logseq/pages/SALERNO MEETING.md

type:: meeting external-links::
tags:: PROJECTS/PRIN-EMELIOT
people:: date:: 15-06-2023 location:: Salerno isconsortium:: true

- ### My Slides
	- https://docs.google.com/presentation/d/1nO3AzY2hLir93rUUXP5Ck06ixz37obloj_ZZJuE__LE/edit#slide=id.p1
- ### Invited Talks
	- *Software Engineering for Machine Learning, some First Experiences (L. Baresi)*
		- **Data Scientist vs Software Engineering (Christian Kaenstner)**
			- SEs are concerned about cost, performance, stability, safety, security, and release time. DEs are not.
		- **Interesting papers**
			- [[2105.01984] Software Engineering for AI-Based Systems: A Survey (arxiv.org)](https://arxiv.org/abs/2105.01984)
			  collapsed:: true
				- ![image.png](../assets/image_1686813801082_0.png)
				- ![image.png](../assets/image_1686813885068_0.png)
			- [[2011.03751] Software engineering for artificial intelligence and machine learning software: A systematic literature review (arxiv.org)](https://arxiv.org/abs/2011.03751)
			  collapsed:: true
				- ![image.png](../assets/image_1686814007762_0.png)
		- **Federated Machine Learning**
			- They approach the problem of configuring and use ML systems as a self-adapting problem
				- HYPERFL
					- Extension of TensorFlow they have developed
			- Main issues
				- Privacy
					- Start analyzing local and send to the cloud service the output of the analysis
				- Network ovrhead
		- **DeepNurse**
			- It's not clear how this works. How is it able to detect new domains? By replacing neural networks at run-time? WHat do you means with self-adapt neural networks? Changes them or refine the values of the hyoerparameters?
		- **Feature model + reinforcement learning for self-adaptive services**
		-
	- *Machine Learning in Software Engineering: How the Software Engineering Lanscape Has Changed in the Last 20 Years (M. Di Penta)*
		- **Interesting papers**
			- [A validation of object-oriented design metrics as quality indicators | IEEE Journals & Magazine | IEEE Xplore (oclc.org)](https://ieeexplore-ieee-org.univaq.idm.oclc.org/document/544352)
			- [A comparative analysis of the efficiency of change metrics and static code attributes for defect prediction | IEEE Conference Publication | IEEE Xplore (oclc.org)](https://ieeexplore-ieee-org.univaq.idm.oclc.org/document/4814129)
		- **Notes**
			- Ensemble classifier
				- Creates multiple models each with different charactertics
					- THis is when a simple model does not work. Depending on the source model you decide which one performs better
				- Textual analysis to scource code. Andrea has been among the first to use such a kind of analysis.
					- [Recovering Traceability Links between Code and Documentation](https://ieeexplore-ieee-org.univaq.idm.oclc.org/document/1041053)
			- Recent paper from David Lo to detect Technical Dept
			- A recommender systems should be able to explain, but with deep learning this is difficult
			- Rise of deep learning applications to SE over the last 7 papers
				- Better hardware, etc.
				- Many data sources
					- World of Code - it's a recent data source
				- Better AI models
				- We have pre-trained models
					- ICSE2023 had a full session on this
				- [[LLMs]]
			- OUR ROLSE AS SE EXPERTS
				- SE for AI
					- Understanding
					- Refactoring
					- Testing (e.g. of DL models)
						- Search the slides of Paola Tonella who delivered an invited talk at ICSE
					- Debugging
				- Problem and data knowledge
					- Preprocess the data,
					- combine data from different sources,
					- pretrainend models might not be ok as they are for the problem at hand and thus need to be refined
						- Prompting in SE
							- [[2207.11680] No More Fine-Tuning? An Experimental Evaluation of Prompt Tuning in Code Intelligence (arxiv.org)](https://arxiv.org/abs/2207.11680)
				- Process and context knowledge
				- System Engineering
				- Usefulness evaluation
				- Legal Issues
					- LIcenses regulate derivative work. Different licenses (more or less restrictive)
						- GPL is the most restrictive, the most permissibe are MIT, BSD, Apache.
					- There are discussios:
						- AI generative code is not derivative code
					- Software bill of materials
						- Similar to the list of ingredients used in a food product, or to the Bill o Material available for manufacturing products
						- AI-related challenges
				- Dimensions to investigate
					- Productvity
					- code correctenss
					- Code quality
- ### Discussion about the experiments to be done in the project
	- Capire come gli sviluppatori beneficiano  di strumenti tipo CoPilot, ChatGPT, etc
		- 2 unità che si concentrano sullo stesso task
		- A livello di field study le unità possono prendere diversi progettini di interesse
- ### PREVIOUS MEETING IN PREPARATION OF THIS MEETING
	- {{embed ((64885a41-6eb1-4bee-a9eb-ede87be14999))}}