Education 

---

Research and Teaching Experience 

---

Selected Publications 

---

1. Books and Book Chapters:

2. Papers in Refereed Journals and Collections:

3. Papers in Proceedings of Refereed Conferences:

4. Reviews

Significant Software 

---

Much of my career has involved the development of complex innovative software systems for both research and production environments. Except where noted, I was the chief architect of the systems and was in most cases the chief or sole developer. One of the more significant accomplishments, the FramerD infrastructure, is the basis for many of these applications and is describe in more detail towards the end of this section.

Web Applications

I have built numerous web applications on the FramerD platform, including:

The "Chopper"

The Chopper is a natural language tagging and analysis engine originally implemented in my work at the MIT Media Laboratory. It provides for high throughput tagging based on optimizing finite state machines (OFSMs). In a nutshell, a hand-coded grammar and hybrid manual/statistical lexicon specify a set of potential paths through any given sentence. The OFSM algorithm simply finds the shortest path(s) in order to analyze the sentence.

This approach also allows the integration of other kinds of ambiguity into the parsing process by adding more paths to the graph of possibilities. For example, choices about compound aggregation (whether "fruit flies" refers to the single lexicon entry "fruit fly" or the two entries "fruit" and "fly") can simply be embedded in the graph. This can also cover simple lexical variation (capitalized words) and some simple misspellings.

The first implementation of the Chopper was in Common Lisp. A subsequent high-performance implementation was done in C, using tables written by the Common Lisp implementation (where the grammar and lexicon were specified). Eventually, the Common Lisp implementation was moved to the FramerD Scheme-based scripting language.

One version of the Chopper can be used experimentally via a web service and various online forms at services.beingmeta.com.

BRICO

BRICO is a broad coverage knowledge base derived from a variety of online sources. The core of BRICO algorithmically combined the public WordNet database with various public translation dictionaries to create an interlingual knowledge base which was then audited by native speakers.

The BRICO knowledge base was subsequently extended by a variety of other public information sources, including national and international gazeteers, public web directories, and Wikipedias in various languages.

This knowledge base has been a key component in many applications and some products. The BRICOBASE (www.bricobase.net) web site provides a rich (AJAX based) browsing experience for the BRICO knowledge base, including versions optimized for Apple's iPhone and Amazon's Kindle.

Work on SteamXML

While working for Clear Methods, I worked extensively on their large Java code base which implemented their SteamXML implementation of the Water language. Among my accomplishments were:

• a foreign library interface using the Java reflection API to expose native Java objects (and the Java namespace) as Water objects;
• a generic object registration facility to maintain object identity (EQness) based on object attributes;
• a module system with autoloading on demand;
• improved JDBC integration and robustness;
• added multithreading and HTTP/1.1 support to their in-application web server;
• integrated SteamXML with the beingmeta/FramerD software stack, through both FramerD's remote execution protocol and through HTTP proxies (for access from firewalled environments).

FramerD

FramerD (www.framerd.org) is an innovative software infrastructure originally developed to support very large scale experience-based common sense reasoning and subsequently becoming the foundation of beingmeta's technology stack. It was implemented to support large scale and high performance computing as well as taking advantage of trends in hardware development (large caches, multiple CPUs/cores). FramerD's core implementation is in portable C and alternative implementions have been implemented in Java, Common Lisp, and Python. The FramerD implementation "layer cake" includes innovations at data, database, inference, scripting, and web application levels.

data level

The FramerD data level consists of standard LISP/Scheme data types with several innovations:

• a low-level typing system which is user/module-extensible for both consed and immediate data structures;
• fine-grained locking to support multi-threaded execution by a combination of structure-based locks and "hash locking" for pointers;
• lightweight (non-consed) native object identifiers (OIDs) to improve processing locality while supporting large object bases;
• very fast native implementation of set objects (called choices) supporting log or linear time operations and amortized mutation;
• ubiquitous unicode to support the canonical representation of multi-lingual and inter-lingual data;
• a distributed execution framework providing for multi-node networked applications.

database

The FramerD database interface includes a simple schema-free object database mapping unique OIDs to arbitrary values (commonly tables). This is combined with an indexing facility based on generic tables mapping key/value pairs to sets of OIDs. The implementation includes:

• network distributed object stores and indices;
• pervasive caching to improve performance and reduce latency;
• provision for block fetches to minimize network latency and optimize disk access;
• seamless specification of virtual objects combining data from multiple sources;
• log-based file structures to improve write performance.

inference

FramerD consists of a frame-based inference facility built on top of the database capabilities. Some innovations include:

• use of the native set implementation (choicesets) to obviate the implementation-level distinction between fields and relations;
• generic slot based facilities which easily support both conventional (OWL) and novel (case-based, resource-limited) inference patterns;
• a dependency maintenance system to ensure the consistency of the database's pervasive caches, extensible to user-specified caches;
• selective thread-local caching of inferences for increased performance, improved memory usage, and greater thread independence;
• iterated partial evaluation: a (patent pending) technique for reducing the effects of network (or device) latency on procedural queries with complex internal dependencies; this can provide average speedups ranging from 100% to over 1000% (based on network latencies) across a broad class of queries.

scripting

FramerD contains a scripting language based on Scheme that offers several useful and unusual features:

• native non-determinism simplifies the expression of procedural queries and many search operations, enabling &lquo;Prolog-style programming&rquo; in an applicative language;
• multi-threading facilities which can take effective advantage of multiple CPUs or cores;
• a simple, effective module system to support programming in the large;
• a formatted output facility (based loosely on Interlisp's PRINTOUT) which simplifies the programmatic generation of reports or documents;
• a generic caching facility taking advantage of the inference layer's dependency maintenance;
• a rich array of text analysis and processing tools based on structured patterns and rules based on LISP (Scheme) objects; this also includes functions for parsing varieties of data file formats and for phonetic (SOUNDEX, metaphone) normalization and simple morphological analysis (with rules implemented for English, French, Spanish, Italian, Portugese, and Finnish).

web applications:

FramerD supports web applications in numerous ways:

• mod_fdserv is a custom Apache module which allows Apache to launch and communicate with persistent fdserv processes which use FramerD's adeptness at multi-threading to provide high-throughput low-latency responses to document and CGI requests;
• XML and HTML generation libraries where the nesting and calling hierarchy mirrors the natural hierarchical structure of XML and HTML documents;
• native JSON parsing and output to simplify the implementation of rich AJAX web applications;
• primitives allowing access to and provision of Web services based on REST or SOAP protocols;
• FDXML, an executable variant of XML/HTML based on FramerD, which embeds computation in HTML documents through either direct scripting in Scheme, specification of programs as XML applications, or the implementation of new XML/HTML tags with underlying implementations in Scheme (or other languages).

Other Cool Stuff

Other interesting accomplishments highlight some more of my skills and interests:

Microniver provided a low-level library for continuation-based or parallel processing (implemented in Z-80 assembler).

Cauldrons implemented part of Minsky's "Society of Mind" theory through a multi-context assertion-based reasoner that provided the assertional equivalent of "procedure calls" and inferential "memoizing."

An animation system providing non-programmer tools for specifying graphical animations; included the coordination of animated characters with musical (MIDI) event streams.

The Dungeon Kit provided an environment for creating online interactive games with responsive characters and visual and audio effects. It was implemented in QLOGO, an object-oriented version of the Logo language.

Various representation language languages starting with ARLO (my Bachelor's thesis in philosophy), that provided flexible inference and storage frameworks for a range of applications.

The information waldo provided a direct manipulation environment for computer data and knowledge bases using a physical device thatx I designed to translate intuitive hand gestures into mouse events.

TYPICAL (my M.S. thesis program) was a type inference system embedded in Scheme which was used to represent mathematical concepts.

CYRANO (my Ph.D. thesis program) was a mathematical discovery program inspired by Lenat's AM and Eurisko programs (their vision, accomplishments, and shortcomings) but based partially on Ray Solomonoff's algorithmic complexity theory (which was in turn based on the work Gregory Chaitin and others (and eventually, Claude Shannon).

The abrasives advisor was an expert system for the custom engineering of abrasive products, codifying the expert knowledge of a key sales engineer at a large manufacturing firm. This was implemented in Common LISP for DOS-based PCs.

The Fed Funds advisor was a tool for traders at a major bank's Money Market desk, using simple calendar-based heuristics to suggest possible factors influencing market fluctuations. This system was the basis of a Harvard Business School case study (it reportedly paid back many times its development costs). This was implemented in TI Scheme for DOS-based PCs.

Reimplementations of classic AI programs (planners, parsers, story understanders, etc) written for various AI programming classes taught at the MIT Media Lab for non-engineers. All written in Common Lisp.

An analogical window system which combined a novel prototype-based object system (Framer) with analogical algorithms to automatically create visual renderings of new data based on past renderings of similar information. This work was joint with the Visual Language Workshop at MIT and linked with the BadWindows environment. Written in C.

A poetry “understander” which implemented a model of aesthetic comprehension which treated poems as analogy teaching machines whose formal structure (rhyme, meter, layout) cued semantic analogies. Written in C and the FramerD scripting language.

Semantic Clay was originally conceived as an art piece at UC Irvine. It was a web-based Java applet which used the BRICO knowledge base and allowed viewers to explore a space of words and meanings by exposing related terms (including across languages) as individual terms were selected or emphasized by user interaction.

Personal Information 

---