Day 1 of the forum is devoted to "Efficient embedded deep learning for vision applications," presented by Marian VERHELST (KU Leuven, Belgium):
- Introduction into deep learning
From neural networks (NN) to deep NN
Benefits & applications
Training and inference with deep NN
Types of deep NN
Sparse connectivity
Residual networks
Separable models
Key enablers & challenges - Computer architectures for deep NN inference
Benefits and limitations of CPU and GPUs
Exploiting NN structure in custom processors
Architecture level exploitation: spatial reuse in efficient datapaths
Architecture level exploitation: temporal reuse in efficient memory hierarchies
Circuit level exploitation: near/in memory compute
Exploiting NN precision in custom processors
Architecture level exploitation: reduced and variable precision processors
Circuit level exploitation: mixed signal neural network processors
Exploiting NN sparsity:
Architecture level exploitation: computational and memory gating
Architecture level exploitation: I/O compression - HW and SW optimization for efficient inference
Co-optimizing NN topology and precision with hardware architectures
Hardware modeling
Hardware-aware network optimization
Network-aware hardware optimization - Trends and outlook
Dynamic application-pipelines
Dynamic SoCs
Beyond deep learning, explainable AI
Outlook
- Data fusion: principles and theory
Bayesian estimation
Priors and likelihood
Information content, redundancy, correlation
Application to image processing: recursive maximum likelihood tracking, pixel fusion - Pixel level fusion
Sampling grids and spatio-temporal aliasing
Multi-modal sensors, interpolation
Temporal fusion and superresolution
Multi-focal fusion - Multi-camera image fusion
Occlusion and inpainting
Uni and multimodal Inter-camera pixel fusion
Fusion of heterogeneous sources: camera, lidar, radar
Applications: time of flight, hyperspectral, hdr, multiview imaging
Fusion of heterogeneous sources: radar, video, lidar - Geometric fusion
Multi-view geometry
Fusion of point clouds
Image stitching
Simultaneous localization and mapping
Applications: remote sensing from drones and vehicles - Inference fusion in camera networks
Multi-camera calibration
Occlusion reasoning for multiple cameras with an overlapping viewpoint
Multi-camera tracking
Cooperative fusion and distributed processing
No comments:
Post a Comment
All comments are moderated to avoid spam and personal attacks.