Darpa Perspective on AI

DARPA may be looking beyond the current Neural Network focus on perception and manipulation but the rest of the field seems to be stuck on NNs. Good-Old-Fashioned-AI (GOFAI) techniques, such as knowledge representation (e.g. SCONE (http://www.aaai.org/ocs/index.php/FSS/FSS11/paper/viewFile/4...) or Subsumption Based systems (http://dl.kr.org/oldproceedings/AIMag11-02-003.pdf), or Rule-Based Programming (http://mprc.pku.cn/mentors/training/ISCAreading/1986/p28-gup...), the whole large field (http://dai.fmph.uniba.sk/~sefranek/kri/handbook/handbook_of_...) are simply ignored these days. (disclaimer, I was involved with several of these systems).

Suppose your problem is to replace a tire on a car using a wrench. Perception is involved in finding the wrench. Manipulation is involved in recognizing torque. NNs are excellent for both of these tasks. In fact, it would seem ideal to bind both kinds of NNs to the concept "WRENCH". They provide grounding for the concept.

But knowing how to find a wrench or how to use one is different from knowing WHY you want to do either. This is where DARPA has lost the thread. In my opinion, the future belongs to self-modifying systems that can do knowledge representation and planning using NNs as the I/O subsystems. Self-modification implies that the system gradually evolves as it acts in the world. The mistakes it made yesterday will change the system to avoid those mistakes. GOFAI techniques combined with self-modification fundamentally changes the game.

An interesting side-effect is that, given two identical systems, they will eventually diverge. One will "know" that a table has 4 legs. The second will know that a table has rows and columns. Through self-modification their knowledge nets will gradually diverge until they are nowhere the same. The side-effect is that you can no longer "learn by copying" but have to "learn by teaching".

Direct links

PDF slides: http://www.darpa.mil/attachments/AIFull.pdf

Video: https://www.youtube.com/watch?v=-O01G3tSYpU