Steve R. Howell, Ph.D.
Updated January 21st , 2009
Download the SRNEngine neural network simulator for
Windows here!
(Use of SRNEngine is free for research or educational purposes only!)
Download the Sensorimotor Feature
Vectors from:
Howell, S. R., Jankowicz, D., & Becker, S. (2005). A Model of Grounded Language Acquisition: Sensorimotor Features Improve Lexical and Grammatical Learning, Journal of Memory and Language, 53(2), 258-276.
Please notify me at howell@kutztown.edu and cite this paper
if you use these features for anything!
Kutztown University
I am presently an assistant professor of psychology at Kutztown University of Pennsylvania, in Kutztown, Pennsylvania, USA. In 2007-2008 I was at KU as a full-time contract faculty member, and starting Fall 2008 I became a tenure-track faculty member. I teach primarily statistics and research design, both undergraduate and graduate, plus cognitive psychology, child psychology, or the psychology of language depending on the term. I run a research lab of 10 undergraduate and graduate research assistants, research students, and programmers, divided among four projects:
Inference of Syntax from Semantics: Continuation of my work on the interplay between syntax and semantics in human language.
Life Legacies: A project devoted to developing a structured interview technique designed to capture and store in perpetuity the personality and wisdom of individuals, especially the elderly. This is by far the largest project my lab is working on, and is now grant-funded thanks to the KU Research Committee.
Serial Killer Profiling: We are examining archival data sets to see if we can improve on the classification system currently used to classify serial offenders based on features of the perpetrator, victim, crime, and crime scene. We intend to use neural network analysis on the data sets.
Consciousness: We are in the early stages of developing a research program addressing the “hard problem” of consciousness, from the perspective of language skill and distributed sensorimotor memory representation.
First Teaching Position
I was a Visiting Assistant Professor
at Albion College in Albion, Michigan, USA
from Fall 2005 to December 2006. I
taught Research Design and Analysis I and II, Introductory Psychology, and the
Psychology of Language.
My research program at Albion
included cognitive experiments with human participants as well as continuing
connectionist modelling of language acquisition. I began an experiment using undergraduate
participants which investigated the human ability to infer meaning from context. This experiment was intended to test the
predictions of previous modelling work (Howell, Becker, and Jankowicz, 2005).
Postdoctoral Work
From August 2004 to July 2005 I was
a Research Associate at the University of
Wisconsin –Madison, working in the lab of Dr. Mark Seidenberg and Dr.
Maryellen MacDonald. While there I
developed an extended model of English past-tense production (based on Joanisse
and Seidenberg, 1999). I also developed
a model of the inference of noun class (mass/count/proper) from
determiners. Manuscripts on both
projects are in preparation.
Dissertation Research
From 1998 to 2004 I was a graduate student in the psychology
department at McMaster university (
My work consisted of designing neural networks to simulate the processes of child language acquisition, both lexicon and ‘grammar’. I'm actually more interested in mature language processing. However, after a hard look at the failures of artificial intelligence research efforts over the past decades, many of them aimed at understanding language, I have concluded that any successful model of language processing must be a developmental model. That is, we can't hope to model the complexity of adult language; we have to start with the simplest children's language and proto-language and build up successively from there. The constraints of earlier learning will direct later learning, limiting the "space" that must be "searched" for the "solution" to the language processing problem.
Of course, even the earliest child language (from the first words in the last part of the first year, up to the burst of grammar at 28 months or so) is preceded by yet another critical aspect: Concept learning. Children must have a concept for something before they can attach a word to it (although, arguably, later more complex words may in fact be learned in an interactive process of learning the word at the same time as learning about the concept itself). So, before one can model child language, one must model child concepts. Of course, language itself has a precurser, phonetics and phoneme learning. Must a developmental model of language acquisition account for all child learning back to the womb?
Luckily, it does not appear so. Some abstracted, high level representation of concepts does seem to be the minimum necessary to provide for a grounding of the meaning of words in real-world, sensory-and-motor afforded objects and actions. And likewise, language is inherently intertwined with phonemes and phonemic processing, and so a good model of language processing should include at least phonemic representations of words. However, things such as phonemic coarticulation, acoustical learning, etc. can, I believe, be safely excluded form our analysis at this level.
Thus, models of child language acquisition include as givens a representation of basic concepts in terms of their sensorimotor features, and a representation of words in terms of phonemes. We conceptualize sensorimotor features as being the most basic kind of semantics, an enumeration of features of objects that the youngest child would quickly gain based on the evidence of their senses, such as "dog is furry; dog has four legs". These prepositional pieces of sensory knowledge combined yield a long vector of featural representation for each concept. It would be nice to be able to see inside children's brains at the images they have of the world, and create our sensorimotor representations based on that. In practice, we create these representations from adults' ratings of concepts on a variety of features that would be accessible to children, such as "size" or "texture". We currently have nearly a 500 word vocabulary which is grounded in such sensorimotor features. We have shown our featural representations of words to be successful at clustering our basic vocabulary in ways that are very consistent with overall adult semantics, as well as agreeing very well with independent categorizations of the words.
The second representation of words is their linguistic label, which for us is a phonemic one. We have developed a system of phonemic coding that can represent nearly any word as a 140 bit vector, maintaining information about phoneme order, start and end effects, etc. Representing words at the input to our models in this fashion gives us great scalability, which is necessary if our models are to have a large enough vocabulary that we can study the emergence of grammar, which happens at lexicon sizes of 300-600 words. Our neural network architecture is an extension of the classic simple recurrent network, allowing for more complex inputs and outputs, and different levels of context learning.
We have recently been analyzing models with a larger vocabulary, trained on a real-world corpus drawn from transcripts of child-directed speech. We are investigating: A) the rate of lexical learning, and whether the "age" of acquisition for words is similar from the model to the child data, and B) whether the acquisition of grammatical relationships is tied firmly to increasing vocabulary size, as many have suggested. We also continue to work on demonstrating the "Propagation of Grounding" effect for instilling grounded, embodied, meaning into high dimensional meaning spaces.
RESEARCH HISTORY
From January to June 2000 I was a visiting student at the University of California at Berkeley,
working with Dr. Jerome
Feldman and his students. Their group, the Neural Theory of Language (NTL)
Project (which is based at the International
Computer Science Institute in
Thanks to the NTL group I have incorporated the work of George Lakoff on language, metaphor, and categorization, and am seeking to integrate it within the PDP (non-structured) connectionist approach. Which is not to say that the NTL people haven't persuaded me that some structure is important even to PDP nets; Terry Regier's book "The Human Sematic Potential" is incredibly powerful in that regard. I just think that the PDP unstructured neural nets, through their very operation, provide the mechanism that is necessary to underlie the categorization effects that are so important in Lakoff's work and other aspects of language, whether that be metaphor, constructions, or lexical semantic categorization. In fact, I'm becoming convinced that in studying language, categorization as is done automatically in PDP nets might be one of the most important mechanisms (See Rogers and McClelland, 2004) for motivation on this concept.
I am also working on a better understanding of such aspects of linguistics as construction grammar, aspect, and embodiment. These are currently being incorporated in a structured way into the NTL's system-building exercise, and I hope to incorporate them into an equivalent PDP approach that is mostly learned instead of structured in advance.
I've been partially focusing for the last several years on the completion of
the SRNEngine, a simulation platform for extended Elman-style nets that allows
multiple layers of all kinds, varying hysteresis values, and simulated
annealing where useful. The software is in a useable state and will enter
public beta soon. The distributed computing screensaver version is now
available for a limited beta, and will also go into public beta soon. Follow
the link at the top of the page to the McMaster Language Modelling Project to
download the screensaver version and participate in running simulations for our
lab on your spare processor cycles! Apart from language modelling, this engine
has also been used to perform time-series prediction on corporate sales data,
allowing forecasting of future sales figures, by a software company in
In 2001, I began running larger simulations that test some of the relationships between human learning of vocabulary and grammar, or the lexicon/syntax split. In human children, there is a high correlation between early lexical size and later grammatical proficiency, and I’ve built a model that demonstrates this same correlation. The paper was published in the proceedings of the 2001 conference of the Cognitive Science Society (See below). Statistical tests over many runs of the network indicate that grammatical performance is significantly improved though the use of a semantic feature input layer in addition to the usual lexical prediction. I discuss this aspect of language acquisition, as well as the symbol grounding problem and how a developmental approach to language acquisition can get around it using our methods, in a second paper, published in the proceedings of the Developmental and Embodied Cognition Workshop 2001 (See below).
In Fall of 2001, I begin doing feedforward neural net categorization using backpropagation for various emotional tasks with various collaborators. We were able to categorize infant’s experience of emotion based on aggregate EEG data, to a moderate extent. Also, we were able to categorize the emotion in speech directed to adults or infants, although we found no significant difference between child-directed speech, and emotion-laden adult-directed speech. Both projects were presented as posters at the ISIS 2002 conference (see below). Work is continuing intermittently; we would like to build an on-line system that could identify infant emotion while they are attached to the EEG machine. It could be an extremely valuable tool to infancy researchers.
I have explored Genetic Algorithms as a method for training the neural network brains of agents. I have successfully replicated the first half of Nolfi, Elman, and Paresi (1994), evolving creatures (Gulpers?) that learn to navigate a gridworld and eat food. Further development involving coevolution, including competition for resources, avoidance of predators, and the like was interesting, but much less successful.
As part of a graduate course, I developed a neural network model of melody composition, which learned some semblance of musical key and generated novel melodies that were in accordance with Narmour's Implication-Realization (IR) model of melodic perception, if not particularly beautiful.
PAPERS, POSTERS,
and PRESENTATIONS (pdf format)
Publications
- Howell, S. R., Jankowicz, D., & Becker, S. (2005). A Model of Grounded Language Acquisition: Sensorimotor Features Improve Lexical and Grammatical Learning, Journal of Memory and Language, 53(2), 258-276.
Forthcoming
- Howell, S.R., and Angrave, L. (In preparation). The importance of neurotrophic factors in learning high-order functions in hebbian cortex.
- Howell, S. R., Joanisse, M., and Seidenberg, M.S. (In preparation). Impairments in verb morphology after brain injury: A connectionist model using distributed semantics.
- Howell, S. R., MacDonald, M. M., and Seidenberg, M. S. (In preparation). Learning noun syntactic classes from semantic representations
- Howell, S.R. & Becker, S. (Under Review). Grounding words in meaning indirectly: A connectionist model of the propagation of grounding. Cognitive Science
- Howell,
S. R., & Becker, S. (Under Review). Grammar from the Lexicon: Evidence
from Neural Network Simulations of Language Acquisition. Language and Speech
· Howell, S. R. & Becker, S. (In Revision). SRNEngine: A Windows-based neural network simulation tool for the non-programmer. Behavior Research Methods, Instruments, and Computers
· Jankowicz, D., Becker, S., and Howell, S. R. (Under Review), Modeling Semantic Category-Specific Deficits Using Topographic, Corpus-Derived Representations. Journal of Experimental Psychology: General
Conference
Papers
- Jankowicz,
D., Becker, S. & Howell, S. R. (2003). Modelling Category-Specific Deficits using Topographic,
Corpus-Derived Representations. Poster presented at the 44th
annual meeting of the Psychonomic Society,
- Jankowicz,
D., Becker, S., & Howell, S. R. (2003). A model of sensorimotor
grounding in lexical learning. Presented
at the Behavior, Brain, and Cognitive Sciences Conference,
- Howell,
S. R., Becker, S., & Jankowicz, D. (2003). A model of sensorimotor
grounding in lexical learning. Presented
at the Behavior, Brain, and Cognitive Sciences Conference,
- Jankowicz,
D., Becker, S., & Howell, S. (2003). Modeling category-specific
deficits using topographic, corpus-derived representations. Presented at the Conference of the
Southern Ontario Neuropsychology Group,
- Howell,
S. R., Schmidt, L., Trainor, L., Santesso, D. L. (2002). Neural Network
Categorization of EEG Measurements of Infant Emotional States. Proceedings
of the International Society for Infant Studies Conference, 2002
- Howell,
S. R., Schmidt, L. (2002). Neural Network Categorization of Emotion in
Infant Directed and Adult Directed Speech.
Proceedings of the
International Society for Infant Studies Conference, 2002
- Howell,
S. R., Becker, S., & Jankowicz, D. (2001). Modelling
language acquisition: Lexical grounding through perceptual features, Proceedings of the Developmental and
Embodied Cognition Conference, 2001
- Howell,
S. R. & Becker, S. (2001). Modelling
language acquisition: Grammar from the lexicon? Proceedings
of the Cognitive Science Society, 2001.
- Howell,
S. R. & Becker, S. (2000). Modelling language
acquisition at multiple temporal scales. Proceedings
of the Cognitive Science Society, 2000.
- Howell,
S. R. (1999) Gender Differences in Relationship Conflict: The Role of
Self-Construal. Presented at the Annual Conference of
the Canadian Psychological Association,
- McKinnon-Doan,
H. & Howell, S. R. (1998). Antecedents
to Parenthood - A. Presented at the Canadian Psychological
Association National Conference,
- McKinnon-Doan,
H. & Howell, S. R. (1998). Antecedents
to Parenthood - B. Presented at the Canadian Psychological
Association National Conference,
- McKinnon-Doan,
H., Zimmerman, A. & Howell, S. R. (1997). Antecedents to Parenthood - Preliminary Results. Presented
at the Canadian Psychological Association National Conference,
Invited Talks
- June 2001. Invited Speaker, Symposium on Language and Memory, Canadian Society for Brain and Behavioral Sciences Conference 2001. Symposium Organizer: Ken McCrae.
Dissertation
Title: Sensorimotor Representations of Meaning in early
language acquisition
Committee: Dr. Suzanna Becker, Dr. Lee Brooks, and Dr. B. A. Levy
Other Papers
I wrote an integration paper (as a comprehensive module) while at
While at
I wrote a review paper (as a comprehensive module) on the outlook for
Computer-Assisted Psychotherapy for Dr. Rob Muller at
I wrote an integration paper (as a comprehensive module) entitled "Perspectives on the Emergence of Language" for Dr. B. A. Levy of McMaster. It is available here in pdf format.
A class paper "Connectionism and Philosophy: Why Aristotle was a connectionist." resulted in this article (pdf format), which I plan to revise soon and submit for publication. It is an application of connectionist heuristics to classic problems in moral philosophy.
A departmental seminar in Microsoft Powerpoint format that I gave in October, 1999 is available as well, which includes work that I did applying SRN's to language learning with letter-triple input representations (a more scalable approach, as well as possibly an easier-learned one than localist words).
I spent a lot of effort creating a very colorful, interesting slideshow for the first ever presentation of my plan of research. It is available here as a Microsoft Powerpoint presentation, although most of it is outdated.
My undergraduate thesis, supervised by Dr. David Reid of York University was entitled "The
Relation of Self-construal Typology to Conflict Resolution in Committed
Relationships". It was presented at the Canadian Psychological Association
1999 conference in
While I was an undergraduate I worked with Dr. Helen Doan of York University on child developmental research, specifically issues of maternal-fetal and paternal-fetal attachment. I presented three co-authored conference presentations (CPA 97; CPA 98;CPA98a) on our results.
Yes, I realize that my current research is a far cry from my undergraduate
research, but that was only the completion of my undergrad career. I starting
in 1989 at the University of Waterloo in
Systems Design Engineering, and completed two years of that program. After that
I worked as an application developer and manager in the computer industry for
several years until entering
Athletics
Since returning to academics I also found my way into athletics. I am a past
member of the McMaster
University Fencing Team - Men's Foil. While in
While at UW-Madison, I fenced and coached with the UW-Madison fencing club
and the
I am now the head coach and Faculty Advisor to the new Kutztown University Fencing Club. We offer recreational training and competition to the university community, and competitive training for club members.
I occasionally practice Karate (Shotokan or Isshinryu) as well. I like to play basketball also, and played sort of weekly while at UW-Madison with the psychology weekend team. I’ve not yet succeeded in organizing the Kutztown Psychology Faculty in to a team yet.
In good weather, I like to bicycle or run outdoors.
Personal Time
As to my scarce leisure time, well, I engage in the socially questionable J hobby of writing action/adventure and intrigue science fiction and fantasy. I have three projects currently underway that never seem to get finished. I have also written several as-yet-unpublished short stories, although one was almost published, before being cut from the anthology that it had been solicited for. L
I'd probably get more writing done, too, if it wasn't for my fondness for computer games.
I love chess, too, and play whenever I can, sometimes on Yahoo.
When I need to get away from the keyboard and unwind with something slower-paced, I work at carpentry. Aside from minor home repairs, I have crafted a Shaker-style maple coffee table, a built-in bookcase wall and fireplace mantle, a freestanding clapboard garden shed, a breakfast bar, and a large built-in bookcase and office-in-a-closet in Michigan. I’m presently learning the fascinating art of steam-bending wood, and have nearly finished a traditional style skin-on-frame rowboat. I’m looking forward to putting the skin on the boat and launching it this summer in Beltzville Lake in Pennsylvania near where I live.
Oh, and when I don't feel like doing anything else listed above, I struggle at learning vocal and guitar music.
Comments or questions? Please e-mail me!