DARPA launches the PAL program and SRI begins CALO

Background

By the early 2000s, software that helped people keep track of messages, appointments and tasks existed, but such digital assistants relied on models that designers had written out explicitly, so their behavior was fixed in advance ^[2]. The Defense Advanced Research Projects Agency (DARPA) took a different view of what an assistant should be, and in 2003 launched the Personalized Assistant that Learns, or PAL, program ^[3].

What happened

SRI International began work on its own contribution to PAL in 2003. That project was called CALO — the Cognitive Assistant that Learns and Organizes — and SRI describes it as its component of the larger DARPA effort ^[1][2].

The scale was unusual for a software research effort. SRI describes PAL as a five-year, 150 million dollar project that brought together more than 300 researchers from 22 leading research institutions ^[1]. AI Magazine, writing about integrated AI systems, describes CALO itself as a significant development effort involving more than 400 researchers from more than 25 universities ^[3]. Either way, the work was distributed across many labs rather than concentrated in one place.

What it was meant to do

AI Magazine describes PAL's main goal as integrating machine learning, natural language processing and human-computer interaction so that the resulting cognitive assistants could emulate an executive assistant: managing email, handling calendar appointments and preparing for meetings ^[3].

SRI frames the ambition more broadly. The aim was a new generation of cognitive assistants that could reason, learn from experience, execute commands, explain their own actions, reflect on their experience, and respond robustly to surprise ^[1]. Explaining actions and reflecting on experience are the parts that go beyond ordinary automation. A system that can say why it filed a message in one folder rather than another, or notice that a pattern of requests has changed, is doing something different from a program following a fixed script.

The clearest break with earlier assistants lay in where the knowledge came from. Rather than having its models explicitly programmed, CALO was rooted in learning in the wild — learning from the material generated as the system was actually used, instead of from rules written down by its designers ahead of time ^[2]. In practice this means the assistant's behavior is derived from observing what its user does with mail, schedules and documents, and adjusting accordingly.

What came out of it

A multi-year program of that size produced research as well as software. According to AI Magazine, DARPA-funded publications and code contributed to a set of techniques that became standard equipment in machine learning and knowledge representation: statistical relational learning, Markov logic networks, probabilistic relational models, conditional random fields, latent Dirichlet allocation and Bayesian inference ^[3]. These are different tools with different jobs — conditional random fields, for instance, are used for labeling sequences such as text, while latent Dirichlet allocation is a way of discovering topics in a document collection — but they share a concern with reasoning under uncertainty, which is exactly what an assistant working from noisy, partial observations of a person's working life has to do.

SRI describes CALO as one of the most significant AI projects of the early 2000s ^[1]. Beyond any single technique, the program is remembered as an early attempt to make an AI system's competence come from data and experience rather than from hand-written rules, and to do so at a scale that required many institutions to coordinate their work.