Lecture notes Academic and Professional Skills
Course organizer: Makiko Sadakata
Lecture 1
Preferentie.
Er was discussie over hypothese veranderen. Integer?
Mijn vraag: Hoe werkt dat in data science? Normaal formuleer je een hypothese en dan verzamel je data. Als je vervolgens je onderzoeksvraag of hypothese aanpast nadat je data hebt gezien, is dat vaak niet wetenschappelijk integer. Cf. "p-hacking".
In welke gevallen is dat wel in orde?
En hoe werkt dat in Data Science, waar data wordt verzameld zonder een specifieke doelstelling?
Example of wrong generalization:
1) Possible conclusion (too generalized): Music makes you smarter 2) What is smart? IQ test only? 3) Is it the music itself, or just something positive?
Lecture 2: Giulio
Current level: between 1 and 2, between driver assistance and partial automation.
Focus on: accountability
We don't want that solutions become worse than problems.
If you read the manual and still have an accident, are you morally to blame?
Difference accountable and morally responsible? Latter requires a moral wrongdoing.
In current situation, there is room to remove any notion of moral wrongdoing.
"Legal solutionism": e.g. sue the car itself.
Scapegoat: because of unclarity, always blame e.g. the driver.
Story: avoid utilitarian attitude, about Ford Pinto.
1960s, Ford design horrible cheap Pinto, realize design mistake too late: if fuel tank is hit, it explodes on impact. Asked manufactorer what is the value of a human life? 5 doller per car tax. Made the calculation to not fix it and just pay in case of accidents.
This attitude is also becomes more problematic for AVs nowadays, to avoid dealing with complications. Is the problem really solved if Mercedes says: we will pay in any case.
Instead: moral accountability and -responsibility is about intrinsic values: self-understanding + duty of explaining behavior in terms of reasons. (This is a basis of society).
This also has an instrumental value: having a sense of moral responsibility increases safety because you put more effort and attention in what you do.
Risk of machine learning and problem with safety testing
Example: the "king's day":
1000 avs go out during king's day where everyone is dressed orange. It's also windy and it rains, and there is an eclipse. The AVs were not assessed on safety under these conditions. And neither was it trained on this situation.
The unpredictable stays unpredictable.
Under normal conditions it takes 11 billion miles (500 years) to have 95% confidence and 80% statistical power to "prove" that their failure rate is 20% less than with human drivers.
How are accidents reported?
Coordination problems in traffic causing accidents are not reported as accidents, because the car itself has no damage.
Example of volvo emergency breaking due to shadow at +100km/h. Normal driver hits the back: that driver is "guilty" due to not having enough headroom. But how is that realistic, to do an emergency break when nothing is happening in front.
Major problem with safety are humans. In the literature three problems that humans cause for semi automated systems:
- Cognitive limitations in interaction (naturally lazy, malleable)
- Epistemic inadequacy
*(we don't fully know what's going, you don't see everything, we filter a lot out)
- Might cause to relying too much on system
- Temporal limitations: we don't have enough time to reflect.
What distinguishes AVs from humans? AVs are:
- Rational
- Fast
"Ethical switch" solution.
Counter-argument: would you really know what you would do in those scenarios, in advance?
Giulio: high sci-fi quality. In reality an AI would have to perfectly recognize situations and apply ethical setting.
Giulio: in practice issues do not arise because AVs took the wrong rational decision, but instead because they go ballooney.
Statement from Mercedes: always saves the driver.
Meaningful Human Control
Not a solution to a specific problem, but proposed as a general framework that avoids situations such as the trolley dilemma where decisions are premeditated.
Two main conditions for meaningful human control:
- The tracking condition for control
- A form of attunement (including between intentions, dispositions of driver and system)
- Tracks reasons to act of users
- The tracing condition for control
Basically, the old human notion of control (e.g. over a hammer) is outdated for this problem. E.g. operational control: a fundamentally causal story.
This operational notion of control does not apply to AI systems. We do not want to control AI systems; we want them to control themselves!
We need to redefine the conditions to speak of control.
Lecture 3: Johan Kwisthout
See Kwisthout_Frugal_Explanations.md
Theoretical argument is not possible, because results depend on MAP distribution (uniform is bad news)
Problem with random nets: how do you compare? These nets need to be benchmarks that the state of the art is also tested on.
They are not randomly chosen but hopefully representative.
Method of my proposal: find relevant benchmark.
Benchmarks.
How do you determine relevant intermediate variables and background knowledge? E.g. ask domain expert.
Practical issue: can a domain expert give enough background knowledge?
First I have to show that if we have background knowledge, there is improvement. Before we go to a domain expert.
I can also compute intrinsic solutions, to compare approximations with.
1) Given background knowledge, faster results? 2) If yes: * a) ask domain experts about relevance --> mostly about accuracy! * b) estimate relevance --> mostly about speed! Is it still faster after we find background knowledge ourselves? - sample: if 0% I expect 0/10 samples to make a difference. - if 90%, I expect 9/10 to make a difference. * But if relevance distribution is uniform, I need exponential amount of samples to be confident.
If not, the second experiment(s) is/are useless.
When
stat. inference makes sense if you work with population subset and you want to generalize to whole poulatoin.
So if you do an exhaustive search over parameters, you already have the whole population. With static params: not random params, so also not relevant. You already picked representative sample.
When is an algorithm better? Running time? On which instances?
Vs. human participants: you can look into the algorithm to explain and interpret them. (Not possible with human experiments.)
Make more clear: topics have to be related to topics in class.
Lecture 4: Serge Thill on "Embodied cognition and robots"
See Chemero-Radical_Embodied_Cog_Science.md
Background cogn. science. On theor. level --> embodied cognition.
Robotics as potential way of embodying models of cognition.
Go way back: cognition is interaction with env.
Recently: cognition as manipulation of formal symbols.
Chinees room --> Strong AI: not suitable for modeling human cogntion. t
To perceive with your eyes, you alreday need to move your eyes. You need to act in order to perceive. Enbodiment in a nutshell.
We all agree body contributes something but disagree on what that exactly is.
Suprisingly many theories on ebodiment do not really discuss the body. A sensori-motor interface is already enough for some to speak of embodiment. Some would even thus argue you don't need a real body.At best it can help you understand i
What about the shape?
Crazy bird: two motors, but behavior cannot be understood based on motors alone. behavior is becaues of the specific morphology of the birds body.
Swiss robots. Broken front sensor.
Cricket robots: due to morphology (audio in legs), automatically turns towards loudest male cricket without procesing. Complex sound processing problem, or solved easily with the right type of body?
Fly eyes solve parallax problem with structure of the eye.
Morphological computation: what computions can I offload?
PMC pre-activity suggest that motor system is involved in decision making. Problem for computational theory.
Pen between your teeth make a smiling movement: rate jokes higher.
This is not yet embodiment of cognition though. Studies are also flakey.
Dynamic system question: how do I maintain my existence given pertubations between internal dynamics with environment and other agents. No reference to symbolic manipulation.
Distributed cognition: what is the cognitive agent running a war ship? There is no single agent.
Prpoer embodiment takes internal body into account. If that is really fundamental to cognition, that is a hard boundary for cognitive robots ability to model cognition.
Limit of representationalism: do symbols need to have meaning? Misses in GOFAI.
How do symbols get meaning? Interaction with env. etc.
See graph in slides: he calls non-functionalist what I expected to be functinoalist (left branch)
Claim that humans do not use representation is indefensible, irregardless of the non-representationlist discussion. You can never falsify this position.
second half
Models for grounding concepts.
Eliasmith:
Semantic pointer architecture: layers of representation (cf NNs). highest level is a symbol, but ultimately grounded in perception. This compression is reversible (so reactivation of perceptions is possible)..
Goal of embodied cognition: main interest is that models should be explanatory of biological cognition. Robotics can be a tool to achieve that.
Main point of Stramandinoli et. al. that does symbol grounding. Output NN nodes are actions, they can be combined. The actions ground the symbols.
But the model is predesigned. And embodiment is a joke: immediately replaced with activation action node.
Question: does the model change if you remove the robot body? Does it change the cognitive aspects of the model?
jkkk
Computational linguistics: "grass is never green". If you look at how often the word grsas co-occurs with green, you do not find this information. The non-informativeness problem.
How to perceive a container? Use your hands to put objects in another, see if they dissappear.
Windridge & Kittel 2010:
perception-action learning: learning to perceive from the consequences of your actions.
Lecture 5: Peelen - naturalistic vision
See Olivia-Scene_perception.md
Discuss specialization in ventral stream (what) between objects and scenes.
Multivariate Pattern Analysis: for picking up subtle differences within a particular brain region. E.g. in v1 same amount of activation for chairs and tables, but different patterns in object representations in LOC.
Compare voxel activation vectors: if they are very similar the objects are very similar.
James McQueen - Presentation and writing skills
I consolidated the background readings for this session in index_writing_principles.md .
These are some extra highlights from the lecture.
- don't show off: you are not trying to impress a teacher anymore
- "audience design" --> psycholinguistic term
- No meta-level padding: don't tell a structure that is completely obvious and already know.
- This does not mean structure is not important!
Use the hourglass model for the structure of your article.
Tip 12: focus on research not on researchers. Not: authors x and y found in paper Z that [claim]. Instead say: We know that [claim] (refer to paper Z)
Tip 18: Make contact at the beginning of a talk. Trick: not needed to look into eyes, can also look at forehead or so.
Apply mini-habits / habit-chunking.