A lecture exploring alternatives to using labeled training data. Labeled training data is often scarce, unavailable, or can be very costly to obtain. To circumvent this problem, there is a growing interest in developing methods that can exploit sources of information other than labeled data, such as weak-supervision and zero-shot learning. While these techniques obtained impressive accuracy in practice, both for vision and language domains, they come with no theoretical characterization of their accuracy. In a sequence of recent works, we develop a rigorous mathematical framework for constructing and analyzing algorithms that combine multiple sources of related data to solve a new learning task. Our learning algorithms provably converge to models that have minimum empirical risk with respect to an adversarial choice over feasible labelings for a set of unlabeled data, where the feasibility of a labeling is computed through constraints defined by estimated statistics of the sources. Notably, these methods do not require the related sources to have the same labeling space as the multiclass classification task. We demonstrate the effectiveness of our approach with experimentations on various image classification tasks. Creative Commons Attribution-Non-Commercial-Share Alike 2.0 UK: England & Wales; http://creativecommons.org/licenses/by-nc-sa/2.0/uk/
All content for Department of Statistics is the property of Oxford University and is served directly from their servers
with no modification, redirects, or rehosting. The podcast is not affiliated with or endorsed by Podjoint in any way.
A lecture exploring alternatives to using labeled training data. Labeled training data is often scarce, unavailable, or can be very costly to obtain. To circumvent this problem, there is a growing interest in developing methods that can exploit sources of information other than labeled data, such as weak-supervision and zero-shot learning. While these techniques obtained impressive accuracy in practice, both for vision and language domains, they come with no theoretical characterization of their accuracy. In a sequence of recent works, we develop a rigorous mathematical framework for constructing and analyzing algorithms that combine multiple sources of related data to solve a new learning task. Our learning algorithms provably converge to models that have minimum empirical risk with respect to an adversarial choice over feasible labelings for a set of unlabeled data, where the feasibility of a labeling is computed through constraints defined by estimated statistics of the sources. Notably, these methods do not require the related sources to have the same labeling space as the multiclass classification task. We demonstrate the effectiveness of our approach with experimentations on various image classification tasks. Creative Commons Attribution-Non-Commercial-Share Alike 2.0 UK: England & Wales; http://creativecommons.org/licenses/by-nc-sa/2.0/uk/
Rapid training of deep neural networks without skip connections or normalization layers using Deep Kernel Shaping. Using an extended and formalized version of the Q/C map analysis of Pool et al. (2016), along with Neural Tangent Kernel theory, we identify the main pathologies present in deep networks that prevent them from training fast and generalizing to unseen data, and show how these can be avoided by carefully controlling the "shape" of the network's initialization-time kernel function. We then develop a method called Deep Kernel Shaping (DKS), which accomplishes this using a combination of precise parameter initialization, activation function transformations, and small architectural tweaks, all of which preserve the model class. In our experiments we show that DKS enables SGD training of residual networks without normalization layers on Imagenet and CIFAR-10 classification tasks at speeds comparable to standard ResNetV2 and Wide-ResNet models, with only a small decrease in generalization performance. And when using K-FAC as the optimizer, we achieve similar results for networks without skip connections. Our results apply for a large variety of activation functions, including those which traditionally perform very badly, such as the logistic sigmoid. In addition to DKS, we contribute a detailed analysis of skip connections, normalization layers, special activation functions like RELU and SELU, and various initialization schemes, explaining their effectiveness as alternative (and ultimately incomplete) ways of "shaping" the network's initialization-time kernel. Creative Commons Attribution-Non-Commercial-Share Alike 2.0 UK: England & Wales; http://creativecommons.org/licenses/by-nc-sa/2.0/uk/
Department of Statistics
A lecture exploring alternatives to using labeled training data. Labeled training data is often scarce, unavailable, or can be very costly to obtain. To circumvent this problem, there is a growing interest in developing methods that can exploit sources of information other than labeled data, such as weak-supervision and zero-shot learning. While these techniques obtained impressive accuracy in practice, both for vision and language domains, they come with no theoretical characterization of their accuracy. In a sequence of recent works, we develop a rigorous mathematical framework for constructing and analyzing algorithms that combine multiple sources of related data to solve a new learning task. Our learning algorithms provably converge to models that have minimum empirical risk with respect to an adversarial choice over feasible labelings for a set of unlabeled data, where the feasibility of a labeling is computed through constraints defined by estimated statistics of the sources. Notably, these methods do not require the related sources to have the same labeling space as the multiclass classification task. We demonstrate the effectiveness of our approach with experimentations on various image classification tasks. Creative Commons Attribution-Non-Commercial-Share Alike 2.0 UK: England & Wales; http://creativecommons.org/licenses/by-nc-sa/2.0/uk/
