Research Abstracts - 2007

Motivation

Animals routinely solve extremely difficult, nonparametric learning problems during development. How do they do this? Can we formalize their learning in a computational model? What insights would this model provide about learning in both biological and computational systems? The goal of this work is to develop biologically-inspired approaches to machine learning and reciprocally, to use these approaches to better understand learning in biological systems.

Approach

This work presents a new model of self-supervised machine learning. It is inspired by the notion that the sensory information gathered by animals is inherently redundant. This redundancy can enable learning without requiring explicit teaching (as in supervised learning) or statistical modeling (as in unsupervised learning). In Nature, these are frequently unavailable and yet animals learn anyway. In other words, redundancy allows animals to supervise their own learning, hence the designation of self-supervised learning. It also enables a powerful computational framework to provide this ability to machines.

Specifically, we present a new mathematical framework called cross-modal clustering [2,4] for understanding learning in both animals and machines based upon unsupervised interactions with the world. This approach draws upon evidence gathered by the brain and cognitive sciences demonstrating the extraordinary degree of interaction between sensory modalities during the course of ordinary perception.

This work provides evidence these interactions are fundamental to solving some of the difficult developmental learning problems faced by both animals and artificial systems. It further demonstrates that a biologically-inspired approach can help answer what are historically challenging computational problems. These include, for example, how to cluster non-parametric data corresponding to an unknown number of categories and how to learn motor control through observation.

Results

We have applied this framework to problems in a wide variety of domains. These include:

1. Separating unknown mixture distributions based on their co-occurrences with other mixture distributions [2,4]. The figure below shows the progression of the cross-modal clustering algorithm. (A) shows the initial randomized Voronoi partitioning of mixture models in two independent but co-occurring modalities. Our algorithm reconstructs the actual distributions by combining Voronoi regions based on co-occurrence information. Figures (B) and (C) show intermediate region formation. (D) shows the correctly clustered outputs, with the confusion region between the categories indicated by the yellow region in the center. Notably, the algorithm is entirely non-parametric and knows neither the number of clusters nor the distributions from which they are drawn.

   
   

2. Learning the vowel structure of American English &ndash the number of vowels and their formant structures &ndash simply by watching and listening to someone speak [3]. This approach is entirely non-parametric and has no a priori knowledge of the number of categories (vowels) nor the distributions of their individual presentations; it also has no prior linguistic knowledge. This work is the first example of unsupervised phonetic acquisition of which we are aware, outside of that done by human infants, who solve this problem easily.

   We plan to extend this result to cover a complete set of phonemes in English to develop a deeper understanding of protolanguage - the poorly studied prelexical states through which infants pass as they acquire word usage. In particular, we would like to construct a system that babbles in English, using the framework provided in (iii) below.

   

3. Acquiring sensorimotor control. We have constructed a system that learns to sing like a real zebra finch looking for a mate, following the developmental stages of a fledgling zebra finch [1]. It first learns the song of an adult male corresponding to its "father" and then listens to its own initially nascent attempts at mimicry through an articulatory synthesizer. By recursively reapplying the cross-modal clustering framework described in [2], the system demonstrates the acquisition of sensorimotor control through what was initially a perceptual framework. Spectrograms displaying the vocalizations of the real bird used to train this system and the resulting learned output of its artificial "son" are shown below. (The data for this experiment was provided by Ofer Tchernichovski, CCNY.)

   
   

4. Unsupervised segmentation of human functional MRI (fMRI) data. The cross-modal clustering framework offers a new way to identify functional regions in the brain without knowing of their existence in advance. This is an especially attractive approach to this problem, because we lack any detailed functional models of modular brain structure. Not having such models makes it difficult to apply standard machine learning techniques to elucidate functional regions within the brain, which are almost entirely unknown.

   The figure below presents very recent results demonstrating the automatic detection of the fusiform face area (FFA) via this approach, which only uses fMRI data and not the actual experimental inputs describing the images or their categories. (The data is this experiment was provided by N. Kanwisher and L. Reddy, Department of Brain and Cognitive Sciences, MIT, and C. Baker, NIH.)

   

   Among our future research goals is the generation of a distributed, geographic atlas of modular brain functions, using self-supervised learning techniques. We are currently locating other functional areas within human brains, such as the parahippocampal place area (PPA). We are also clustering fMRI data from rats looking for similar functional regions. Most importantly, this work demonstrates that a biologically-inspired theory of machine learning can symbiotically help us understand the very systems providing its inspiration.

Support

This work has been funded through the AFOSR under contract #F49610-03-1-0213 and through the AFRL under contract #FA8750-05-2-0274.

References:

[1] Michael Coen. Learning to sing like a bird: An architecture for self-supervised sensorimotor learning. In submission. 2007.

[2] Michael Coen. Multimodal Dynamics: Self-Supervised Learning in Perceptual and Motor Systems. Ph.D. Dissertation. Massachusetts Institute of Technology. 2006.

[3] Michael Coen. Self-Supervised Acquisition of Vowels in American English. In Proceedings of the Twentieth National Conference on Artificial Intelligence (AAAI'06). Boston, MA. 2006.

[4] Michael Coen. Cross-Modal Clustering. In Proceedings of the Twentieth National Conference on Artificial Intelligence(AAAI'05), pp. 932-937. Pittsburgh, PA. 2005.