Forecasting the state of vegetation in response to climate and weather events is a major challenge. Its implementation will prove crucial in predicting crop yield, forest damage, or more generally the impact on ecosystems services relevant for socio-economic functioning, which if absent can lead to humanitarian disasters. Vegetation status depends on weather and environmental conditions that modulate complex ecological processes taking place at several timescales. Interactions between vegetation and different environmental drivers express responses at instantaneous but also time-lagged effects, often showing an emerging spatial context at landscape and regional scales. We formulate the land surface forecasting task as a strongly guided video prediction task where the objective is to forecast the vegetation developing at very fine resolution using topography and weather variables to guide the prediction. We use a Convolutional LSTM (ConvLSTM) architecture to address this task and predict changes in the vegetation state in Africa using Sentinel-2 satellite NDVI, having ERA5 weather reanalysis, SMAP satellite measurements, and topography (DEM of SRTMv4.1) as variables to guide the prediction. Ours results highlight how ConvLSTM models can not only forecast the seasonal evolution of NDVI at high resolution, but also the differential impacts of weather anomalies over the baselines. The model is able to predict different vegetation types, even those with very high NDVI variability during target length, which is promising to support anticipatory actions in the context of drought-related disasters.

We study the problem of planning under model uncertainty in an online meta-reinforcement learning (RL) setting where an agent is presented with a sequence of related tasks with limited interactions per task. The agent can use its experience in each task and across tasks to estimate both the transition model and the distribution over tasks. We propose an algorithm to meta-learn the underlying structure across tasks, utilize it to plan in each task, and upper-bound the regret of the planning loss. Our bound suggests that the average regret over tasks decreases as the number of tasks increases and as the tasks are more similar. In the classical single-task setting, it is known that the planning horizon should depend on the estimated model's accuracy, that is, on the number of samples within task. We generalize this finding to meta-RL and study this dependence of planning horizons on the number of tasks. Based on our theoretical findings, we derive heuristics for selecting slowly increasing discount factors, and we validate its significance empirically.

Self-supervised learning is a popular and powerful method for utilizing large amounts of unlabeled data, for which a wide variety of training objectives have been proposed in the literature. In this study, we perform a Bayesian analysis of state-of-the-art self-supervised learning objectives and propose a unified formulation based on likelihood learning. Our analysis suggests a simple method for integrating self-supervised learning with generative models, allowing for the joint training of these two seemingly distinct approaches. We refer to this combined framework as GEDI, which stands for GEnerative and DIscriminative training. Additionally, we demonstrate an instantiation of the GEDI framework by integrating an energy-based model with a cluster-based self-supervised learning model. Through experiments on synthetic and real-world data, including SVHN, CIFAR10, and CIFAR100, we show that GEDI outperforms existing self-supervised learning strategies in terms of clustering performance by a wide margin. We also demonstrate that GEDI can be integrated into a neural-symbolic framework to address tasks in the small data regime, where it can use logical constraints to further improve clustering and classification performance.

State-of-the-art performance in electroencephalography (EEG) decoding tasks is currently often achieved with either Deep-Learning or Riemannian-Geometry-based decoders. Recently, there is growing interest in Deep Riemannian Networks (DRNs) possibly combining the advantages of both previous classes of methods. However, there are still a range of topics where additional insight is needed to pave the way for a more widespread application of DRNs in EEG. These include architecture design questions such as network size and end-to-end ability as well as model training questions. How these factors affect model performance has not been explored. Additionally, it is not clear how the data within these networks is transformed, and whether this would correlate with traditional EEG decoding. Our study aims to lay the groundwork in the area of these topics through the analysis of DRNs for EEG with a wide range of hyperparameters. Networks were tested on two public EEG datasets and compared with state-of-the-art ConvNets. Here we propose end-to-end EEG SPDNet (EE(G)-SPDNet), and we show that this wide, end-to-end DRN can outperform the ConvNets, and in doing so use physiologically plausible frequency regions. We also show that the end-to-end approach learns more complex filters than traditional band-pass filters targeting the classical alpha, beta, and gamma frequency bands of the EEG, and that performance can benefit from channel specific filtering approaches. Additionally, architectural analysis revealed areas for further improvement due to the possible loss of Riemannian specific information throughout the network. Our study thus shows how to design and train DRNs to infer task-related information from the raw EEG without the need of handcrafted filterbanks and highlights the potential of end-to-end DRNs such as EE(G)-SPDNet for high-performance EEG decoding.

In-context learning (ICL) enables large language models (LLMs) to perform new tasks by prompting them with a sequence of training examples. However, ICL is very sensitive to the choice of training examples: randomly sampling examples from a training set leads to high variance in performance. In this paper, we show that curating a carefully chosen subset of training data greatly stabilizes ICL performance. We propose two methods to choose training subsets, both of which score training examples individually and then select the highest-scoring ones. CondAcc scores a training example by its average ICL accuracy when combined with random training examples, while Datamodels learns a linear proxy model that estimates how the presence of each training example influences LLM accuracy. On average, CondAcc and Datamodels outperform sampling from the entire training set by 7.7% and 6.3%, respectively, across 5 tasks and two LLMs. Our analysis shows that stable subset examples are no more diverse than average, and are not outliers in terms of sequence length and perplexity.

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

Underwater images are altered by the physical characteristics of the medium through which light rays pass before reaching the optical sensor. Scattering and strong wavelength-dependent absorption significantly modify the captured colors depending on the distance of observed elements to the image plane. In this paper, we aim to recover the original colors of the scene as if the water had no effect on them. We propose two novel methods that rely on different sets of inputs. The first assumes that pixel intensities in the restored image are normally distributed within each color channel, leading to an alternative optimization of the well-known \textit{Sea-thru} method which acts on single images and their distance maps. We additionally introduce SUCRe, a new method that further exploits the scene's 3D Structure for Underwater Color Restoration. By following points in multiple images and tracking their intensities at different distances to the sensor we constrain the optimization of the image formation model parameters. When compared to similar existing approaches, SUCRe provides clear improvements in a variety of scenarios ranging from natural light to deep-sea environments. The code for both approaches is publicly available at https://github.com/clementinboittiaux/sucre .

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

As causal inference becomes more widespread the importance of having good tools to test for causal effects increases. In this work we focus on the problem of testing for causal effects that manifest in a difference in distribution for treatment and control. We build on work applying kernel methods to causality, considering the previously introduced Counterfactual Mean Embedding framework (\textsc{CfME}). We improve on this by proposing the \emph{Doubly Robust Counterfactual Mean Embedding} (\textsc{DR-CfME}), which has better theoretical properties than its predecessor by leveraging semiparametric theory. This leads us to propose new kernel based test statistics for distributional effects which are based upon doubly robust estimators of treatment effects. We propose two test statistics, one which is a direct improvement on previous work and one which can be applied even when the support of the treatment arm is a subset of that of the control arm. We demonstrate the validity of our methods on simulated and real-world data, as well as giving an application in off-policy evaluation.

We study representation learning for efficient imitation learning over linear systems. In particular, we consider a setting where learning is split into two phases: (a) a pre-training step where a shared $k$-dimensional representation is learned from $H$ source policies, and (b) a target policy fine-tuning step where the learned representation is used to parameterize the policy class. We find that the imitation gap over trajectories generated by the learned target policy is bounded by $\tilde{O}\left( \frac{k n_x}{HN_{\mathrm{shared}}} + \frac{k n_u}{N_{\mathrm{target}}}\right)$, where $n_x > k$ is the state dimension, $n_u$ is the input dimension, $N_{\mathrm{shared}}$ denotes the total amount of data collected for each policy during representation learning, and $N_{\mathrm{target}}$ is the amount of target task data. This result formalizes the intuition that aggregating data across related tasks to learn a representation can significantly improve the sample efficiency of learning a target task. The trends suggested by this bound are corroborated in simulation.