Previous transfer learning methods based on deep network assume the knowledge should be transferred between the same hidden layers of the source domain and the target domains. This assumption doesn't always hold true, especially when the data from the two domains are heterogeneous with different resolutions. In such case, the most suitable numbers of layers for the source domain data and the target domain data would differ. As a result, the high level knowledge from the source domain would be transferred to the wrong layer of target domain. Based on this observation, "where to transfer" proposed in this paper should be a novel research frontier. We propose a new mathematic model named DT -LET to solve this heterogeneous transfer learning problem. In order to select the best matching of layers to transfer knowledge, we define specific loss function to estimate the corresponding relationship between high-level features of data in the source domain and the target domain. To verify this proposed cross-layer model, experiments for two cross-domain recognition/classification tasks are conducted, and the achieved superior results demonstrate the necessity of layer correspondence searching.

Transfer-learning and meta-learning are two effective methods to apply knowledge learned from large data sources to new tasks. In few-class, few-shot target task settings (i.e. when there are only a few classes and training examples available in the target task), meta-learning approaches that optimize for future task learning have outperformed the typical transfer approach of initializing model weights from a pre-trained starting point. But as we experimentally show, meta-learning algorithms that work well in the few-class setting do not generalize well in many-shot and many-class cases. In this paper, we propose a joint training approach that combines both transfer-learning and meta-learning. Benefiting from the advantages of each, our method obtains improved generalization performance on unseen target tasks in both few- and many-class and few- and many-shot scenarios.

In this special guest feature, Ellexus CEO Rosemary Francis writes that the convergence of HPC and Ai is already changing the landscape in the world of supercomputing. AI and machine learning are hot words at the moment in many sectors. We're all very aware just how much a company can predict about us by plugging information into a cunning algorithm and there are thousands of start-ups racing to use machine learning to create the next big-win app. As the dust settles on the initial hype, we are starting to see some of the areas where this technology has real benefits. One of those areas has to be high performance computing (HPC) and the quest for exascale.

The size of publicly available data in cognitive neuro-imaging has increased a lot in recent years, thanks to strong research and community efforts. Exploiting this wealth of data demands new methods to turn the heterogeneous cognitive information held in different task-fMRI studies into common-universal-cognitive models. In this paper, we pool data from large fMRI repositories to predict psychological conditions from statistical brain maps across different studies and subjects. We leverage advances in deep learning, intermediate representations and multi-task learning to learn universal interpretable low-dimensional representations of brain images, usable for predicting psychological stimuli in all input studies. The method improves decoding performance for 80% of studies, by permitting cognitive information to flow from every study to the others: it notably gives a strong performance boost when decoding studies of small size. The trained low-dimensional representation-task-optimized networks-is interpretable as a set of basis cognitive dimensions relevant to meaningful categories of cognitive stimuli. Our approach opens new ways of extracting information from brain maps, overcoming the low power of typical fMRI studies.

Learning robot tasks or controllers using deep reinforcement learning has been proven effective in simulations. Learning in simulation has several advantages. For example, one can fully control the simulated environment, including halting motions while performing computations. Another advantage when robots are involved, is that the amount of time a robot is occupied learning a task---rather than being productive---can be reduced by transferring the learned task to the real robot. Transfer learning requires some amount of fine-tuning on the real robot. For tasks which involve complex (non-linear) dynamics, the fine-tuning itself may take a substantial amount of time. In order to reduce the amount of fine-tuning we propose to learn robustified controllers in simulation. Robustified controllers are learned by exploiting the ability to change simulation parameters (both appearance and dynamics) for successive training episodes. An additional benefit for this approach is that it alleviates the precise determination of physics parameters for the simulator, which is a non-trivial task. We demonstrate our proposed approach on a real setup in which a robot aims to solve a maze puzzle, which involves complex dynamics due to static friction and potentially large accelerations. We show that the amount of fine-tuning in transfer learning for a robustified controller is substantially reduced compared to a non-robustified controller.

Human activity recognition plays an important role in people's daily life. However, it is often expensive and time-consuming to acquire sufficient labeled activity data. To solve this problem, transfer learning leverages the labeled samples from the source domain to annotate the target domain which has few or none labels. Unfortunately, when there are several source domains available, it is difficult to select the right source domains for transfer. The right source domain means that it has the most similar properties with the target domain, thus their similarity is higher, which can facilitate transfer learning. Choosing the right source domain helps the algorithm perform well and prevents the negative transfer. In this paper, we propose an effective Unsupervised Source Selection algorithm for Activity Recognition (USSAR). USSAR is able to select the most similar $K$ source domains from a list of available domains. After this, we propose an effective Transfer Neural Network to perform knowledge transfer for Activity Recognition (TNNAR). TNNAR could capture both the time and spatial relationship between activities while transferring knowledge. Experiments on three public activity recognition datasets demonstrate that: 1) The USSAR algorithm is effective in selecting the best source domains. 2) The TNNAR method can reach high accuracy when performing activity knowledge transfer.

Inductive transfer learning has greatly impacted computer vision, but existing approaches in NLP still require task-specific modifications and training from scratch. We propose Universal Language Model Fine-tuning (ULMFiT), an effective transfer learning method that can be applied to any task in NLP, and introduce techniques that are key for fine-tuning a language model. Our method significantly outperforms the state-of-the-art on six text classification tasks, reducing the error by 18- 24% on the majority of datasets.

Abstract--Autonomous Vehicles(AV) are one of the brightest promises of the future which would help cut down fatalities and improve travel time while working in harmony. Autonomous vehicles will face with challenging situations and experiences not seen before. These experiences should be converted to knowledge and help the vehicle prepare better in the future. Online Transfer Learning will help transferring prior knowledge to a new task and also keep the knowledge updated as the task evolves. This paper presents the different methods of transfer learning, online transfer learning and organic computing that could be adapted to the domain of autonomous vehicles. Autonomous Vehicles(AV) or Driver-less Cars are one of the most widely discussed emerging technology in the present day. An autonomous vehicle can be explained as a vehicle that adapts to its surroundings and can navigate itself by sensing its environment and minimal human input.

As a promising area in artificial intelligence, a new learning paradigm, called Small Sample Learning (SSL), has been attracting prominent research attention in the recent years. In this paper, we aim to present a survey to comprehensively introduce the current techniques proposed on this topic. Specifically, current SSL techniques can be mainly divided into two categories. The first category of SSL approaches can be called "concept learning", which emphasizes learning new concepts from only few related observations. The purpose is mainly to simulate human learning behaviors like recognition, generation, imagination, synthesis and analysis. The second category is called "experience learning", which usually co-exists with the large sample learning manner of conventional machine learning. This category mainly focuses on learning with insufficient samples, and can also be called small data learning in some literatures. More extensive surveys on both categories of SSL techniques are introduced and some neuroscience evidences are provided to clarify the rationality of the entire SSL regime, and the relationship with human learning process. Some discussions on the main challenges and possible future research directions along this line are also presented.

Electronic phenotyping, which is the task of ascertaining whether an individual has a medical condition of interest by analyzing their medical records, is a foundational task in clinical informatics. Increasingly, electronic phenotyping is performed via supervised learning. We investigate the effectiveness of multitask learning for phenotyping using electronic health records (EHR) data. Multitask learning aims to improve model performance on a target task by jointly learning additional auxiliary tasks, and has been used to good effect in disparate areas of machine learning. However, its utility when applied to EHR data has not been established, and prior work suggests that its benefits are inconsistent. Here we present experiments that elucidate when multitask learning with neural networks can improve performance for electronic phenotyping using EHR data relative to well-tuned single task neural networks. We find that multitask networks consistently outperform single task networks for rare phenotypes but underperform for more common phenotypes. The effect size increases as more auxiliary tasks are added.