A motorcycle talk founded the CMMRS. I am wondering if small applications like these are enough in having actual impact. It does have some impacts on the people there mmm. The seed was eaten, as his metaphor, industry is eating out most of the talent. And academia is still important as a propulsion engine.

I don’t think it is something general, you need something specific for sure. They say its not recruiting informercial and repeat it a lot. But it seems very strange. They say its more christian like opportunity to

Fields of Expertise

Distributed Systems

  • Peter Druschel seems a systems guy, decentralized systems and bluethoot things
  • Lorenzo Alvisi

Good questions

“How do you think industry and academia should ideally collaborate in [your field]?”

  • “What can academic research learn from the speed or pragmatism of industry—and vice versa?”

  • “What makes a collaboration or mentorship productive in your experience?”

  • “What are the qualities for a good student or employee?”

    • On the other side what is a good supervisor or good manager?
    • What are the differences and commonalities or being supervisor and manager or student or employee between academia and industry? “How do you balance reading, building, and writing to stay productive?”

  • “What are the qualities for a good student or employee?”

    • On the other side what is a good supervisor or good manager?
    • What are the differences and commonalities or being supervisor and manager or student or employee between academia and industry? Company business goals is within company. Good manager implements the company objective (I think it depends on the level of the manager). The company can shift the direction, and the manager ’s job is able to unite people to work towards that goal. Finish task as the employee in a company is the goal. In university you have more flexible goals, research excellence is the main goal there, self-motivated, big -thinkers. I think it’s the case also in nice industries.

In her view execution is more important in industry. Advisors’ job is well-being of the student and deliver research agenda is the good think there. But this is also a good think for the manager. Advisor’s another goal is to pick students job which is feasible within his/her competence or PhD. So this is very difficult for an advisor (could be sized correctly) And should be brought along. Industry has smaller time frames. This is still biased towards the academia in my opinion.

Talks

28 July

Christina Giannoula: Fast Software

Talks about fast processing in hardware and software, something that is very similar to what we have done in Fast Linear Algebra.

They cite Minuet and ColorTM as some examples in doing this. NVIDIA designs both algorithms and hardware, this is one other reason why it goes fast. One thing that we know is that the processor speed is growing faster than the memory transfers.

One of the examples is point cloud networks (sparse convolution kinds of computation), so there is a lot of this work to make it closer to the hardware, and making it more efficient. So they use sparse convolution to compute with that. The normal sparse convolution has lots of data movements. One note you can look at is Sparse Matrix Vector Multiplication. We have ways to encode the sparse matrix so that the data movement is lesser.

Hash tables are currently used for modern runs. TorchSparse is one example of modern framework that uses it. But hash tables have high data movement! which is not good.

Lorenzo Alvisi PhD panel

Articulate, defend, and execute on an ill-defined question. The discourse is very biased anyways Nothing to write…

29 July

Meeyoung Cha: Ethical challenges with big-data

Data science for Humanity, she works mostly with NGOs and social good purposes, WHO for example in pandemic analysis. See is very warm in her talk, perhaps this is important.

She is trying to answer the question of how to use big-data to give best value for the society. A nice thing is that she is very practical, high impact.

It seems there is the technology for many things, but not the students or the knowledge to do that. It is surprising that some people don’t even know that these problems exist.

Digital transformation of customs: other problems in customs, since tariffs are by product and by country. They wanted to maximize tax revenue for tax in different developing countries in sub-saharian countries. But officers did not use that because it loses authority, something human so.

Fraud detection: Illicit traders used covid to make more trades, so there was a lot of fraud detection works there.

So a lot of works on social good.

Information has an unique trace from what it starts and who it interacts too. In this case information has a thumbnail (one of the first algorithms for fake news in social media). Facts and true news are more connected, while fake news are consumed alone, but both spreads fast.

Geographic Information Science We have one pixel for 10 meters, but if one pixel for 10 centimeter or 1 meter we can do stuff, this data is becoming available, and it is quite exciting to see what can be done for this. AI for social good. We can say size of the building (wealth?) how many people live there, how many people it can hold, so population density. For example storying things at the border, and estimate of trades between countries (streets, small roads). For example estimate poverty rate of north korea from satellite images. You can decide how sanctions affect economic region activities, but this is only for visible economic things.

Every point on hearth is being monitored every 20 minutes (with different resolutions).

Data is becoming cheaper and cheaper, we lack structured algorithms to extract insights from that unstructured data (from different sources and things like that).

Ulaanbaatar a new form of slum, but also announcing the news and things like that. Because people don’t like to hear that you live in a slum.

Next problem is AI prevalence between real people and fake people. Broken window theory.

Johannesburg south africa time magazine: !GUtOPekXcAAeCRm.jpg|369

Also humans in virtual space is a big thing.

Suggestions for future scientists:

  1. Find your why. When research gets messy (and it will), your “why” keeps you anchored. Don’t chase trends - chase purpose. Think in 3-steps: what do I want, why do I want it, and how.
  2. Smash mental walls. If you ever find yourself saying “I can’t do this because… “, then, challenge it imemdiately and break the wall. Rewrite your story.
  3. Set big goals. Forget “achievable goals” Aim 2x, 5x, 10x higher goals. Meet the version of you who dreams bigger - and walks taller. Get advice from the future self.

Alexandra Silve: Algebraic Network verification

NetKAT

We want to have verification in network routing policies. Verification using software defined languages are more main streem This is about Control Plane and Data Plane level of routing. More precisely in data plane, where it is decided where packets are forwarded in the network. Every router has a routing table, which match action logic.

We want to answer questions like: from initial source, can I reach a certain destination? This is the reachability problem. Are all host reachable from every host. Is a subnet isolated from another subnet? These are questions that are useful. Data is isolated in certain geographic types (Cloud providers for example might want to provide these guarantees). From a programming point of view, these things can be answered in a similar manner. This can be compiled into some structure that can be efficiently checked automatically.

Domain specific languages are good if they help in reasoning for specific problems.

There is some program optimization procedure, a map between this specific language and some program properties, in a way its generalizing that property.

We can express control flows and while loops with kleene algebra:

$$ \text{if b then p else q } \to b \cdot p + \tilde{b} \cdot q $$

We can also compute the while loops using kleene algebras. Then you add semantics, with actions this is how we are executing this.

30 July

Manuel Gomez-Rodriguez: Counterfactuals in Minds and Machines

Counterfactuals are the ability of imagining the consequences of another case (the what if case). In cognitive science it has many usages:

  • Modulation of emotions
  • formation of intentions
  • causality and explanation
  • Responsability and blame (what you should have done).

This is helpful to inform the model to decide what is normal, recency reasons and controllability. And the how thing.

Normality is another way to measure that thing, which is an hypothetical counterfactual. https://psycnet.apa.org/record/1986-21899-001. Recency: they are closer to what you are experiencing. Controllability: depends how controllable are those things.

Positive and negative counterfactuals

There are also different types of counterfactuals. One easy thing is defining the outcome of one thing as better or worse.

Downward counterfactuals lead to positive emotions (Unlucky victims or lucky survivors)

This is in a paper -> Teigen & Jensen, Unlucky victims or lucky survivors? The opposite is valid too:

Upward counterfactuals lead to negative emotions

In paper “When less is more counterfactual thinking and satisfaction among Olympic medalists”. But it is useful to aid self-improvement and learning from mistakes.

reinforcement learning only penalizes or rewards actions that you actually did. ~Manuel Gomez-Rodriguez.

Mental simulation

Craik -> nature of explanation in 1943, says that we have small-scale model of the external world that we use to decide what is reality. The important takeaway here is the correlation between counterfactuals and our idea of cause and explanation.

One another thing is that you can define something along the line of responsibility judgements with this and (Baker et al. 2011) similar inverse planning. This is called person inference, and it is close to game theory stuff.

Causal Machine Learning

We use counterfactual reasoning with outputs, data, users, of these things….

Meta analysis

From https://arxiv.org/pdf/2206.15475 Looks like causality suffers from:

  • Absence of ground truth data
  • Absence of nice frameworks for causal inference (there is a lot of contribution we can have from this point of view, but need to work out some framework)

Main Concepts

Structural Causal Models

$$ X_{i} := f_{i}(PA_{i}, U_{i}) $$

Where $PA_{i}$ are direct causes, $U$ are jointly independent variables which condition the deterministic approach. The nice thing here is that with this model we can reason about counterfactual queries.

This is defined in section 7.1 of the book.

Interventions

We can actually say something like do semantics, where we can have some intervention on the causal system.

$$ P^{\mathcal{M} : do(T = 1)} (B= 1) = \text{ num} $$

To write about interventional probabilities about some model. It’s difficult to say if this is a intervencion or a real counterfactual explanation, because here we are forgetting about the causal relationships.

Counterfactual Fairness

$$ P^{\mathcal{M} \mid X =x, A=a; do(A=a')} (\hat{Y}) = P^{\mathcal{M} \mid X=x, A=a}(\hat{Y}) $$

Meaning if $a'$ is a discriminator factor, for example the demographic group, we would like to have that probability to be the same, so that we are not accounting for that value.

Counterfactual Harm

Doctors prefer things that do not have any harm with respect to the medicine they are creating. You need a concept of utility here with actions and similar things.

31 July

How to write

He says elements of style are too general. Here its constructive-> easy to check and easily improvable.

  • Style: Toward clarity and grace. By Joseph williams
  • Learn technical writing in two hours pwer week with course notes.
  • Simon Peyton Jones How to write a great research paper

Flow

Flow: clear each sentence and paragraph relates to the adjacent ones, and also in general. Information has to connect with previous text, and this is a flow problem.

Begin with old information and end with new information to continue, this is a classical writing technique.

Coherence

Everything should related to the big picture. Both sentence and paragraph level. One principle is one paragraph and one point, expressed in a single sentence, called the point sentence, typically at the beginning of the paragraph, which has more importance usually. This also helps other people to make sense of the rest of the paragraph. One paragraph is around 6 sentences most of the time.

Name and Time

We want unique names for the concepts we have and introduce the concepts only when its needed (students tend to give every information before).

Principles

  • Top-Down is the best thing, start with accessible papers and then go down with the technical details to convince.
  • Tell them what they want to know:
    • The overall contribution (The novelty)
    • The scientific soundness and the results achieved. (Important?)
    • Why should I care (Why is it interesting)

01 August

Rediet Abebe

When does allocation require prediction? As a policy maker I care about actual outcome: predictive vs societal outcome are not the same, as the policy maker, it is more important to take care about the real outcomes of the thing. We want to integrate these frameworks within the existing decision-making ecosystem, if not that would be difficult to do. Predictive and allocative mechanisms are different from each other.

Early Warning Systems (EWS)

Predict educational outcome of individual students, and do small interventions to analyze this. They wanted to identify resources and how you can improve the societal outcomes (individual risk scores for everybody). This was a result of a bipartisan agreements, they faced a lot of dropouts, and they collected data and build systems. But people don’t have consensus if the systems are working.

Wisconsin’s DEWS work and model to create risk category and probability of dropout EWS no longer exist because of their work. 200k students impact in all schools in Wisconsin. Tracked demographics, academic performance, community-level. Early systems where underconfident.

The planner has to trade off relying on earlier and potentially noisier predictions to intervene before individuals experience undesirable outcomes, or they may wait to gather more observations to make more precise allocations.

02 August

Krishna P. Gummadi: Better foundations for foundational Models

The Pathetic Dot Theory

The term “pathetic” doesn’t mean pitiful; it means acted upon (from the Greek pathos).

There are some forces that constrain what we can do in society. For example:

  • Laws
  • Norms (not encoded in law, but enough to ostracise you)
  • Markets (exchanging goods and services)
  • Architecture (physics, biology geography and computer code, in the sense of similar constraining force as law, because they are automatically enforced, like preventing you to post certain things)
      Norms
        ↑
Law ← [You] → Market
        ↓
   Architecture

This diagram shows how each force pulls or shapes you, often simultaneously.

There are systems in AI that influence what you consume and what you will do.

  • Digital Public Spaces: are influenced by search and recommender AI.
  • Digital market places: AI market makers for example.
  • Decision Support Systems: Risk prediction AIs that decide who to admit to certain school and similar dilemmas.

In the past AI was programming, then it was machine learning (statistical learning), then deep learning (NNs), now we have foundational models (just zero-shot tasks, instead of task specific learning). Fairness, robustness, hallucinations are other problems. In the past they were not problems since we did not have this ability to automate.

Main issues in modern world are:

  • Governance
  • accountability
  • privacy

For the professor we have layers of AI:

  1. Pre-training
    1. Learning useful representations of the world knowledge (this seems to be true, even if you learn on random labelled data ->https://arxiv.org/pdf/2210.14019)
  2. Finetuning
    1. Soft skills and domain specific knowledge
  3. Inference
    1. System level optimization (KV caching and stuff, see Optimizations for DNN).
    2. Quantization and dropout. This is not exactly lobotomizing as the professor presents it.
  4. Configuring
  5. Prompting
  6. Agentifying.

Systems with Goals?

In Simon’s terms, an artifact is anything that is artificially designed to fulfill a purpose.

It is useful to reason with these systems as something that has goals. Herbert Simon: psychology as the science of the artifact.

More efficient to get the LLM to get the things you want done, if we analyze it with a psychological view. But we are not antropophizing them.

Here Theory of Mind becomes important, like my paper on ToM.

Testing

How can we test the models in a correct unbiased manner?

  • Test administration
    • Could be subject to overfitting by prompt engineering
  • Test scoring

Post

For the past few days I have been attending CMMRS at the Max Planck Institute for Software Systems and in Saarbrücken 🇩🇪. The main aim of this Summer School is to guide undergrads and grads in CS through the decision of how to chose a PhD program, or not doing one at all.

Throughout the week, we covered a wide range of topics: Panels from testimonials of PhD vs industry experiences, applications tips to grad school, social events with the fellow students and lectures about (very) different topics ranging from Structural Causal Models, Data Science for Social Good, programming languages, network algebras and even more.

Among all the talks during this past week, I personally liked the talks by Meeyoung (Mia) Cha where she showed how modern technology can have real impact on society and public policy, (1, 2, 3), Manuel Gomez Rodriguez’s work on Causality, showing how we can use these tools to think formally about ideas of fairness, harm and aid decision making (4, 5) and Rediet Abebe’s social commitment in helping Winsconsin’s schools to provide better education (6). I believe they are modern role models showing how we can leverage data analysis, mathematical modeling and personal drive to have a tangible impact on the people around us, and remember many of us why (some) research is really important.

I leave the city today with a stronger faith in what we, as Computer Scientists, can contribute in this world, an expanded view on some research directions, and a thankful salute to all the students, organizers, and staff that hosted me during this week for the shared memories :).

A shoutout to Shreya Kochar with her personal testimonial on the event https://substack.com/@shreyak164865/note/p-169922537

For the past few days, I’ve been attending CMMRS at the Max Planck Institute for Software Systems (MPI-SWS) in Saarbrücken 🇩🇪.

This Summer School is designed to help undergraduate and graduate students in Computer Science navigate the decision of whether or not to pursue a PhD.

Throughout the week, we covered a wide range of topics: panels with testimonials comparing PhD and industry experiences, grad school application tips, social events with fellow students, and lectures on subjects ranging from Structural Causal Models and Data Science for Social Good to programming languages and network algebras.

Among all the talks, I particularly appreciated those by Meeyoung (Mia) Cha, who demonstrated how modern technology can shape society and public policy ([1], [2], [3]); Manuel Gomez Rodriguez, whose work on causality showed how we can formalize ideas like fairness, harm, and decision-making ([4], [5]); and Rediet Abebe, whose social engagement is making a real difference in Wisconsin’s public education system ([6]).

To me, they represent a new kind of role model—showing how data analysis, mathematical modeling, and personal drive can lead to a tangible impact on the world around us, and reminding us why (some) research truly matters.

I’m leaving Saarbrücken today with a stronger belief in what we, as Computer Scientists, can contribute to the world, a broader view of emerging research directions, and deep gratitude to the students, organizers, and staff who made this week so memorable.

Shoutout to Shreya Kochar for her personal testimonial on the event: <a href=”/https://substack.com/@shreyak164865/note/p-169922537@shreyak164865/note/p-169922537”>https:/substack.com/

Let me know if you’d like a more casual or more professional tone.

Personally, I especially liked the talk by Meeyoung Cha on Data Science for Social Good. She showed that current technology can have real impact on society and public policy, making border custom’s procedures more efficient (1, 2), , and I was surprised about the level of detail of information extractable by Satellite data, which she calls Geographic Information Science 3.

Comment

I had the chance to reflect about my future during this week. CMMRS enabled me to understand research from the point of view of world-renowned academics and to engage with broad fields within Computer Science but also to just share beautiful experiences with the fellow students there.

CMMRS gave me the chance to reflect on my future in a focused timeframe. It helped me better understand what research looks like through the perspective of world-renowned academics, explore broad fields within Computer Science, and share beautiful experiences with fellow students.

References

[1] Baker et al. “Bayesian Theory of Mind: Modeling Joint Belief-Desire Attribution” Proceedings of the Annual Meeting of the Cognitive Science Society Vol. 33(33) 2011