Intel Collaborative Research Institute for Computational Intelligence
Machine learning with a 2020 sight
Group of machine learning experts from the Hebrew University and the Technion, led by Prof. Shai Shalev Shwartz (HUJI) and Prof. Shie Mannor (Technion) will develop advanced learning algorithms with the goal of meeting the requirements of future applications, in particular intelligent agents and distributed learning. The research team has identified and will explore five fundamental technology gaps, which are believed essential to the next generation of machine learning applications, and can benefit and enhance novel computer platforms and architectures:
Distributed Learning.Most current learning algorithms involve a centralized agent, which collects data and uses it to deduce prediction rules (“the intelligent behavior”). However, this approach is quickly becoming irrelevant, due to the exponential growth in both the available data (e.g., images, texts, sensor data) and computational resources (data centers, clouds, millions of available cores, smart-phones). The purpose of this project is to build novel frameworks for a “decentralized intelligence”. The researchers will define a distributed learning model and will develop novel learning algorithms to tackle the unique features and constraints of decentralized learning, such as lack of centralized authority, resiliency to failures of components, dynamic network topology, asynchrony, power restrictions, communication limitations, privacy, etc.
Light supervision. In many real world applications we are constrained with the amount of information we can receive for each individual training example. The lack of full information on each individual example can be due to various reasons such as partial supervision, latent variables, inherent noise, privacy protected data, or from constraints on the type of interaction with the data source. For example, suppose that we would like an intelligent smart phone to learn to classify our phone calls, say according to work related vs. personal. While the phone can easily record all our calls, it does not receive the “labeling” of which calls are work related and which calls are personal. It may receive labels of few examples, or even may ask the user to label specific calls.
Online Learning. In static learning problems, we have a clear distinction between a learning phase, in which the learning algorithm trains the model based on a batch of training examples, and a prediction phase in which the learned model is applied on new examples. However, in many real world problems, the learner must continue to learn and improve all the time. This is especially true in dynamic environment. For example, in security applications the environment always tries to find weaknessed in the current model. In addition, in next generation applications of machine learning, new examples always keep coming, and it may even be impossible to store all the examples.
Learning sequences.In well structured learning problems, there is a simple way to represent objects as vectors. For example, an image can be represented as a matrix of pixels or as a bag of image descriptors. Things are more involved when we like to classify are sequenced objects, like video, audio, text streams, or Internet traffic. Sequences are common to all modalities of sensory data. New methods and learning models are required for effectively dealing with structure and prediction of sequential data streams.
Process and sensing-acting learning. The most challenging learning scenario, on which there is much progress in recent years, is the integration of continuous multi sensory signals, such as audio and video, into sequences of decisions, actions and long term planning, in a continuously changing stochastic environment. This type of learning setting dominates applications like robotics and autonomous intelligent agents, but is most crucial for natural human machine interaction. We will develop novel paradigms that combine information and control theories with online and semi-supervised learning algorithms, to tackle the very challenging setting of decision making and planning in partially observed environments.
The main goal of this proposal is to tackle aspects of the aforementioned technology gaps by providing novel algorithms for learning in the scenarios mentioned above. For all these central machine learning problems we will investigate special accelerators that can enhance their performance.
The main outcomes of this research are formal models, algorithms and software libraries. Ideally, one would like to build a proof of concept by applying these technologies into real-world application. However, this is conditioned on the availability of real–life data.
During the first year we will develop and analyze a first round of algorithms for the aforementioned projects. In particular, we will build a learning framework for distributed learning (analogous to PAC learning), we will develop efficient algorithms for active learning, we will develop algorithms for online learning with smart experts, for finding meaningful patterns in sequences, and for long term planning. In subsequent years we will implement and experiment with the algorithms, underscore possible practical deficiencies, and will develop improved versions of the algorithms.
The team has recently developed a first algorithm for active learning (available on arxiv, http://arxiv.org/abs/1208.3561 ) and is in the process of developing an algorithm for distributed primal-dual optimization.