Gemma 2: Improving Open Language Models at a Practical Size

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.

Learning rich touch representations through cross-modal self-supervision

The sense of touch is fundamental in several manipulation tasks, but rarely used in robot manipulation. In this work we tackle the problem of learning rich touch features from cross-modal self-supervision. We evaluate them identifying objects and their properties in a few-shot classification setting. Two new datasets are introduced using a simulated anthropomorphic robotic hand equipped with tactile sensors on both synthetic and daily life objects. Several self-supervised learning methods are benchmarked on these datasets, by evaluating few-shot classification on unseen objects and poses. Our experiments indicate that cross-modal self-supervision effectively improves touch representation, and in turn has great potential to enhance robot manipulation skills.

Temporal Difference Uncertainties as a Signal for Exploration

An effective approach to exploration in reinforcement learning is to rely on an agent's uncertainty over the optimal policy, which can yield near-optimal exploration strategies in tabular settings. However, in non-tabular settings that involve function approximators, obtaining accurate uncertainty estimates is almost as challenging a problem. In this paper, we highlight that value estimates are easily biased and temporally inconsistent. In light of this, we propose a novel method for estimating uncertainty over the value function that relies on inducing a distribution over temporal difference errors. This exploration signal controls for state-action transitions so as to isolate uncertainty in value that is due to uncertainty over the agent's parameters. Because our measure of uncertainty conditions on state-action transitions, we cannot act on this measure directly. Instead, we incorporate it as an intrinsic reward and treat exploration as a separate learning problem, induced by the agent's temporal difference uncertainties. We introduce a distinct exploration policy that learns to collect data with high estimated uncertainty, which gives rise to a curriculum that smoothly changes throughout learning and vanishes in the limit of perfect value estimates. We evaluate our method on hard exploration tasks, including Deep Sea and Atari 2600 environments and find that our proposed form of exploration facilitates both diverse and deep exploration.

Small Data, Big Decisions: Model Selection in the Small-Data Regime

Highly overparametrized neural networks can display curiously strong generalization performance - a phenomenon that has recently garnered a wealth of theoretical and empirical research in order to better understand it. In contrast to most previous work, which typically considers the performance as a function of the model size, in this paper we empirically study the generalization performance as the size of the training set varies over multiple orders of magnitude. These systematic experiments lead to some interesting and potentially very useful observations; perhaps most notably that training on smaller subsets of the data can lead to more reliable model selection decisions whilst simultaneously enjoying smaller computational costs. Our experiments furthermore allow us to estimate Minimum Description Lengths for common datasets given modern neural network architectures, thereby paving the way for principled model selection taking into account Occams-razor.

Continual Unsupervised Representation Learning

Continual learning aims to improve the ability of modern learning systems to deal with non-stationary distributions, typically by attempting to learn a series of tasks sequentially. Prior art in the field has largely considered supervised or reinforcement learning tasks, and often assumes full knowledge of task labels and boundaries. In this work, we propose an approach (CURL) to tackle a more general problem that we will refer to as unsupervised continual learning. The focus is on learning representations without any knowledge about task identity, and we explore scenarios when there are abrupt changes between tasks, smooth transitions from one task to another, or even when the data is shuffled. The proposed approach performs task inference directly within the model, is able to dynamically expand to capture new concepts over its lifetime, and incorporates additional rehearsal-based techniques to deal with catastrophic forgetting. We demonstrate the efficacy of CURL in an unsupervised learning setting with MNIST and Omniglot, where the lack of labels ensures no information is leaked about the task. Further, we demonstrate strong performance compared to prior art in an i.i.d setting, or when adapting the technique to supervised tasks such as incremental class learning.

ReConvNet: Video Object Segmentation with Spatio-Temporal Features Modulation

We introduce ReConvNet, a recurrent convolutional architecture for semi-supervised video object segmentation that is able to fast adapt its features to focus on any specific object of interest at inference time. Generalization to new objects never observed during training is known to be a hard task for supervised approaches that would need to be retrained. To tackle this problem, we propose a more efficient solution that learns spatio-temporal features self-adapting to the object of interest via conditional affine transformations. This approach is simple, can be trained end-to-end and does not necessarily require extra training steps at inference time. Our method shows competitive results on DAVIS2016 with respect to state-of-the art approaches that use online fine-tuning, and outperforms them on DAVIS2017. ReConvNet shows also promising results on the DAVIS-Challenge 2018 winning the $10$-th position.

A guide to convolution arithmetic for deep learning

We introduce a guide to help deep learning practitioners understand and manipulate convolutional neural network architectures. The guide clarifies the relationship between various properties (input shape, kernel shape, zero padding, strides and output shape) of convolutional, pooling and transposed convolutional layers, as well as the relationship between convolutional and transposed convolutional layers. Relationships are derived for various cases, and are illustrated in order to make them intuitive.

Multi-View Stereo with Single-View Semantic Mesh Refinement

While 3D reconstruction is a well-established and widely explored research topic, semantic 3D reconstruction has only recently witnessed an increasing share of attention from the Computer Vision community. Semantic annotations allow in fact to enforce strong class-dependent priors, as planarity for ground and walls, which can be exploited to refine the reconstruction often resulting in non-trivial performance improvements. State-of-the art methods propose volumetric approaches to fuse RGB image data with semantic labels; even if successful, they do not scale well and fail to output high resolution meshes. In this paper we propose a novel method to refine both the geometry and the semantic labeling of a given mesh. We refine the mesh geometry by applying a variational method that optimizes a composite energy made of a state-of-the-art pairwise photo-metric term and a single-view term that models the semantic consistency between the labels of the 3D mesh and those of the segmented images. We also update the semantic labeling through a novel Markov Random Field (MRF) formulation that, together with the classical data and smoothness terms, takes into account class-specific priors estimated directly from the annotated mesh. This is in contrast to state-of-the-art methods that are typically based on handcrafted or learned priors. We are the first, jointly with the very recent and seminal work of [M. Blaha et al arXiv:1706.08336, 2017], to propose the use of semantics inside a mesh refinement framework. Differently from [M. Blaha et al arXiv:1706.08336, 2017], which adopts a more classical pairwise comparison to estimate the flow of the mesh, we apply a single-view comparison between the semantically annotated image and the current 3D mesh labels; this improves the robustness in case of noisy segmentations.

PixelVAE: A Latent Variable Model for Natural Images

Natural image modeling is a landmark challenge of unsupervised learning. Variational Autoencoders (VAEs) learn a useful latent representation and model global structure well but have difficulty capturing small details. PixelCNN models details very well, but lacks a latent code and is difficult to scale for capturing large structures. We present PixelVAE, a VAE model with an autoregressive decoder based on PixelCNN. Our model requires very few expensive autoregressive layers compared to PixelCNN and learns latent codes that are more compressed than a standard VAE while still capturing most non-trivial structure. Finally, we extend our model to a hierarchy of latent variables at different scales. Our model achieves state-of-the-art performance on binarized MNIST, competitive performance on 64x64 ImageNet, and high-quality samples on the LSUN bedrooms dataset.

Mollifying Networks

The optimization of deep neural networks can be more challenging than traditional convex optimization problems due to the highly non-convex nature of the loss function, e.g. it can involve pathological landscapes such as saddle-surfaces that can be difficult to escape for algorithms based on simple gradient descent. In this paper, we attack the problem of optimization of highly non-convex neural networks by starting with a smoothed -- or \textit{mollified} -- objective function that gradually has a more non-convex energy landscape during the training. Our proposition is inspired by the recent studies in continuation methods: similar to curriculum methods, we begin learning an easier (possibly convex) objective function and let it evolve during the training, until it eventually goes back to being the original, difficult to optimize, objective function. The complexity of the mollified networks is controlled by a single hyperparameter which is annealed during the training. We show improvements on various difficult optimization tasks and establish a relationship with recent works on continuation methods for neural networks and mollifiers.