Workshop
https://www.youtube.com/watch?v=gfGTd7PXK54&list=PLMPUUgLIYH1ZJVEXTTZT82ipTprSu1hn8&index=1
Learning at the Wireless Edge
Vince Poor (Princeton University)
Two aspects
Using ML to optimize communication networks
Learning on mobile devices (the focus of today's talk)
Today's talk: focus on federated learning
Motivation
Federated learning over wireless channels (scheduling)
Privacy protection in federated learning (differential privacy)
Some research issues
ML
Tremendous progress in recent years (more data, increase in computational power)
Standard ML: implement in centralized manner (data center, cloud), full access to the data
SOTA models:
Standard software tools, specialized hardware
Wireless edge
Centralized ML are not suitable for many emerging applications
Self-driving cars, first responder networks, healthcare networks
What makes the application different:
Data is born at the edge (phone, IoT devices)
Limited capacity uplinks
Low latency & high reliability
Data privacy / security
Scalability & locality
Motivate moving learning closer to the network edge
Federated learning over wireless channels (scheduling)
Wireless: communication to the AP needs to go through wireless channels
Shared, resource-constrained
Only limited number of devices can be selected in each update round
Transmissions are not reliable due to interference
Questions
How should we schedule devices to update the trained weights?
How does the interference affect the training?
Scheduling mechanisms
Random scheduling: aggregator select N out of K users at random
Round Robin: divide into group
Proportional Fair: strongest SNRs
Design metric: age of information (AoI)
Age-based scheduling scheme for federated learning in mobile edge networks
Optimization algorithm in each iteration round
Wireless round robin
Privacy in federated learning
"privacy preserving": data remains on end-user devices
But end-user data can be inferred from the parameter (or gradient) updates
Approach: use differential privacy to protect end-user data
Refers to a type of privacy in which two datasets, one with private information and one without it, but otherwise identical, cannot be distinguished by a statistical query (with high probability)
Trade-off between privacy and accuracy
Other issues
Model efficiency
Resources on end-user devices are limited (e.g., energy, storage, computational power)
Trade-offs between # of layers, # of neurons per layer, accuracy
Communication efficiency
Limited data at the edge
Local data is sparse
Incorporating domain and physical knowledge
Security & Privacy
Robustness to malicious end-user devices & adversarial training examples
Other approaches to end-user privacy
Challenges in ML and the way forward
Ariela Zeira, Intel Labs
Challenges in DL
Compute efficiency
Memory overhead
Data efficiency
Online learning
Robustness
Knowledge Representation
Hyper Dimensional Computing
New paradigm for energy-efficient, noise-robust and fast alternatives to standard ML
Adventures in Learning-Based Rate Control
Brighten Godfrey, UIUC
TCP protocols: point solutions designed for specific environments, and far from optimal
Why does traditional CC architecture struggle?
TCP Reno, CUBIC, FAST, Scalable, HTCP
"Hardwired" control actions
Underlying conditions --> ideal control action
Not enough information about what's happening in the network
What is the right rate to send?
Network is a blackbox: but we can send at some rate and see what happens
Collect observations: throughput, loss rate, latency. We can summarize that in a utility function
A change in perspective
Traditional perspective: simple network model + well-crafted rules --> predictable results
"Black-box" perspective: the world is complex. Quantifying goal and observing effect of actions yields good decisions
A fit for learning!
Diverse, opaque environments
Only a trickle of information
Infer good action at millisecond timescales
Software components
Paper: PCC
Control algorithm: heuristic hill-climbing algorithm
Noise in measurement? randomized controlled trials
Where is the congestion control?
Equilibrium depends on utility function
Selfish utility-maximizing decision --> non-cooperative game
Promising performance
Upgrade: PCC Vivace (NSDI 2018)
Leveraging powerful tools from online learning theory
New utility function framework
Latency-awareness
Strictly concave --> equilibrium guarantee
Weighted fairness among senders
New control algorithm
Gradient-ascent --> convergence speed / stability
Deals with measurement noise
Performance: great improvement in latency & responsiveness, but still suboptimal in extremely dynamic networks (i.e. wireless)
Deep RL on congestion control (ICML 2019)
scale-free values to aid robustness
History length: what lengths work well
Training
Simulated environment
Order of magnitude faster than emulation
Each episode chooses link parameters from a range
Setting appropriate discount factor
Maximize expected cumulative discounted return
Future
Multi-gent scenarios: training, competition
Online training: challenge is to improve outcomes with limited additional training data
New uses
Scavenger transport
SIGCOMM 2020: Proteus: Scavenger Transport and Beyond
Different applications: software updates, CDN warmup, cloud storage replication, online video, real-time streaming, search
Elastic timing, inelastic timing
Scavenger design goals
Yielding: minimally impact primary flows
Performance: high utilization, low latency when only scavengers exist
Flexibility: dynamically switch, avoid separate implementation
Utility functions
Primary, Scavenger, Hybrid
RTT deviation as a competition indicator
Definition: standard deviation of observed RTT samples
Intuition: earlier signal of dynamics of buffer occupancy during flow competition
Proteus: yields more effectively
Cross-layer design
Dynamic threshold based on the application requirement (buffer occupancy)
Hybrid mode: to get the bandwidth when they need it
Improving QoE (rebuffer raito)
Rate control robustness (HotNets 2019)
Keeping up with rapid change
Recent acceleration of innovation in rate control
Approach: ML as an adversary
It gets rewarded when it finds environmental parameters that cause this algorithm under test to perform poorly (looking for the hard cases)
Do this carefully, rewarded for suboptimal performance of the algorithm, and smoothing to make it more useful results to see
Reward = -1 * protocol score + optimal score - smoothing penalty
Implement: ABR video, CC
Lessons learned
What worked
Modular architecture
New control algorithms
New utility functions open new issues
Learning-based control can improve performance over traditional protocols
Industry implementations
Open challenges & opportunities
Performance: fast decisions v.s careful decisions
Even one RTT can be a long time - especially in fluctuating wireless environments
Understanding protocol robustness
Existing opportunities in system design
Complexity in systems, environments, and application needs lead to opportunity for inference
Restructure systems for a learning mindset
Last updated
