Vladlen Koltun	Learning to Act

• Reinforcement Learning

  • Reinforcement learning is learning what to do - how to map situations to actions - so as to maximise a numerical reward signal.

• Why is it called Reinforcement ? —> Conditioned reflexes - Pavlov 1927

  • reward and punishment
  • History (Pavlov and the dog , Skinner and pigeons - behaviorism -> animal behavior can be explained in terms of conditioning)

• Natural Supervision Example

  • act and observe the consequences
  • learn to predict the consequences of actions
  • Learn to map observarion x action -> future observation

• LEARNING TO ACT (Markov decition proces)

• Problems:

  • State - “Grid world” don’t work with 3D real world situations (“perspective”)
  • Reward - “What is the reward for life?”,
  • Discounting - Its extremely aggressive. In reality you can chose something with smaller discount
  • Markov Assumption - No history

[comment]: <> (

Emilio Frazzoli	Title

• Mobility as a service is the way to go
  • Easier to scale / unroll
  • Cost efficient for consumer and provider
  • Controlled Environoment
• Levels of autonomy is misleading
  • OEMs try to increasy autonomy step by step
  • MaaS (AVs as a service (MaaS)) approach allows to unroll by scaling up the scope
• Utilize prior Knowledge where applicable
  • Rules of the Road don't need to be learned
  • Use formal logic for decision making
  • Utilize graph theory for planning
  • Combine traditional methods with learning

Column 1	Column 2	Column 3

Hao Li	Title

• Allow realistic digital communication
  • with everyday hardware
• Requires to capture fine grane details
  • to transport non-verbal aspects of communication
• Use deep learning to infer those details
  • instead of expensive sensoring equipment

Trevor Darrell	Title

• Domain adaptation
  • Use adversial network for cross-domain adaptation
  • Force the network to learn a representation which is indistinguisable between two domains
  • (future project?)
• Natural Supervision
• Curious learning
  • Use loss as reward in RL
• Visual Explanations
  • Use attention to make model interpretable

Sanja Fidler	Learning Embettings of Images

[comment]: <> (

• Use word2vec (or similar) to learn embetting for words
• RNNs for sequential models
• Tasks:
  • Image Captioning
  • Classification using description (zero short)

Vittorio Ferrari	Learning Embettings of Images

• Weakly supervised learning: Learn from a lower degree of supervision
• Semi-supervised learning: combine labeled unlabeled data (useful project?)
• Utilize that different combinations of labels appear
  • Face recoginition example
  • Training: EM
• Weakly Supervised Deep Detection Networks
• One point per class (Use object model)

Ross Girshick	From Visual Perception to Visual Reasoning

• Respect Variation:
  • Translational Variation
  • Rotational Variation
  • Scale Variation
  • Deformation Variation
  • Intra-class Variation
• Problem might be Invariances or Equivariance w.r.t. variation
• Look at old papers
  • You might find good ideas
• Mask-RCNN
  • Modularised Approach
  • Feature Pyramid Networks for Object Detection
• Kluger Hans & Visual Question Answering
  • is VQA only nearest neighboor?
  • How do we know our models are as sophisticated as we think
• Diagnosing Errors in Object Detectors
  • Useful Error diagnosation tool
• CLEVR: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning
  • Dataset for analysing VQA

Abhinav Gupta	Self-supervised Learning of Visual Representations

• Use auxiliary tasks as pretraining
• For classification the following tasks are useful:
  • predict color channel (given RGB channel)
  • predict relative positions of text
  • jigsaw puzzl predict position of patches
  • predict content of hole
• Kluger Hans Issue (shortcuts)
  • low level image structures (edges)
  • camera artifacts
• Solution: Jitter input
  • Augment position
  • drop color channel
• Natural supervised training not yet competative with ImageNet pretraining
  • (ImageNet labels as natural supervision for some tasks?)
• Validate weakly supervision
  • only finetune few layers
  • use view data

Column 1	Column 2	Column 3	Column 4

Jan Peters	Learning Visuomotor Skills in Robotics

• Machine Learning is not Robot Learning
  • Robot Learning has its own unique challanges
• Use prior knowledge (e.g. about physics) where avaible
  • we have very good models, e.g. for motion planning
  • not everythink needs to be learned from scratch

Raquel Urtasun	Towards Affordable Self Driving Cars

• combine computer vision knowledge with
  • e.g. solve stereo by computing matchings
  • use flow for detection / segmentation
  • geometric aware cnn
• prediction (of road user behaviour) is an important task
• Visual Odometry (localization / mapping) needs more research
• Cityscapes offers poor meta-data
• 3D detection is required

Column 1	Column 2	Column 3

Andrew Fitzgibbon	Optimiztion

• Finding solution fast (speed) is the key goal
• To find global optima, search for local optima
  • on some problem spaces (quasi-convex) reaching global optima is garantied
  • otherwise use random research
• optimization is an ongoing field of research
• better optimization algorithms can improve machine learning

Column 1	Column 2