Cross-Sectional Team Deep Learning
The Cross-Sectional Team Deep Learning conducts basic and applied research in the field of adaptive multi-layer neural network architectures that learn complex tasks from very large amount of raw, unprocessed data.
The group's special focus is on:
- Deep unsupervised learning and generative models to enable processing of unlabeled data
- Deep reinforcement learning for complex control task / control optimization
- Multi-modal learning
- Transfer learning
- Multi-task learning
- One-shot learning (learning from few examples only)
- Learning to generalize from synthetic to real-world data
- Learning to learn (meta-learning: data-driven adaptation of learning procedure and network architecture)
- Active learning (active selection of data to be processed)
Background
Recently, major breakthroughs were achieved by deep neural networks in complex learning scenarios such as natural image understanding, speech processing, and control such as autonomous car driving or playing GO. One of the main benefits is hereby the ability of the networks to learn task relevant features directly from the raw data, in contrast to the necessity to hand craft features in previous state-of-the-art machine learning approaches.
However, most of the successful state-of-the-art deep learning architectures still rely on heavily supervised learning mode where millions of hand labeled examples are necessary for the training and the data is fed into the networks in passive, carefully well-prepared fashion without the network being able to actively request and select data of interest.
Fields of Activity
Group activities are therefore aimed towards advancing methods for active, unsupervised and reinforcement learning. In these modes, networks should be able to process the available data without any labels, exploiting data internal statistics, like for instance inherent spatial, temporal or other structural relations, or different transformations inherent in the data. These can aid to drive unsupervised learning and build predictive generative models of the observed data by inferring most probable underlying latent causes.
Moreover, in the active mode, a network should be able to execute different actions that influence what data will arrive at the network's input. Generated actions can also produce positive or negative consequences following action selection.
Using this closed loop, the network gains the ability to guide learning via prediction error signals. These signals are generated when the network is making a mistake in attempt to predict an outcome - upcoming sensory events or their valencies, good or bad - that follow previous sensory events or an executed action. Prediction error signals can then be used to incrementally correct the network's internal model and continuously improve its prediction capability.
The long term aim of this research is to establish a generic artificial intelligence type of neural network architecture that fuses supervised, unsupervised and reinforcement learning, providing following capabilities:
- Transfer learning: skills learned on one task can be easily transfered on performing well on a similar, related task
- Multi-task learning: learning on multiple tasks results in a network that can recognize task structure and switch rapidly to corresponding task frame when encountered (task structure learning)
- One snapshot learning: learning a new task from a few examples
- Off-line memory reprocessing: learning in absence of the input data by internaly generated replay
- Learning to learn, meta-learning: the objective cost function and resulting learning procedure (synaptic weight and bias update rules, network architecture organization) are not fixed but are also subject to adaptation driven by the data
- Active learning: instead of being a passive receipient of input data, the network is actively looking and selecting data sources and data relevant for the task, and actively discovers relevant tasks (learning what constitutes a relevant task)
The group also provides support for implementation and deployment of deep neural networks for different scientific and technological applications.
The CST Deep Learning is jointly led by Dr. Jenia Jitsev and Prof. Morris Riedel.