Daffodil International University
Faculties and Departments => Teaching & Research Forum => Topic started by: rashidacse on November 28, 2015, 12:10:46 PM
-
Quoting "Structured Machine Learning: Ten Problems for the Next Ten Years" by Pedro Domingos (Professor at University of Washington)
These seem to be top 10 problems in machine learning research.
Statistical Predicate Invention
Predicate invention in ILP and hidden variable discovery in statistical learning are really two faces of the same problem. Researchers in both communities generally agree that this is a key (if not the key) problem for machine learning. Without predicate invention, learning will always be shallow.
Generalizing Across Domains
Machine learning has traditionally been defined as generalizing across tasks from the same domain, and in the last few decades we’ve learned to do this quite successfully. However, the glaring difference between machine learners and people is that people can generalize across domains with great ease.
Learning Many Levels of Structure
So far, in statistical relational learning (SRL) we have developed algorithms for learning from structured inputs and structured outputs, but not for learning structured internal representations. In both ILP and statistical learning, models typically have only two levels of structure. For example, in support vector machines the two levels are the kernel and the linear combination, and in ILP the two levels are the clauses and their conjunction. While two levels are in principle sufficient to represent any function of interest, they are an extremely inefficient way to represent most functions.
Deep Combination of Learning and Inference
Inference is crucial in structured learning, but research on the two has been largely separate to date. This has led to a paradoxical state of affairs where we spend a lot of data and CPU time learning powerful models, but then we have to do approximate inference over them, losing some (possibly much) of that power. Learners need biases and inference needs to be efficient, so efficient inference should be the bias. We should design our learners from scratch to learn the most powerful models they can, subject to the constraint that inference over them should always be efficient (ideally realtime).
Learning to Map between Representations
Three major problems in this area are entity resolution (matching objects), schema matching (matching predicates) and ontology alignment (matching concepts).
Learning in the Large
Structured learning is most likely to pay off in large domains, because in small ones it is often not to difficult to hand-engineer a “good enough” set of propositional features. So far, for the most part, we have worked on micro-problems (e.g., identifying promoter regions in DNA); our focus should shift increasingly to macro-problems (e.g., modeling the entire metabolic network in a cell).
Structured Prediction with Intractable Inference
Max-margin training of structured models like HMMs and PCFGs has become popular in recent years. One of its attractive features is that, when inference is tractable, learning is also tractable. This contrasts with maximum likelihood and Bayesian methods, which remain intractable. However, most interesting AI problems involve intractable inference. How do we optimize margins when inference is approximate? How does approximate inference interact with the optimizer? Can we adapt current optimization algorithms to make them robust with respect to inference errors, or do we need to develop new ones? We need to answer these questions if max-margin methods are to break out of the narrow range of structures they can currently handle effectively.
Reinforcement Learning with Structured Time
The Markov assumption is good for controlling the complexity of sequential decision problems, but it is also a straitjacket. In the real world systems have memory, some interactions are fast and some are slow, and long uneventful periods alternate with bursts of activity. We need to learn at multiple time scales simultaneously, and with a rich structure of events and durations. This is more complex, but it may also help make reinforcement learning more efficient. At coarse scales, rewards are almost instantaneous, and RL is easy. At finer scales, rewards are distant, but by propagating rewards across scales we may be able to greatly speed up learning.
Expanding SRL to Statistical Relational AI
We should reach out to other subfields of AI, because they have the same problems we do: they have logical and statistical approaches, each solves only a part of the problem, and what is really needed is a combination of the two. We want to apply learning to larger and larger pieces of a complete AI system. For example, natural language processing involves a large number of subtasks (parsing, coreference resolution, word sense disambiguation, semantic role labeling, etc.).
Learning to Debug Programs
One area that seems ripe for progress is automated debugging. Debugging is extremely time-consuming, and was one of the original applications of ILP. However, in the early days there was no data for learning to debug, and learners could not get very far. Today we have the Internet and huge repositories of open-source code. Even better, we can leverage mass collaboration. Every time a programmer fixes a bug, we potentially have a piece of training data. If programmers let us automatically record their edits, debugging traces, compiler messages, etc., and send them to a central repository, we will soon have a large corpus of bugs and bug fixes