Recent work has shown that using unlabeled data in semi-supervised learning is not always beneficial and can even hurt generalization, especially when there is a class mismatch between the unlabeled and labeled examples. We investigate this phenomenon for image classification on the CIFAR-10 and the ImageNet datasets, and with many other forms of domain shifts applied (e.g. salt-and-pepper noise). Our main contribution is Split Batch Normalization (Split-BN), a technique to improve SSL when the additional unlabeled data comes from a shifted distribution. We achieve it by using separate batch normalization statistics for unlabeled examples. Due to its simplicity, we recommend it as a standard practice. Finally, we analyse how domain shift affects the SSL training process. In particular, we find that during training the statistics of hidden activations in late layers become markedly different between the unlabeled and the labeled examples.

Is all of machine learning supervised to some degree? The field of machine learning has traditionally been categorized pedagogically into $supervised~vs~unsupervised~learning$; where supervised learning has typically referred to learning from labeled data, while unsupervised learning has typically referred to learning from unlabeled data. In this paper, we assert that all machine learning is in fact supervised to some degree, and that the scope of supervision is necessarily commensurate to the scope of learning potential. In particular, we argue that clustering algorithms such as k-means, and dimensionality reduction algorithms such as principal component analysis, variational autoencoders, and deep belief networks are each internally supervised by the data itself to learn their respective representations of its features. Furthermore, these algorithms are not capable of external inference until their respective outputs (clusters, principal components, or representation codes) have been identified and externally labeled in effect. As such, they do not suffice as examples of unsupervised learning. We propose that the categorization `supervised vs unsupervised learning' be dispensed with, and instead, learning algorithms be categorized as either $internally~or~externally~supervised$ (or both). We believe this change in perspective will yield new fundamental insights into the structure and character of data and of learning algorithms.

We first pose the Unsupervised Continual Learning (UCL) problem: learning salient representations from a non-stationary stream of unlabeled data in which the number of object classes varies with time. Given limited labeled data just before inference, those representations can also be associated with specific object types to perform classification. To solve the UCL problem, we propose an architecture that involves a single module, called Self-Taught Associative Memory (STAM), which loosely models the function of a cortical column in the mammalian brain. Hierarchies of STAM modules learn based on a combination of Hebbian learning, online clustering, detection of novel patterns, forgetting outliers, and top-down predictions. We illustrate the operation of STAMs in the context of learning handwritten digits in a continual manner with only 3-12 labeled examples per class. STAMs suggest a promising direction to solve the UCL problem without catastrophic forgetting.

Deep neural networks are known to suffer from catastrophic forgetting in class-incremental learning, where the performance on previous tasks drastically degrades when learning a new task. To alleviate this effect, we propose to leverage a continuous and large stream of unlabeled data in the wild. In particular, to leverage such transient external data effectively, we design a novel class-incremental learning scheme with (a) a new distillation loss, termed global distillation, (b) a learning strategy to avoid overfitting to the most recent task, and (c) a sampling strategy for the desired external data. Our experimental results on various datasets, including CIFAR and ImageNet, demonstrate the superiority of the proposed methods over prior methods, particularly when a stream of unlabeled data is accessible: we achieve up to 9.3% of relative performance improvement compared to the state-of-the-art method.

Data augmentation is rapidly gaining attention in machine learning. Synthetic data can be generated by simple transformations or through the data distribution. In the latter case, the main challenge is to estimate the label associated to new synthetic patterns. This paper studies the effect of generating synthetic data by convex combination of patterns and the use of these as unsupervised information in a semi-supervised learning framework with support vector machines, avoiding thus the need to label synthetic examples. We perform experiments on a total of 53 binary classification datasets. Our results show that this type of data over-sampling supports the well-known cluster assumption in semi-supervised learning, showing outstanding results for small high-dimensional datasets and imbalanced learning problems.

Unsupervised learning may sound like a fancy way to say "let the kids learn on their own not to touch the hot oven" but it's actually a pattern-finding technique for mining inspiration from your data. It has nothing to do with machines running around without adult supervision, forming their own opinions about things. This post is beginner-friendly, but assumes you're familiar with the story so far: Check out the six instances above. These photographs are not accompanied by labels. No worries, your brain is pretty good at unsupervised learning.

In traditional generative modeling, good data representation is very often a base for a good machine learning model. It can be linked to good representations encoding more explanatory factors that are hidden in the original data. With the invention of Generative Adversarial Networks (GANs), a subclass of generative models that are able to learn representations in an unsupervised and semi-supervised fashion, we are now able to adversarially learn good mappings from a simple prior distribution to a target data distribution. This paper presents an overview of recent developments in GANs with a focus on learning latent space representations.

With digital video content creation going viral and assuming the bulk of Internet traffic, how can the deluge of video content be analyzed effectively to derive insights and ROI? After all, video is not only huge in size, but it is complex given various visual, audio and temporal elements. Video summarization (a mechanism for generating a short video summary via key frame analysis or video skimming) has become a popular research topic industry-wide and across academia. Video thumbnail generation and summarization has been developed for years, but deep learning and reinforcement learning is changing the landscape and emerging as the winner for optimal frame selection. Recent advances in Generative Adversarial Networks (GANs) are improving the quality, aesthetics and relevancy of the frames to represent the original videos.

Now we have the probability that each data point belongs to each cluster. If we need hard cluster assignments, we can just choose for each data point to belong to the cluster with the highest probability. But the nice thing about EM is that we can embrace the fuzziness of the cluster membership. We can look at a data point and consider the fact that while it most likely belongs to Cluster B, it's also quite likely to belong to Cluster D. This also takes into account the fact that there may not be clear cut boundaries between our clusters. These groups consist of overlapping multi-dimensional distributions, so drawing clear cut lines might not always be the best solution.

PageRank for Semi-Supervised Learning has shown to leverage data structures and limited tagged examples to yield meaningful classification. Despite successes, classification performance can still be improved, particularly in cases of fuzzy graphs or unbalanced labeled data. To address such limitations, a novel approach based on powers of the Laplacian matrix $L^\gamma$ ($\gamma > 0$), referred to as $L^\gamma$-PageRank, is proposed. Its theoretical study shows that it operates on signed graphs, where nodes belonging to one same class are more likely to share positive edges while nodes from different classes are more likely to be connected with negative edges. It is shown that by selecting an optimal $\gamma$, classification performance can be significantly enhanced. A procedure for the automated estimation of the optimal $\gamma$, from a unique observation of data, is devised and assessed. Experiments on several datasets demonstrate the effectiveness of both $L^\gamma$-PageRank classification and the optimal $\gamma$ estimation.