Machine Learning Study Group
 |
Welcome! We meet from 4:00-4:40 p.m. Central Time. Anyone can join. Feel free to attend any or all sessions, or ask to be removed from the invite list as we have no wish to send unneeded emails of which we all certainly get too many. |  |
Contacts: jdberleant@ualr.edu and mgmilanova@ualr.edu
Agenda & Minutes
108th meeting, Apr. 5, 2024
- Three CS masters student project proposal presentations today! Back to our regular program next week (or maybe the week after).
- SS:
- Amazon product ratings analysis. Question: Why do you think linear regression will work? Answer: Simplicity. It was suggested that simplicity might not be enough of a reason. Question: What criteria for evaluating the results? Answer: (this was not really addressed). Question: what metric will you use, or how will you choose features to use? Answer: (this was not really addressed).
- SB:
- Topic: Hybrid feature fusion, model selection, and coffee leaf disease classification.
- Question: Do we need to decide the model before research? Sounds backwards, research should dictate the model approach.
- Answer: (this was not addressed).
- Question: 1) Genealogy of coffee leaf data set, normalization, basis for integration and comparison.
2) diagnosis implies a decision tree. What is the space of possible outcomes and what are the range of interventions that take place? How many treatments are there for how many diseases? At what scale are the leaf features recorderd? Is this scale similar among disparate datasets?
3) CNNs are a good choice for this problem. One can train with the same disease under different lighting and sampling conditions, and conversely different diseases under similar lighting and sampling conditions.
4) Why bother with the Nano? Why not just upload the training set to the cloud, and then acquire the test set, upload it to the many-GPU cloud and then receive the results. This is not an application which requires real-time or near real-time decision making, so you don’t really need to worry about edge computing either because the deadlines are soft.
Who labels the test set? Are there culture methods that can give you high accuracy in the test set? - Answer: Feature fusion will be tested using multiple methods. There will be three disease categories.
- Question: 6) Hardware is a don’t care provided the environment you are using is portable. Jupiter is a good choice. Flask is unnecessary. You can build out a compact client environment without creating too grandiose a framework.
1A) Consider choosing one methodology and doing it very well and ensure that its coverage is adequate for the spectrum of disease states you are concerned with. - Answer:
- Question: What if you reduce the scope of the project? How would you do that? Why would you not want to do that. For example, why bother requiring to do it on a nano processor?
- Answer:
- Question: How are you going to label your training set reliably? Since you are not an expert in coffee plant diseases.
- Answer:
- PA:
- Topic: detection, diagnosis and treatment of mental health disorders. Use machine learning.
- Question: There is a cycle: diagnosis, then treatment, then evaluation of whether the treatment helped. How much of DSM-V would be covered by this project? In more detail: 1) This is a fantastic topic and of major importance to the world.
2) Diagnosis —> Intervention —> Evaluation —> Go to 1.
3) How much of the DSM-5 would be covered by this methodology? Bipolar, OCD, Tourette’s Spectrum Disorders, Schizophrenia and subtypes (catatonic, paranoid, etc), Anxiety, Depression, Mania.
4) Would detection of emotions be the same as detection of mental illness? Everyone shows all the listed emotions, but when any emotion predominates we may be concerned about a behavior pathology. How do you get form emotion detection to mental illness diagnosis? Maybe you are doing emotion classification, not illness diagnosis.
5) Decision trees, ensembles of weak learners, the success of XGboost could be used. - Answer:
- Consider working with the VA, as they are very concerned with veterans who return from military service with mental health problems.
- What about suicide prevention? If you can somehow prevent even a few, it is very worthwhile. You don't have to prevent all of them.
- Please contact Grace to discuss data and etc.
