Scalable, Sample Efficient and Interpretable Bayesian Deep Learning

我的工作将集中在EPSRC的“人工智能技术”研究领域。摘要如下。神经网络（NN）是一类机器学习模型，由于其对大量数据的灵活性和可扩展性，最近越来越受欢迎。神经网络最常用的训练方法是使用最大后验（MAP）参数估计。该框架为模型权重提供了点估计，经常导致过拟合、预测过度自信和样本效率低下。此外，神经网络通常表现得像黑盒子。众所周知，他们的预测很难解释。贝叶斯方法提供了一种原则性的方法来解决模型参数的过度置信度问题。在贝叶斯神经网络（BNN）中，权点估计用概率分布代替。通过考虑所有可能的参数值，将权重边缘化来进行预测。在这些模型中，权重空间的不确定性转化为预测的不确定性，为我们提供了一种对“我们不知道的”进行建模的方法。不幸的是，对于复杂模型的精确推理往往是难以处理的，近似推理方法往往依赖于粗糙的近似，这在不确定性估计的准确性和可扩展性之间存在权衡。我的目标是为神经网络开发一类新的灵活的近似推理方法，这些方法能够拟合复杂的数据，产生可靠的不确定性估计，并扩展到大型数据集。我想使用这些模型产生的不确定性估计来自动生成非专家可以理解的关于这些模型决策的解释。最后，我想使用这些近似推理方法来减少模型偏差，并构建基于样本高效模型的强化学习算法。

My work will be focused around the EPSRC's "Artificial intelligence technologies" research area. An abstract can be read bellow.Neural Networks (NN) are a class of machine learning models which have recently soared in popularity due to their flexibility and scalability to large amounts of data. NNs are most commonly trained using Maximum a Posteriori (MAP) parameter estimation. This framework provides point estimates for model weights, often leading to overfitting, overconfidence in predictions and sample inefficiency. Additionally, NNs often behave like black boxes. Their predictions are notoriously difficult to interpret.Bayesian methods provide a principled way of tackling the issue of overconfidence in model parameters. In Bayesian Neural Networks (BNN), weight point estimates are substituted by probability distributions. Predictions are made by marginalising the weights, considering all possible parameter values. In these models, uncertainty in weight space is translated into uncertainty in predictions, giving us a way to model 'what we do not know.' Unfortunately, exact inference is often intractable for complex models and approximate inference methods tend to rely on crude approximations which present a trade-off between accuracy of uncertainty estimates and scalability. My goal is to develop a new class of flexible approximate inference methods for neural networks which are able to fit complex data, produce reliable uncertainty estimates, and scale to large datasets. I want to use the uncertainty estimates produced by these models to automatically generate explanations that are understandable to non-experts about these model's decisions. Finally, I want to use these approximate inference methods to reduce model-bias and build sample-efficient model-based reinforcement learning algorithms.