This paper introduces a universal federated learning framework that enables over-the-air computation via digital communications, using a new joint source-channel coding scheme. Without relying on channel state information at devices, this scheme employs lattice codes to both quantize model parameters and exploit interference from the devices. A novel two-layer receiver structure at the server is designed to reliably decode an integer combination of the quantized model parameters as a lattice point for the purpose of aggregation. Numerical experiments validate the effectiveness of the proposed scheme. Even with the challenges posed by channel conditions and device heterogeneity, the proposed scheme markedly surpasses other over-the-air FL strategies.
Even though large language models (LLMs) have demonstrated remarkable capability in solving various natural language tasks, the capability of an LLM to follow human instructions is still a concern. Recent works have shown great improvements in the instruction-following capability via additional training for instruction-following tasks. However, the mechanisms responsible for effective instruction-following capabilities remain inadequately understood. Here, we introduce a simplified instruction-following task and use synthetic datasets to analyze a Transformer-based causal language model. Our findings suggest that the model learns task-specific information by clustering data within its hidden space, with this clustering process evolving dynamically during learning. We also demonstrate how this phenomenon assists the model in handling unseen instances and validate our results in a more realistic setting.
Large-scale pretraining and instruction tuning have been successful for training general-purpose language models with broad competencies. However, extending to general-purpose vision-language models is challenging due to the distributional diversity in visual inputs. A recent line of work explores vision-language instruction tuning, taking inspiration from the Query Transformer (QFormer) approach proposed in BLIP-2 models for bridging frozen modalities. However, these approaches rely heavily on large-scale multi-modal pretraining for representation learning before eventual finetuning, incurring a huge computational overhead, poor scaling, and limited accessibility. To that end, we propose a more efficient method for QFormer-based vision-language alignment and demonstrate the effectiveness of our strategy compared to existing baselines in improving the efficiency of vision-language pretraining.
Abid et al. (2021) showed a tendency in GPT-3 to generate violent completions when prompted about Muslims, compared with other religions. Two pre-registered replication attempts found few violent completions and only the weakest anti-Muslim bias in the Instruct version, fine-tuned to eliminate biased and toxic outputs. However, more pre-registered experiments showed that using common names associated with the religions in prompts increases several-fold the rate of violent completions, revealing a highly significant second-order bias against Muslims. Our content analysis revealed religion-specific violent themes containing highly offensive ideas regardless of prompt format. Replications with ChatGPT suggest that any effects of GPT-3's de-biasing have disappeared with continued model development, as this newer model showed both a strong Muslim-violence bias and rates of violent completions closer to Abid et al. (2021). Our results show the need for continual de-biasing of models in ways that address higher-order associations.
Constructing model-agnostic group equivariant networks, such as equitune (Basu et al., 2023b) and its generalizations (Kim et al., 2023), can be computationally expensive for large product groups. We address this by providing efficient model-agnostic equivariant designs for two related problems: one where the network has multiple inputs each with potentially different groups acting on them, and another where there is a single input but the group acting on it is a large product group. For the first design, we initially consider a linear model and characterize the entire equivariant space that satisfies this constraint. This characterization gives rise to a novel fusion layer between different channels that satisfies an invariance-symmetry (IS) constraint, which we call an IS layer. We then extend this design beyond linear models, similar to equitune, consisting of equivariant and IS layers. We also show that the IS layer is a universal approximator of invariant-symmetric functions. Inspired by the first design, we use the notion of the IS property to design a second efficient model-agnostic equivariant design for large product groups acting on a single input. For the first design, we provide experiments on multi-image classification where each view is transformed independently with transformations such as rotations. We find equivariant models are robust to such transformations and perform competitively otherwise. For the second design, we consider three applications: language compositionality on the SCAN dataset to product groups; fairness in natural language generation from GPT-2 to address intersectionality; and robust zero-shot image classification with CLIP. Overall, our methods are simple and general, competitive with equitune and its variants, while also being computationally more efficient.
The Transformer architecture has become prominent in developing large causal language models. However, mechanisms to explain its capabilities are not well understood. Focused on the training process, here we establish a meta-learning view of the Transformer architecture when trained for the causal language modeling task, by explicating an inner optimization process that may happen within the Transformer. Further, from within the inner optimization, we discover and theoretically analyze a special characteristic of the norms of learned token representations within Transformer-based causal language models. Our analysis is supported by experiments conducted on pre-trained large language models and real-world data.
We find limits to the Transformer architecture for language modeling and show it has a universal prediction property in an information-theoretic sense. We further analyze performance in non-asymptotic data regimes to understand the role of various components of the Transformer architecture, especially in the context of data-efficient training. We validate our theoretical analysis with experiments on both synthetic and real datasets.
Efficient transfer learning algorithms are key to the success of foundation models on diverse downstream tasks even with limited data. Recent works of \cite{basu2022equi} and \cite{kaba2022equivariance} propose group averaging (\textit{equitune}) and optimization-based methods, respectively, over features from group-transformed inputs to obtain equivariant outputs from non-equivariant neural networks. While \cite{kaba2022equivariance} are only concerned with training from scratch, we find that equitune performs poorly on equivariant zero-shot tasks despite good finetuning results. We hypothesize that this is because pretrained models provide better quality features for certain transformations than others and simply averaging them is deleterious. Hence, we propose $\lambda$-\textit{equitune} that averages the features using \textit{importance weights}, $\lambda$s. These weights are learned directly from the data using a small neural network, leading to excellent zero-shot and finetuned results that outperform equitune. Further, we prove that $\lambda$-equitune is equivariant and a universal approximator of equivariant functions. Additionally, we show that the method of \cite{kaba2022equivariance} used with appropriate loss functions, which we call \textit{equizero}, also gives excellent zero-shot and finetuned performance. Both equitune and equizero are special cases of $\lambda$-equitune. To show the simplicity and generality of our method, we validate on a wide range of diverse applications and models such as 1) image classification using CLIP, 2) deep Q-learning, 3) fairness in natural language generation (NLG), 4) compositional generalization in languages, and 5) image classification using pretrained CNNs such as Resnet and Alexnet.
Discontinuities can be fairly arbitrary but also cause a significant impact on outcomes in social systems. Indeed, their arbitrariness is why they have been used to infer causal relationships among variables in numerous settings. Regression discontinuity from econometrics assumes the existence of a discontinuous variable that splits the population into distinct partitions to estimate the causal effects of a given phenomenon. Here we consider the design of partitions for a given discontinuous variable to optimize a certain effect previously studied using regression discontinuity. To do so, we propose a quantization-theoretic approach to optimize the effect of interest, first learning the causal effect size of a given discontinuous variable and then applying dynamic programming for optimal quantization design of discontinuities that balance the gain and loss in the effect size. We also develop a computationally-efficient reinforcement learning algorithm for the dynamic programming formulation of optimal quantization. We demonstrate our approach by designing optimal time zone borders for counterfactuals of social capital, social mobility, and health. This is based on regression discontinuity analyses we perform on novel data, which may be of independent empirical interest in showing a causal relationship between sunset time and social capital.
Learning models that are robust to test-time distribution shifts is a key concern in domain generalization, and in the wider context of their real-life applicability. Invariant Risk Minimization (IRM) is one particular framework that aims to learn deep invariant features from multiple domains and has subsequently led to further variants. A key assumption for the success of these methods requires that the underlying causal mechanisms/features remain invariant across domains and the true invariant features be sufficient to learn the optimal predictor. In practical problem settings, these assumptions are often not satisfied, which leads to IRM learning a sub-optimal predictor for that task. In this work, we propose the notion of partial invariance as a relaxation of the IRM framework. Under our problem setting, we first highlight the sub-optimality of the IRM solution. We then demonstrate how partitioning the training domains, assuming access to some meta-information about the domains, can help improve the performance of invariant models via partial invariance. Finally, we conduct several experiments, both in linear settings as well as with classification tasks in language and images with deep models, which verify our conclusions.