This work explores the degree to which grammar acquisition is driven by language `simplicity' and the source modality (speech vs. text) of data. Using BabyBERTa as a probe, we find that grammar acquisition is largely driven by exposure to speech data, and in particular through exposure to two of the BabyLM training corpora: AO-Childes and Open Subtitles. We arrive at this finding by examining various ways of presenting input data to our model. First, we assess the impact of various sequence-level complexity based curricula. We then examine the impact of learning over `blocks' -- covering spans of text that are balanced for the number of tokens in each of the source corpora (rather than number of lines). Finally, we explore curricula that vary the degree to which the model is exposed to different corpora. In all cases, we find that over-exposure to AO-Childes and Open Subtitles significantly drives performance. We verify these findings through a comparable control dataset in which exposure to these corpora, and speech more generally, is limited by design. Our findings indicate that it is not the proportion of tokens occupied by high-utility data that aids acquisition, but rather the proportion of training steps assigned to such data. We hope this encourages future research into the use of more developmentally plausible linguistic data (which tends to be more scarce) to augment general purpose pre-training regimes.
Solving program induction problems requires searching through an enormous space of possibilities. DreamCoder is an inductive program synthesis system that, whilst solving problems, learns to simplify search in an iterative wake-sleep procedure. The cost of search is amortised by training a neural search policy, reducing search breadth and effectively "compiling" useful information to compose program solutions across tasks. Additionally, a library of program components is learnt to express discovered solutions in fewer components, reducing search depth. In DreamCoder, the neural search policy has only an indirect effect on the library learnt through the program solutions it helps discover. We present an approach for library learning that directly leverages the neural search policy, effectively "decompiling" its amortised knowledge to extract relevant program components. This provides stronger amortised inference: the amortised knowledge learnt to reduce search breadth is now also used to reduce search depth. We integrate our approach with DreamCoder and demonstrate faster domain proficiency with improved generalisation on a range of domains, particularly when fewer example solutions are available.
Conditional neural processes (CNPs) are a flexible and efficient family of models that learn to learn a stochastic process from observations. In the visual domain, they have seen particular application in contextual image completion - observing pixel values at some locations to predict a distribution over values at other unobserved locations. However, the choice of pixels in learning such a CNP is typically either random or derived from a simple statistical measure (e.g. pixel variance). Here, we turn the problem on its head and ask: which pixels would a CNP like to observe? That is, which pixels allow fitting CNP, and do such pixels tell us something about the underlying image? Viewing the context provided to the CNP as fixed-size latent representations, we construct an amortised variational framework, Partial Pixel Space Variational Autoencoder (PPS-VAE), for predicting this context simultaneously with learning a CNP. We evaluate PPS-VAE on a set of vision datasets, and find that not only is it possible to learn context points while also fitting CNPs, but that their spatial arrangement and values provides strong signal for the information contained in the image - evaluated through the lens of classification. We believe the PPS-VAE provides a promising avenue to explore learning interpretable and effective visual representations.
This work explores the utility of explicit structure for representation learning in NLP by developing StrAE -- an autoencoding framework that faithfully leverages sentence structure to learn multi-level node embeddings in an unsupervised fashion. We use StrAE to train models across different types of sentential structure and objectives, including a novel contrastive loss over structure, and evaluate the learnt embeddings on a series of both intrinsic and extrinsic tasks. Our experiments indicate that leveraging explicit structure through StrAE leads to improved embeddings over prior work, and that our novel contrastive objective over structure outperforms the standard cross-entropy objective. Moreover, in contrast to findings from prior work that weakly leverages structure, we find that being completely faithful to structure does enable disambiguation between types of structure based on the corresponding model's performance. As further evidence of StrAE's utility, we develop a simple proof-of-concept approach to simultaneously induce structure while learning embeddings, rather than being given structure, and find that performance is comparable to that of the best-performing models where structure is given. Finally, we contextualise these results by comparing StrAE against standard unstructured baselines learnt in similar settings, and show that faithfully leveraging explicit structure can be beneficial in lexical and sentence-level semantics.
Learning general-purpose representations from perceptual inputs is a hallmark of human intelligence. For example, people can write out numbers or characters, or even draw doodles, by characterizing these tasks as different instantiations of the same generic underlying process -- compositional arrangements of different forms of pen strokes. Crucially, learning to do one task, say writing, implies reasonable competence at another, say drawing, on account of this shared process. We present Drawing out of Distribution (DooD), a neuro-symbolic generative model of stroke-based drawing that can learn such general-purpose representations. In contrast to prior work, DooD operates directly on images, requires no supervision or expensive test-time inference, and performs unsupervised amortised inference with a symbolic stroke model that better enables both interpretability and generalization. We evaluate DooD on its ability to generalise across both data and tasks. We first perform zero-shot transfer from one dataset (e.g. MNIST) to another (e.g. Quickdraw), across five different datasets, and show that DooD clearly outperforms different baselines. An analysis of the learnt representations further highlights the benefits of adopting a symbolic stroke model. We then adopt a subset of the Omniglot challenge tasks, and evaluate its ability to generate new exemplars (both unconditionally and conditionally), and perform one-shot classification, showing that DooD matches the state of the art. Taken together, we demonstrate that DooD does indeed capture general-purpose representations across both data and task, and takes a further step towards building general and robust concept-learning systems.
We propose ADIOS, a masked image model (MIM) framework for self-supervised learning, which simultaneously learns a masking function and an image encoder using an adversarial objective. The image encoder is trained to minimise the distance between representations of the original and that of a masked image. The masking function, conversely, aims at maximising this distance. ADIOS consistently improves on state-of-the-art self-supervised learning (SSL) methods on a variety of tasks and datasets -- including classification on ImageNet100 and STL10, transfer learning on CIFAR10/100, Flowers102 and iNaturalist, as well as robustness evaluated on the backgrounds challenge (Xiao et al., 2021) -- while generating semantically meaningful masks. Unlike modern MIM models such as MAE, BEiT and iBOT, ADIOS does not rely on the image-patch tokenisation construction of Vision Transformers, and can be implemented with convolutional backbones. We further demonstrate that the masks learned by ADIOS are more effective in improving representation learning of SSL methods than masking schemes used in popular MIM models.
Multimodal VAEs seek to model the joint distribution over heterogeneous data (e.g.\ vision, language), whilst also capturing a shared representation across such modalities. Prior work has typically combined information from the modalities by reconciling idiosyncratic representations directly in the recognition model through explicit products, mixtures, or other such factorisations. Here we introduce a novel alternative, the MEME, that avoids such explicit combinations by repurposing semi-supervised VAEs to combine information between modalities implicitly through mutual supervision. This formulation naturally allows learning from partially-observed data where some modalities can be entirely missing -- something that most existing approaches either cannot handle, or do so to a limited extent. We demonstrate that MEME outperforms baselines on standard metrics across both partial and complete observation schemes on the MNIST-SVHN (image-image) and CUB (image-text) datasets. We also contrast the quality of the representations learnt by mutual supervision against standard approaches and observe interesting trends in its ability to capture relatedness between data.
We introduce a simple and effective method for learning VAEs with controllable inductive biases by using an intermediary set of latent variables. This allows us to overcome the limitations of the standard Gaussian prior assumption. In particular, it allows us to impose desired properties like sparsity or clustering on learned representations, and incorporate prior information into the learned model. Our approach, which we refer to as the Intermediary Latent Space VAE (InteL-VAE), is based around controlling the stochasticity of the encoding process with the intermediary latent variables, before deterministically mapping them forward to our target latent representation, from which reconstruction is performed. This allows us to maintain all the advantages of the traditional VAE framework, while incorporating desired prior information, inductive biases, and even topological information through the latent mapping. We show that this, in turn, allows InteL-VAEs to learn both better generative models and representations.
Machine learning systems typically assume that the distributions of training and test sets match closely. However, a critical requirement of such systems in the real world is their ability to generalize to unseen domains. Here, we propose an inter-domain gradient matching objective that targets domain generalization by maximizing the inner product between gradients from different domains. Since direct optimization of the gradient inner product can be computationally prohibitive -- requires computation of second-order derivatives -- we derive a simpler first-order algorithm named Fish that approximates its optimization. We demonstrate the efficacy of Fish on 6 datasets from the Wilds benchmark, which captures distribution shift across a diverse range of modalities. Our method produces competitive results on these datasets and surpasses all baselines on 4 of them. We perform experiments on both the Wilds benchmark, which captures distribution shift in the real world, as well as datasets in DomainBed benchmark that focuses more on synthetic-to-real transfer. Our method produces competitive results on both benchmarks, demonstrating its effectiveness across a wide range of domain generalization tasks.