Transfer Learning for Segmentation Problems: Choose the Right Encoder and Skip the Decoder

It is common practice to reuse models initially trained on different data to increase downstream task performance. Especially in the computer vision domain, ImageNet-pretrained weights have been successfully used for various tasks. In this work, we investigate the impact of transfer learning for segmentation problems, being pixel-wise classification problems that can be tackled with encoder-decoder architectures. We find that transfer learning the decoder does not help downstream segmentation tasks, while transfer learning the encoder is truly beneficial. We demonstrate that pretrained weights for a decoder may yield faster convergence, but they do not improve the overall model performance as one can obtain equivalent results with randomly initialized decoders. However, we show that it is more effective to reuse encoder weights trained on a segmentation or reconstruction task than reusing encoder weights trained on classification tasks. This finding implicates that using ImageNet-pretrained encoders for downstream segmentation problems is suboptimal. We also propose a contrastive self-supervised approach with multiple self-reconstruction tasks, which provides encoders that are suitable for transfer learning in segmentation problems in the absence of segmentation labels.

Similarity of Pre-trained and Fine-tuned Representations

In transfer learning, only the last part of the networks - the so-called head - is often fine-tuned. Representation similarity analysis shows that the most significant change still occurs in the head even if all weights are updatable. However, recent results from few-shot learning have shown that representation change in the early layers, which are mostly convolutional, is beneficial, especially in the case of cross-domain adaption. In our paper, we find out whether that also holds true for transfer learning. In addition, we analyze the change of representation in transfer learning, both during pre-training and fine-tuning, and find out that pre-trained structure is unlearned if not usable.

A modular framework for object-based saccadic decisions in dynamic scenes

Visually exploring the world around us is not a passive process. Instead, we actively explore the world and acquire visual information over time. Here, we present a new model for simulating human eye-movement behavior in dynamic real-world scenes. We model this active scene exploration as a sequential decision making process. We adapt the popular drift-diffusion model (DDM) for perceptual decision making and extend it towards multiple options, defined by objects present in the scene. For each possible choice, the model integrates evidence over time and a decision (saccadic eye movement) is triggered as soon as evidence crosses a decision threshold. Drawing this explicit connection between decision making and object-based scene perception is highly relevant in the context of active viewing, where decisions are made continuously while interacting with an external environment. We validate our model with a carefully designed ablation study and explore influences of our model parameters. A comparison on the VidCom dataset supports the plausibility of the proposed approach.

Exploring the Similarity of Representations in Model-Agnostic Meta-Learning

In past years model-agnostic meta-learning (MAML) has been one of the most promising approaches in meta-learning. It can be applied to different kinds of problems, e.g., reinforcement learning, but also shows good results on few-shot learning tasks. Besides their tremendous success in these tasks, it has still not been fully revealed yet, why it works so well. Recent work proposes that MAML rather reuses features than rapidly learns. In this paper, we want to inspire a deeper understanding of this question by analyzing MAML's representation. We apply representation similarity analysis (RSA), a well-established method in neuroscience, to the few-shot learning instantiation of MAML. Although some part of our analysis supports their general results that feature reuse is predominant, we also reveal arguments against their conclusion. The similarity-increase of layers closer to the input layers arises from the learning task itself and not from the model. In addition, the representations after inner gradient steps make a broader change to the representation than the changes during meta-training.

Training Generative Networks with general Optimal Transport distances

We propose a new algorithm that uses an auxiliary Neural Network to calculate the transport distance between two data distributions and export an optimal transport map. In the sequel we use the aforementioned map to train Generative Networks. Unlike WGANs, where the Euclidean distance is implicitly used, this new method allows to use any transportation cost function that can be chosen to match the problem at hand. More specifically, it allows to use the squared distance as a transportation cost function, giving rise to the Wasserstein-2 metric for probability distributions, which has rich geometric properties that result in fast and stable gradients descends. It also allows to use image centered distances, like the Structure Similarity index, with notable differences in the results.

The NIGENS General Sound Events Database

Computational auditory scene analysis is gaining interest in the last years. Trailing behind the more mature field of speech recognition, it is particularly general sound event detection that is attracting increasing attention. Crucial for training and testing reasonable models is having available enough suitable data -- until recently, general sound event databases were hardly found. We release and present a database with 714 wav files containing isolated high quality sound events of 14 different types, plus 303 `general' wav files of anything else but these 14 types. All sound events are strongly labeled with perceptual on- and offset times, paying attention to omitting in-between silences. The amount of isolated sound events, the quality of annotations, and the particular general sound class distinguish NIGENS from other databases.

Non-Deterministic Policy Improvement Stabilizes Approximated Reinforcement Learning

This paper investigates a type of instability that is linked to the greedy policy improvement in approximated reinforcement learning. We show empirically that non-deterministic policy improvement can stabilize methods like LSPI by controlling the improvements' stochasticity. Additionally we show that a suitable representation of the value function also stabilizes the solution to some degree. The presented approach is simple and should also be easily transferable to more sophisticated algorithms like deep reinforcement learning.

Regression with Linear Factored Functions

Many applications that use empirically estimated functions face a curse of dimensionality, because the integrals over most function classes must be approximated by sampling. This paper introduces a novel regression-algorithm that learns linear factored functions (LFF). This class of functions has structural properties that allow to analytically solve certain integrals and to calculate point-wise products. Applications like belief propagation and reinforcement learning can exploit these properties to break the curse and speed up computation. We derive a regularized greedy optimization scheme, that learns factored basis functions during training. The novel regression algorithm performs competitively to Gaussian processes on benchmark tasks, and the learned LFF functions are with 4-9 factored basis functions on average very compact.

Risk-sensitive Markov control processes

We introduce a general framework for measuring risk in the context of Markov control processes with risk maps on general Borel spaces that generalize known concepts of risk measures in mathematical finance, operations research and behavioral economics. Within the framework, applying weighted norm spaces to incorporate also unbounded costs, we study two types of infinite-horizon risk-sensitive criteria, discounted total risk and average risk, and solve the associated optimization problems by dynamic programming. For the discounted case, we propose a new discount scheme, which is different from the conventional form but consistent with the existing literature, while for the average risk criterion, we state Lyapunov-like stability conditions that generalize known conditions for Markov chains to ensure the existence of solutions to the optimality equation.

Risk-sensitive Reinforcement Learning

We derive a family of risk-sensitive reinforcement learning methods for agents, who face sequential decision-making tasks in uncertain environments. By applying a utility function to the temporal difference (TD) error, nonlinear transformations are effectively applied not only to the received rewards but also to the true transition probabilities of the underlying Markov decision process. When appropriate utility functions are chosen, the agents' behaviors express key features of human behavior as predicted by prospect theory (Kahneman and Tversky, 1979), for example different risk-preferences for gains and losses as well as the shape of subjective probability curves. We derive a risk-sensitive Q-learning algorithm, which is necessary for modeling human behavior when transition probabilities are unknown, and prove its convergence. As a proof of principle for the applicability of the new framework we apply it to quantify human behavior in a sequential investment task. We find, that the risk-sensitive variant provides a significantly better fit to the behavioral data and that it leads to an interpretation of the subject's responses which is indeed consistent with prospect theory. The analysis of simultaneously measured fMRI signals show a significant correlation of the risk-sensitive TD error with BOLD signal change in the ventral striatum. In addition we find a significant correlation of the risk-sensitive Q-values with neural activity in the striatum, cingulate cortex and insula, which is not present if standard Q-values are used.