastic gradient descent with momentum (SGDm) rely on iid sampling. Yet, SGDm is often used outside this regime, in settings with temporally correlated input samples such as continual learning and reinforcement learning. Existing work has shown that SGDm with a decaying step-size can converge under Markovian temporal correlation. In this work, we show that SGDm under covariate shift with a fixed step-size can be unstable and diverge. In particular, we show SGDm under covariate shift is a parametric oscillator, and so can suffer from a phenomenon known as resonance. We approximate the learning system as a time varying system of ordinary differential equations, and leverage existing theory to characterize the system's divergence/convergence as resonant/nonresonant modes. The theoretical result is limited to the linear setting with periodic covariate shift, so we empirically supplement this result to show that resonance phenomena persist even under non-periodic covariate shift, nonlinear dynamics with neural networks, and optimizers other than SGDm.

Dhruv Mullick, Bilal Ghanem, Alona Fyshe

Customer feedback is invaluable to companies as they refine their products. Monitoring customer feedback can be automated with Aspect Level Sentiment Classification (ALSC) which allows us to analyse specific aspects of the products in reviews. Large Language Models (LLMs) are the heart of many state-of-the-art ALSC solutions, but they perform poorly in some scenarios requiring Coreference Resolution (CR). In this work, we propose a framework to improve an LLM's performance on CR-containing reviews by fine tuning on highly inferential tasks. We show that the performance improvement is likely attributed to the improved model CR ability. We also release a new dataset that focuses on CR in ALSC.

Bilal Ghanem, Alona Fyshe

Multiple choice questions (MCQs) are an efficient and common way to assess reading comprehension (RC). Every MCQ needs a set of distractor answers that are incorrect, but plausible enough to test student knowledge. Distractor generation (DG) models have been proposed, and their performance is typically evaluated using machine translation (MT) metrics. However, MT metrics often misjudge the suitability of generated distractors. We propose DISTO: the first learned evaluation metric for generated distractors. We validate DISTO by showing its scores correlate highly with human ratings of distractor quality. At the same time, DISTO ranks the performance of state-of-the-art DG models very differently from MT-based metrics, showing that MT metrics should not be used for distractor evaluation.

Cory Efird, Alex Murphy, Joel Zylberberg et al.

We introduce a method that takes advantage of high-quality pretrained multimodal representations to explore fine-grained semantic networks in the human brain. Previous studies have documented evidence of functional localization in the brain, with different anatomical regions preferentially activating for different types of sensory input. Many such localized structures are known, including the fusiform face area and parahippocampal place area. This raises the question of whether additional brain regions (or conjunctions of brain regions) are also specialized for other important semantic concepts. To identify such brain regions, we developed a data-driven approach to uncover visual concepts that are decodable from a massive functional magnetic resonance imaging (fMRI) dataset. Our analysis is broadly split into three sections. First, a fully connected neural network is trained to map brain responses to the outputs of an image-language foundation model, CLIP (Radford et al., 2021). Subsequently, a contrastive-learning dimensionality reduction method reveals the brain-decodable components of CLIP space. In the final section of our analysis, we localize shared decodable concepts in the brain using a voxel-masking optimization method to produce a shared decodable concept (SDC) space. The accuracy of our procedure is validated by comparing it to previous localization experiments that identify regions for faces, bodies, and places. In addition to these concepts, whose corresponding brain regions were already known, we localize novel concept representations which are shared across participants to other areas of the human brain. We also demonstrate how this method can be used to inspect fine-grained semantic networks for individual participants. We envisage that this extensible method can also be adapted to explore other questions at the intersection of AI and neuroscience.

Richard C. Gerum, Cassidy Pirlot, Alona Fyshe et al.

Recent studies suggest that artificial neural networks (ANNs) that match the spectral properties of the mammalian visual cortex -- namely, the $\sim 1/n$ eigenspectrum of the covariance matrix of neural activities -- achieve higher object recognition performance and robustness to adversarial attacks than those that do not. To our knowledge, however, no previous work systematically explored how modifying the ANN's spectral properties affects performance. To fill this gap, we performed a systematic search over spectral regularizers, forcing the ANN's eigenspectrum to follow $1/n^α$ power laws with different exponents $α$. We found that larger powers (around 2--3) lead to better validation accuracy and more robustness to adversarial attacks on dense networks. This surprising finding applied to both shallow and deep networks and it overturns the notion that the brain-like spectrum (corresponding to $α\sim 1$) always optimizes ANN performance and/or robustness. For convolutional networks, the best $α$ values depend on the task complexity and evaluation metric: lower $α$ values optimized validation accuracy and robustness to adversarial attack for networks performing a simple object recognition task (categorizing MNIST images of handwritten digits); for a more complex task (categorizing CIFAR-10 natural images), we found that lower $α$ values optimized validation accuracy whereas higher $α$ values optimized adversarial robustness. These results have two main implications. First, they cast doubt on the notion that brain-like spectral properties ($α\sim 1$) \emph{always} optimize ANN performance. Second, they demonstrate the potential for fine-tuned spectral regularizers to optimize a chosen design metric, i.e., accuracy and/or robustness.

Dhruv Mullick, Alona Fyshe, Bilal Ghanem

Aspect-based Sentiment Analysis (ABSA) helps to explain customers' opinions towards products and services. In the past, ABSA models were discriminative, but more recently generative models have been used to generate aspects and polarities directly from text. In contrast, discriminative models commonly first select aspects from the text, and then classify the aspect's polarity. Previous results showed that generative models outperform discriminative models on several English ABSA datasets. Here, we evaluate and contrast two state-of-the-art discriminative and generative models in several settings: cross-lingual, cross-domain, and cross-lingual and domain, to understand generalizability in settings other than English mono-lingual in-domain. Our more thorough evaluation shows that, contrary to previous studies, discriminative models can still outperform generative models in almost all settings.

Saeed Najafi, Alona Fyshe

Pre-trained Language Models (PLMs) can be accurately fine-tuned for downstream text processing tasks. Recently, researchers have introduced several parameter-efficient fine-tuning methods that optimize input prompts or adjust a small number of model parameters (e.g LoRA). In this study, we explore the impact of altering the input text of the original task in conjunction with parameter-efficient fine-tuning methods. To most effectively rewrite the input text, we train a few-shot paraphrase model with a Maximum-Marginal Likelihood objective. Using six few-shot text classification datasets, we show that enriching data with paraphrases at train and test time enhances the performance beyond what can be achieved with parameter-efficient fine-tuning alone. The code used for our experiments can be found at https://github.com/SaeedNajafi/RIFF.

Alex Murphy, Joel Zylberberg, Alona Fyshe

Centred Kernel Alignment (CKA) has recently emerged as a popular metric to compare activations from biological and artificial neural networks (ANNs) in order to quantify the alignment between internal representations derived from stimuli sets (e.g. images, text, video) that are presented to both systems. In this paper we highlight issues that the community should take into account if using CKA as an alignment metric with neural data. Neural data are in the low-data high-dimensionality domain, which is one of the cases where (biased) CKA results in high similarity scores even for pairs of random matrices. Using fMRI and MEG data from the THINGS project, we show that if biased CKA is applied to representations of different sizes in the low-data high-dimensionality domain, they are not directly comparable due to biased CKA's sensitivity to differing feature-sample ratios and not stimuli-driven responses. This situation can arise both when comparing a pre-selected area of interest (e.g. ROI) to multiple ANN layers, as well as when determining to which ANN layer multiple regions of interest (ROIs) / sensor groups of different dimensionality are most similar. We show that biased CKA can be artificially driven to its maximum value when using independent random data of different sample-feature ratios. We further show that shuffling sample-feature pairs of real neural data does not drastically alter biased CKA similarity in comparison to unshuffled data, indicating an undesirable lack of sensitivity to stimuli-driven neural responses. Positive alignment of true stimuli-driven responses is only achieved by using debiased CKA. Lastly, we report findings that suggest biased CKA is sensitive to the inherent structure of neural data, only differing from shuffled data when debiased CKA detects stimuli-driven alignment.

Rohan Saha, Abrar Fahim, Alona Fyshe et al.

For specialized domains, there is often not a wealth of data with which to train large machine learning models. In such limited data / compute settings, various methods exist aiming to $\textit{do more with less}$, such as finetuning from a pretrained model, modulating difficulty levels as data are presented to a model (curriculum learning), and considering the role of model type / size. Approaches to efficient $\textit{machine}$ learning also take inspiration from $\textit{human}$ learning by considering use cases where machine learning systems have access to approximately the same number of words experienced by a 13 year old child (100M words). We investigate the role of 3 primary variables in a limited data regime as part of the multimodal track of the BabyLM challenge. We contrast: (i) curriculum learning, (ii), pretraining (with text-only data), (iii) model type. We modulate these variables and assess them on two types of tasks: (a) multimodal (text+image), and (b) unimodal (text-only) tasks. We find that curriculum learning benefits multimodal evaluations over non-curriclum learning models, particularly when combining text-only pretraining. On text-only tasks, curriculum learning appears to help models with smaller trainable parameter counts. We suggest possible reasons based on architectural differences and training designs as to why one might observe such results.

Rohan Saha, Farzane Aminmansour, Alona Fyshe

While transformers can discover latent structure from context, the dynamics of how they acquire different components of the latent structure remain poorly understood. In this work, we use the Alchemy benchmark, to investigate the dynamics of latent structure learning. We train a small decoder-only transformer on three task variants: 1) inferring missing rules from partial contextual information, 2) composing simple rules to solve multi-step sequences, and 3) decomposing complex multi-step examples to infer intermediate steps. By factorizing each task into interpretable events, we show that the model acquires capabilities in discrete stages, first learning the coarse grained rules, before learning the complete latent structure. We also identify a crucial asymmetry, where the model can compose fundamental rules robustly, but struggles to decompose complex examples to discover the fundamental rules. These findings offer new insights into understanding how a transformer model learns latent structures, providing a granular view of how these capabilities evolve during training.

Saeed Najafi, Alona Fyshe

Aligning Large Language Models (LLMs) with human preferences is crucial, but standard methods like Reinforcement Learning from Human Feedback (RLHF) are often complex and unstable. In this work, we propose a new, simpler approach that recasts alignment through the lens of Maximum Marginal Likelihood (MML) estimation. Our new MML based Preference Optimization (MMPO) maximizes the marginal log-likelihood of a preferred text output, using the preference pair as samples for approximation, and forgoes the need for both an explicit reward model and entropy maximization. We theoretically demonstrate that MMPO implicitly performs preference optimization, producing a weighted gradient that naturally up-weights chosen responses over rejected ones. Across models ranging from 135M to 8B parameters, we empirically show that MMPO: 1) is more stable with respect to the hyperparameter $β$ compared to alternative baselines, and 2) achieves competitive or superior preference alignment while better preserving the base model's general language capabilities. Through a series of ablation experiments, we show that this improved performance is indeed attributable to MMPO's implicit preference optimization within the gradient updates.

Qingfeng Lan, Luke Kumar, Martha White et al.

To effectively perform the task of next-word prediction, long short-term memory networks (LSTMs) must keep track of many types of information. Some information is directly related to the next word's identity, but some is more secondary (e.g. discourse-level features or features of downstream words). Correlates of secondary information appear in LSTM representations even though they are not part of an \emph{explicitly} supervised prediction task. In contrast, in reinforcement learning (RL), techniques that explicitly supervise representations to predict secondary information have been shown to be beneficial. Inspired by that success, we propose Predictive Representation Learning (PRL), which explicitly constrains LSTMs to encode specific predictions, like those that might need to be learned implicitly. We show that PRL 1) significantly improves two strong language modeling methods, 2) converges more quickly, and 3) performs better when data is limited. Our work shows that explicitly encoding a simple predictive task facilitates the search for a more effective language model.

Maryam Hashemzadeh, Greta Kaufeld, Martha White et al.

The representations generated by many models of language (word embeddings, recurrent neural networks and transformers) correlate to brain activity recorded while people read. However, these decoding results are usually based on the brain's reaction to syntactically and semantically sound language stimuli. In this study, we asked: how does an LSTM (long short term memory) language model, trained (by and large) on semantically and syntactically intact language, represent a language sample with degraded semantic or syntactic information? Does the LSTM representation still resemble the brain's reaction? We found that, even for some kinds of nonsensical language, there is a statistically significant relationship between the brain's activity and the representations of an LSTM. This indicates that, at least in some instances, LSTMs and the human brain handle nonsensical data similarly.

Qingfeng Lan, Yangchen Pan, Alona Fyshe et al.

Q-learning suffers from overestimation bias, because it approximates the maximum action value using the maximum estimated action value. Algorithms have been proposed to reduce overestimation bias, but we lack an understanding of how bias interacts with performance, and the extent to which existing algorithms mitigate bias. In this paper, we 1) highlight that the effect of overestimation bias on learning efficiency is environment-dependent; 2) propose a generalization of Q-learning, called \emph{Maxmin Q-learning}, which provides a parameter to flexibly control bias; 3) show theoretically that there exists a parameter choice for Maxmin Q-learning that leads to unbiased estimation with a lower approximation variance than Q-learning; and 4) prove the convergence of our algorithm in the tabular case, as well as convergence of several previous Q-learning variants, using a novel Generalized Q-learning framework. We empirically verify that our algorithm better controls estimation bias in toy environments, and that it achieves superior performance on several benchmark problems.

Callie Federer, Haoyan Xu, Alona Fyshe et al.

The current state-of-the-art object recognition algorithms, deep convolutional neural networks (DCNNs), are inspired by the architecture of the mammalian visual system, and are capable of human-level performance on many tasks. However, even these algorithms make errors. As they are trained for object recognition tasks, it has been shown that DCNNs develop hidden representations that resemble those observed in the mammalian visual system. Moreover, DCNNs trained on object recognition tasks are currently among the best models we have of the mammalian visual system. This led us to hypothesize that teaching DCNNs to achieve even more brain-like representations could improve their performance. To test this, we trained DCNNs on a composite task, wherein networks were trained to: a) classify images of objects; while b) having intermediate representations that resemble those observed in neural recordings from monkey visual cortex. Compared with DCNNs trained purely for object categorization, DCNNs trained on the composite task had better object recognition performance and are more robust to label corruption. Interestingly, we also found that neural data was not required, but randomized data with the same statistics as neural data also boosted performance. Our results outline a new way to train object recognition networks, using strategies in which the brain - or at least the statistical properties of its activation patterns - serves as a teacher signal for training DCNNs.