A Technical Perspective on Cognitive Architectures

Transcription

1 A Technical Perspective on Cognitive Architectures March 14, 2015 Guna Seetharaman Ph.D., FIEEE Information Intelligence and Analysis Division Information Directorate, Rome, NY Integrity Service Excellence Distribution C: 1

2 CONTEXT The ACM and IBM Joint Workshop on Cognitive Architectures Thanks to Pradip and IBM colleagues Thanks to audience for accepting absent presentation Apology for not attending Standard disclaimers 2

3 AFRL Mission Leading the discovery, development, and integration of affordable warfighting technologies for our air, space and cyberspace force. Assured performance, scalable, affordable, sustainable, scalable-on-demand It s not just about the science it s about leadership in S&T 3

4 4

5 5

6 CETE Program: Core Projects C4ISR Enterprise at The Edge Processing & Exploitation Focused Effort Core set of Problems: Focused thrust on processing and exploitation. Detection and tracking of dismounts, proactive & request driven cueing between platforms, reasoning from visual and multimedia information, content-driven analysis indexing and accessing data, and performance focused algorithms. Projects: We are focused on four specific efforts at the core, and will foster active synergy with CETE funded efforts across the DoD Labs and DoD sector, plus Universities. Robust tracking, and 3D mapping across wide area imagery the intermittent loss of visibility and fidelity.. Video based reasoning and deriving who, what and when from theatre derived imagery Content analysis for automated indexing of images with ISR oriented goals and assistive retrievals Mapping of key algorithms on to sensor platforms: geolocation/situation awareness; performance, energy and latency tradeoffs to enable endurances Real-time plus metrics: feed-forward, versatility, comprehensive/contextually compartmentalized, distributed KB representations ingest time vs batch-mode data structures ie episodic learning vs epochal learning included. CORE TOPICS Transform the FORENSIC capacities to TACTICAL advantage through onboard computing. Mitigate the risk of communication. Stay ahead of the adversaries in game-changes Deal with the volume and velocity of data. Level 3 6

7 State of The Art at high-trl Admittedly lots of progress in Architecture Commodity Computation wise Functional know-how Lack of Systems Engineering and Interoperability makes it difficult mature new ideas Watson inspired systems are more likely to impact sooner than critical performance enabler/multiplier constructs/mechanisms Distribution C: 7

8 Distribution C: Eminent Opportunities Importance, Critical necessities, and emerging newtools make it possible. Human-perception mechanism inspired cognitive processors Force multipliers such as neural-net chips The need for flexible memories devices Need for efficient cloning of stateful computing threads Power efficiency Large scale computing imprecision and evolving nature of internal representation cognitive resolution 8

9 Conditions for Growth A set of critical problems A partial set of robust algorithms, clearly limited by hardware New non-von Neumann devices Willing partners to experiment Surrogate systems with incremental insertion of new technology A a new system developer with access to SME Distribution C: 9

10 CETE: Versatile Intelligent sensor node with cueing and embedded computing capability Essential functional blocks of an intelligent, versatile, reconfigurable, sensor envisioned for CETE: Modality agnostic, onboard computations, mission and situation awareness, autonomous decisions to determine when and what to communicate. 10

11 A Wide Range of Computations HIGH VALUE FUNCTIONS Numbers to decisions a stream of computations. No one-core fits for all. Heterogeneity of the core - is our strength. A serious challenge in managing >> 16 asynchronous cores. Model & Architecture: FITTS Chart Multiply and add y(k) = y(k-1) + a_k * x (k) Integral Transforms s.a. Fourier, Chirp-Z Integral histograms Moments Computation on sliding window Affine & Deformable Template Warping Median Filters sorting Hough Transforms -- evidence collection Multi resolution packages Multidimensional indexing k-d trees Clustering via k-nearest Neighbors Time-space equally weighted local operations: FLUX TENSORS, and video change detections. Logarithmic time-scale statistics summaries Episodic learning a time-to-live aspect Epochal learning has time-obliviousness Bounded Brownian-uncertainties in T-S dispersion, aka tight registration across time. GEOINT persistence & coherence exploitation Correlation is still a critical bottleneck Similarity based Matching & exploitations Identifying associations, storing and retrieving related elements is a challenge Distribution C: 11

12 Video Processing Summary (FMV and WAMI) State of the Art CETE capabilities Bandwidth Use Motion Detection Target Tracking Track Initiation Intentionally drop frame-rate, use lossy-compression, or transmit selective areas ; high latency forensics Frame differencing, Optical Flow, Flux tensors Limited number of targets tracked, due to noise Manual initiation of tracks on targets of interest (TOIs) Onboard processing to reduce products to a stream of -attachable, <5MB reports of focused video snippets and track information Flux tensors, research other novel techniques Track all movers, and nominated nonmovers Automated initiation of tracks on all movers, plus additional TOIs nominated manually Feature Extraction color, size, gradient, SIFT All of those, focus on more scale-invariant features Image Registration Geo-registration Leverage other programs for precise georegistration Parallax Estimation Other higher-level products Background averaging; Gaussian mixture models Human on the platform, human on the loop 3D Scene Estimation Intelligent onboard processing to cue the remote operator, or selective summarization 12

13 Challenges in Big Data Analytics Distribution C: Advantage in labor power could converts into global competitive advantages Ability to do more, do better with more intelligent power, less labor power, would be a force multiplier Big data limits human performance in analysis and decision making High-performance information technologies for massive analytics enables the journey from big data to information, knowledge and wisdom. R&D needed to achieve trusted autonomous systems that are capable of learning, reasoning, inferencing and interacting with human. Computing hardware technologies reach physical limits in area, power and performance Three-dimensional integrated circuits and systems with optimized performance under size, weight and power (SWaP) constraints. RDT&E for nano/quantum/neuro device and system technologies. 13

14 Multi-Tiered Approach to HPEC Challenges TECH DEVELOPMENT TRUSTED ARCHITECTURES FUNDAMENTAL SCIENCE Distribution C: 14

15 Getting to Hardware/Software Foundations of Trust Intelligently mix COTS and GOTS Use quantity and proximity to the pointy end of the spear as guides. Cyber-harden the building blocks CPU, DSP, FPGA, memory, A/D, mixed analog-digital Incorporate resilience and agility innovations Bring S&T to high technology readiness (TRL 6) with red teaming Rapidly compose many embedded systems solutions Open hardware and software to third party inspection through intellectual property management Employ and extend COTS innovation for cyber security Distribution C: 15

16 Summary Future embedded systems are challenged to continue delivering extreme performance in small space But cyber security places additional challenges for trust, agility and resilience Powerful technology drivers are still at hand to meet the challenges Computer architecture innovations Nano and quantum advances 3D stacking Algorithm development and mapping to architectures Distribution C: 16

17 17