Database Systems



Artificial Intelligence

CSE 5290/4301, Spring 2021

Instructor: Debasis Mitra, Ph.D.

Office: OEC 350 E-mail: dmitra ‘at’ cs.fit. edu

Class Home Page:

Office Hours: MW 2-4pm I will try to be at but you may need to let me know first by e-mail (or by appointment)

Grading plan:

Graduate stds: Takehomes (5-10): 15%, Exams (3-4) 35-40%, Project: 45-50%

Undergraduate stds: Takehomes (5-10): 20%, Coding exercises/Project: 30%, Exams (3-4) 50%

All assignments/tests/etc. are online, asynchronous.

CLASS coordinates: 8:00 pm - 9:15 pm TR Crawford Bldg 230 Jan 11, 2021 - Apr 30, 2021, Zoom URL:

SYLLABUS FOR AI Fall 2020 FINAL EXAM.

Materials under *---* are excluded. 3.3-3.4.5 means 3.3 through 3.4.5

SEARCH, up to IDA* search, but not SMA*,

Adversarial search (min-max) & alpha-beta pruning is included

Up to materials on my slide Constrained Optimization

Ch 4-4.2

Ch 5-5.3

CONSTRAINTS

Ch 6-6.2.5, 6.3-6.3.2, 6.5 (*use of symmetries on p226*)

Temporal reasoning: My slides

LOGIC

Ch 7.3-7.5.2, *completeness of resolution*, 7.5.3-7.6.2, *walksat algorithm*,

Ch 8.2-8.3

Ch 9-9.2.2, 9.3-9.3.2, 9.4.1, 9.5-9.5.3

PROBABILISTIC REASONING

Ch 13.2-end

Ch 14-14.3, 14.7.2-3

Bayesian inferencing

LEARNING

Decision Tree 18 – 18.3.4

Evaluation 18.4-Model Selection 18.4.1

Regularization 18.4.3

Learning Theory 18.5.0 only

Regression 18.6 – 18.6.2

Classification 18.6.3 – 18.6.4

Neural Network 18.7 – 18.7.4 (exclude exotic varieties of NN on my slides)

Non-parametric models 18.8 – 18.8.4

SVM basics 18.9

Clustering basics (from my slides)

ETHICS

My slides

Florida Tech Academic Calendar:

Detailed activities in Fall 2020 mapped on to the dates of Spring 2021, for now.

This table will act as our curricular map for the semester. It helps me to track progress and plan next few days of the class, assignments, etc., and I edit this frequently. I keep this online in your view so that you may be in sync with me. However, use it at your discretion.

Meetings: ~ 28. Lectures planned: Search 4, Constraints 3, Automated Reasoning 4, Probabilistic Reasoning 4, Machine Learning 5, Ethics 1,

Exams 3 (asynchronous), Std presentations 4

| |Activities (Spring2020) |Comments |

|Jan 12 T |AI an Introduction, and Get-to-know Pre-quiz (0 point but must | |

| |submit) | |

|Jan 14 R |Required CS background list | Graduate projects |

| |Big O-notation (Algo-Intro slide 1-10), |described below this table, Due dates: YTD |

| |Djikstra algorithm (Algo-Graph slide 23-25), Min-spanning tree | |

| |algorithm (Algo-Graph slide 54, 57-8) |Undergrad students pair up |

| | |for the project. |

| |Complexity Theory - lite | |

|Jan 19 T |AI SEARCH (From my slides): 8-puzzle: | |

|(Jan 18 MLK | | |

|Holiday) |BFS- Uniform Cost; DFS-Depth Limited | |

| |Iterative-deepening | |

| |Heuristic Search: ucbfs, rules for heuristic function | |

| |MySlides 15-35 | |

|Jan 21 R |Classes will continue to be online through the semester. I will |UG project announced |

| |lecture from classroom next week. However, given the COVID | |

| |infection rate around, I strongly encourage you to avoid | |

| |congregation. If no one comes to the class, I also can teach from| |

| |safety. It is also FAR more productive for me than the restricted| |

| |in class lecture. | |

| | | |

| |Homework-Search on BFS, DFS on canvas | |

| | | |

| |Heuristic Search A*, IDA* (MySlides) | |

| |Text Slide Ch4a | |

|Jan 26 T |Local search (not recorded, sorry the class started late due to | |

| |security’s being unaware of this class coordinates) | |

|Jan 28 R |Local search continued | |

|Feb 02 T |(Please let me know if you have/had a bio-minor) |UG project updated |

| | |Grad project & Spl. UG project proposals due Feb 5/F |

| |Remind importance of heuristics in PRUNING search tree to reduce | |

| |effective search space | |

| | | |

| |Game (Adversarial) search MySlides | |

| | | |

| |Logic Module starts | |

| | | |

| |Constraint Satisfaction will be covered later:: | |

| |REASONING WITH CONSTRAINTS: Motivating with Map/Graph coloring, | |

| |Backtracking, Forward Checking (TextSlides) | |

|Feb 04 R |Grad & Spl. UG project presentation/discussions 5 min each group| |

| | | |

|Feb 09 T |More on DFS stack scenario: Myslide 13 |Test1-SearchAlgorithms this week online Canvas |

| | | |

| |Automatted reasoning / Logic continues: Models, Inferencing, |No one was present in classroom today, |

| |Logical equivalence, Syntax-Semantics-Model, |I will take the class on zoom from home |

| |Satisfaction-Entailment-Inference procedure-Validity, |the next class. Let me know if any of you |

| |Model-checking algorithm (up to TextSlide 41, ch-7) |wants to attend in-person. |

|Feb 11 R |More on Best-First-Search (A*) PriorityQ scenarios (Myslide 23): |A Canvas test (multiple-choice) on Search, |

| | |opens 9:20pm, 35 min from your start, late after F/11am |

| |Brief intro to artificial neural network: Textslides Ch20b, CNN: |* Some figures are uploaded as PowerPoint (SearchFigures.pptx)  for the test where the syllabus is on Canvas. |

| |Myslides 42, 45-6 |* Use the best answer, i.e. read all choices first. |

| |Brief intro to Bayesian network: Textslides Ch14a (slide 17) |* Closed book test. Do not refer to any material during the test. |

| | |* You may need extra papers for scratch work. |

| |PROP-LOGIC contd: Forward chaining algo, Backward chaining algo. |* Use the honor code to not to talk to anyone about the test during or after the test as long as it is available online. |

| | | |

| |MySlides,; AC-3, SAT, and SAT-DPLL | |

| |SAT in Algo-complexity slides: 34-8 | |

|Feb 16 T |Automated Reasoning: First-Order Logic (FOL) (Ch 8) Syntaxes, |Regular UG project Phase 1.1 submission |

| |ForAll and ThereExists quantifiers, writing FOL statements, |due date announced |

| |equivalences (Ch8, up to sl 23) | |

|Feb 18 R |Test1-Search: Q15, correct answer is a or b, (tie-braking in | |

| |favor of a). Not c. Grades are incremented by 1. | |

| | | |

| |Automated Reasoning: (Ch 8 sl 24, 26, 28) substitution, caveat of| |

| |expressing in FOL; | |

| |(Ch 9) inferencing in FOL: Completeness-Herbrand Universe, | |

| |Unification, Forward chaining algo | |

|Feb 23 T |DPLL algorithm | |

| | |Occur check: Substitution like x/S(x) not allowed in most logic programming languages, Unification-algo syntactically checks for that and |

| |AUTOMATED REASONING: Backward chaining algo (sl#40-66) |fails in that case. Otherwise, x/S(S…(x)…) infinite looping – source of semi-decidability of FOL Occur check makes an algorithm unsouond! |

| |CNF, Resolution Algo (p255), Horn Clause, Definite clause; | |

| |Examples of logic prog languages | |

|Feb 25 R |Examples of logic prog languages |UG Project-1 Due 3-1-2021, upload on Canvas. |

| | | |

| |SAT definition, from my Algo slides, DPLL algorithm |AI & Ethics short essay due 3-8-M-11am |

| | | |

| | | |

| |Probabilistic REASONING: Ch13-Textslides-up to sl 20 | |

|Mar 02 T |Probabilistic REASONING contd: Ch13-Textslides-from sl 20 |Classwork/Homework on Predicate Logic |

| |(planned) | |

|Mar 04 R |Probabilistic Reasoning: Bayesian network |Neurology groups meeting at 7:30pm before class |

|Mar 09 T |Class got cancelled unexpectedly - |AI & Ethics short essay due 3-8-M-11am |

|Mar 11 R |-Will have class, and may continue past the hour- |Test-2 on Automated Reasoning / Logic |

| |Graduate + Spl UG presentations |3/11/R/10pm-3/12/F/11:59am Eastern |

| |(~10 min each group, have slides/demo) |(35+min, 16 questions, including logicEnglish, |

| | |Unification, CNF conversion, …) |

| | | |

| | |Presentations, mostly appeared in line with other semesters. |

| | |You guys have a good sense of the problem and data. |

| | |It should pick up steam from now. |

| | |Some specific group-wise comments below. |

|Mar 16 T |Sheetal-Parth (OCR) project meeting @7:30pm |ProbReasoning-HW1 |

| | |Due 3-18-R-11:59pm |

| |Inferencing with Bayesian network Ch 14. | |

| | | |

| |(Sample exercise Ch13: 13.8, 14, 15 ) | |

|Mar 18 R |Meeting two groups @7:30pm: Paloma-Mike (IVC), Blake-James | UG regular project (Bayes Net) |

| |(Bio-sequence HMM) |Phase 1.2 due date below |

| | | |

| |Machine Learning, from text-slide Ch18 first |Classwork/Homework2: Apply Enumerate-Ask |

| | |Algorithm to draw the call-tree for |

| | |P(j | m), John called given Marry called, similar |

| | |to that on slide 6. Compute the value using the tree. |

|Mar 23 T |Meeting two Special UG groups @7:30pm: | |

| | | |

| |MACHINE LEARNING MySlides: Decision tree, basics | |

|Mar 25 R |Meeting two groups @7:30pm: Blake-James (Bio-sequence HMM), |Homework-2 on Prob reasoning on canvas, |

| |Fred-Chris (Affine motion detection) |submission open through 3/26/F |

| | | |

| |MACHINE LEARNING: linear classification, Perceptron, Artificial | |

| |neural network (ANN) | |

|Mar 30 T |SciComp Fall2021: cs.fit.edu/~dmitra/SciComp | |

| | | |

| |MACHINE LEARNING: ANN from text, Non-parametric learning kNN | |

| |variations | |

| |LSH (start - slide 51), kernel-regression | |

|Apr 01 R |No Class, work on your project, report due soon |Probabilistic Reasoning test this day. 5pm-10am Closed book test. I am planning to use the Canvas' Proctoring option - you need to turn on |

| | |your camera. Choose the best answer. You may need a calculator and scratch paper. Usual upper-lower case semantics apply, e.g., Cavity for |

| | |the proposition, cavity for Cavity=True. |

| | |~ tilde or, - dash may be used for negation. |

|Apr 06 T |Contd. ML Unsupervised learning: | Grad projects intermediate report due |

| |Clustering basics: K-means, from |UG regular Phase1.2 |

| | | |

| |K-median, Slide 64 hierarchical, density-based, hierarchical | |

|Apr 08 R |Constraint Reasoning Ch. 5 (in Final exam) | |

|Apr 13 T |Project intermediate report discussion. | |

| |Well done, so far! | |

| | | |

| |BI-CLUSTERING, Fuzzy logic (in Machine learning module) | |

| | | |

| |A few of myslides on Constraint Reasoning. | |

| | | |

|Apr 15 R |I may get to the class room = may get late to return back home, |Machine learning test Thursday 9+pm through Friday mid-day., |

| |if the room/building not open |I may still try for proctored test |

| | |after figuring out “onboarding |

| |AI and Ethics (my slides are included in syllabus/tests) | |

| | |FINAL EXAM IS COMPREHENSIVE – all modules |

|Apr 20 T |Graduate presentation+demo, up to ~20 min each group (SpecialUG |Graduate Project Final report due |

|(Our last class) |groups no need to present) |April 27-T by 11:59pm on Canvas, zipped file (7zip): |

| | |(i) A report between 4-10 pages, |

|(Apr 22-R Last |A make up ML test for four absentees tonight, may deduct points. |including an abstract and references, |

|day of class, | |(ii) Your source code including |

|Grad presentation|Please do the instructor review. |instructions on how to run and any dependencies and |

|may continue) | |required library details, |

| | |and (iii) if relevant, data files with metadata information |

| | |(to understand data). |

| | |All three in a zip file, named using: |

| | |Sp21_project name’s signature id_one group member’s name |

| | |(iv) Add your final presentation to the package. |

| | | |

| | |Undergraduate report as below, same due date. |

| | | |

| | |FOR GRADING I USE MY OWN SPREADSHEET. |

| | |ON CANVAS, THERE IS A AGGREGATE |

| | |ASSIGNMENT (NULL) THAT CONTAINS |

| | |FORMULAS |

|April 22 R |Optional meet at 8pm: unfinished grad group presentations, any | |

| |grad student planning to continue, any UG volunteer presentation | |

| |(no grades), any class/test question, … | |

| |Above updated for Spring 2021 | |

| (Dec-2 Last day | | |

|of class) | | |

|Exam: |

|Class: 8 p.m. Tuesday and Thursday Exam: Thursday, April 29, 8:30-10:30 p.m. |

Graduate Project:

Submissions includes proposal, intermediate presentation+report, final presentation+report. Details will be posted later. I value your efforts and learning than your results. However, good results may lead to publication.

Project proposal (define the project’s scope, describe your current understanding of the project, its input-output, data source and data augmentation, expected results, references, possible code-source, etc.) is due Feb 5/F’21. Do not copy/paraphrase from past projects.

1. Indus-Valley Civilization OCR with deep learning.

Twenty-six original IVC-images are there in the zip directory.

Train a model to detect the bounding boxes for Graphemes (characters) in the text.

(Similar to SVHN project: , Goodfellow’s (2013) work and google up on Detecto project.)

You will need to create labels first by creating bounding boxes around each grapheme in original images. See:

/Fall2020/murrayjordan_1830146_42861297_Fa20_IRVC_MurrayNguyen.zip

This group used Detecto pre-trained model. I am not sure it is flexible enough for our purpose. Also, they had trouble understanding data augmentation requirements. Do not do too much rotation, low angle rotation is good enough (mimicking an archaeologist taking picture in the field). Rather, include small 3D rotation, if you can.

You will need to augment data for creating large enough training set from 22 original images (some set aside for validation). Augmentation should be with low angle rotation and scaling in 3D. Imagine an archeologist taking a picture of a tablet in the field. A relevant code is in: /Fall2020/sp20_motif_Qi_mike(1).7z

This previous group used RetinaNet, but they only created bounding box around the whole text, rather than each grapheme/character. Also, they used reflection for data augmentation that is meaningless for us.

Additional image resources from Asko Parpola’s corpus:

Two volumes of Parpola's collections seem available from ASI as e-books. Here they are:





This also means, ASI is likely to have the authority to provide us permission.

Otherwise, I am more confident now to obtain permission from the publisher.

The third volume is in print, so, possibly some libraries will have it around.



You may like to check this paper too: “Deep Learning the Indus Script.” Satish Palaniappan1, Ronojoy Adhikari. PLOS submission, arXiv:1702.00523v1 [cs.CV] 2 Feb 2017.

Tons of images here below, you need to extract them. Ignore the document’s dictionary.

cs.fit.edu/~dmitra/ArtInt/2021Spring/INDUS SCRIPT DICTIONARY-HARDBOD.pdf

This report was not on OCR, but had some proposed-Unicode for a few characters: Report_IndusValleyScript_FnlProjRep_AleeshaMishra_RahulDevMishra.pdf

I also downloaded Indus Script’s Unicode. Here it is: IndusUnicode-n1959.pdf

3/9/T: Expecting to see input data ready and some code output

3/12: Two routes, 1- use R-CNN code for 4-corner bounding box, 2- use x% overlap (say, 20%) for quality control on rectangular bounding box. I recommend the second option. At Inference do not provide too-much-rotated images. Primary objective for you should be not bounding box detection, but OCR.

Group 1. Peter Thomas pthomas2019

Brad Costa (wants IVC) bcosta2017

3/12: Detect single IVC characters. Choose 5 of them (plus unknown) for recognition. You need large augmented data set after you crop those 5+1 types of characters from images. No bounding box detection for you.

Group 2. Michael Hon mhon2014

Paloma Achu pvela2016

2. English handwriting optical character recognition with deep learning. Many commercial tools are available. You may explore and report on them, if free trials are available. However, you should use some free code/model to adapt for your project.

3/9/T: Expecting to see input data ready and some code output

3/12: Absent Sheetal. Present at 7:30pm on 3/16/T, before class.

Group 3. Parth Panchal ppanchal2021

Sheetal Ghodake sghodke2020

3. Bio-sequence prediction with Hidden Markov Model (HMM). Use an HMM tool to predict protein or nucleic-acid sequences. Evolution of SARS-CoV-2 is not very rapid (compared to, say, flu virus). Yet, it does evolve. Predicting future sequences may help vaccine developments. An HMM may be trained from existing sequences to predict new sequences with the model. Read up a little bit about gene, nucleotide sequences, and protein (amino acid sequence). Here is a report from the last semester: /Fall2020/ProjectFinalReport-CSE5400.docx However, you do not have to use the same sequences he used. For example, you may use evolution of SARS-CoV-2 genes, especially, the receptor-binding-domain, or flu-virus genes, or HIV. Dig up for such sequence databases.

3/9/T: Expecting to see input data ready and specific generative HMM implementation to work with for new sequence prediction

3/12/R: Data choice looks good to me. Get an HMM working!

3/25/R: Make generative HMM for sequence prediction. You can try that on some simple English words! Developing a validation regime needs deep thinking!

Group 4. Blake Janes bjanes2013@my.fit.edu

James Riswick-Esttelle jriswickeste2018@my.fit.edu

4. Affine motion modeling with deep learning. A 2D object gets translated, rotated, or scaled over time that is periodic: frequency higher than that of sampling, and low amplitude. Resulting image is blurred from motion. Train a CNN model for each transformation type, to predict the motion model parameters: direction and frequency. You will need image generator to create data sets. OpenCV / PILO may have affine transformation generators, but you need to create periodic motion with them. Little advanced: produce the motion free original image, in addition to estimating motion-model parameters.

The original image to work with is here: /2021Spring/Heart_image_82.png

A nice tutorial here, but do not use this code, as yours should be neural network based:

3/9/T: Expecting to see input motion data ready with ground truth, & network structure (input-output layers)

3/12/R: I recommended dropping milestones 3 and 4. Focus on periodic motion parameter’s detection.

3/26/F: Do forget about periodicity.

I was thinking deeper about affine-parameter detection.

You can train from a single input (layer) image, rather than two images.

The training set should contain images with parameter=0, and those with non-zero.

The output, of course, looks like non-zero (x,y) for translation, (theta, +/-) for rotation, and (lambda, +/-) for scaling.

Let me know for more clarification.

Group 5. Fred Gamble fgamble2021@my.fit.edu

Chris Millsap (merge with the above group) cmillsap2013@my.fit.edu

5. Recoding UNITY game engine to apply Genetic Algorithm for optimal rocket landing on lunar surface

3/12/R: No recommendation. Great job! Full group should present. Task sharing between partners.

Group 6. Michael Hoadley mhoadley2018@my.fit.edu

David Nieves-Acaron dnievesacaro2018@my.fit.edu

Nandith Narayan nnarayan2018@my.fit.edu

Special undergraduate project / Junior Design project: Submissions and deadlines will be the same as for the graduate students as above.

Some conscious patients in ICU can hear but cannot speak, often because of partial paralysis or tubings/masks in mouth and nose. Some of these patients die without being capable of communicating. Can we understand at least a few words of them from their EEG signals?

Task: train a deep learning model with brain EEG signals to understand a small fixed set of words or phonemes from whispers or loud thoughts.

References:

UC, San Francisco and Facebook project:

Here is a relevant past Senior Design project:

cs.fit.edu/~dmitra/ArtInt/2021Spring/(dm)BMES2019poster.pptx , cs.fit.edu/~dmitra/ArtInt/2021Spring//2021Spring/BMES_Abstract.pdf

A reference paper on bioRxiv on reconstructing image from brain waves:



Also, google up “Convolutional Neural Network Based Approach Towards Motor Imagery Tasks EEG Signals Classification”

NAPLib toolkit from ColumbiaU: cs.fit.edu/~dmitra/ArtInt/2021Spring/ColumbiaNaplibTool-IEEEicassp2017.pdf

David Cisek is available to talk on night or on weekends.

(1) Search for 2) such systems of your choice and store them with your knowledge base and to answer queries. One of the systems would be simple “burglar alarm” example from that you can hand-compute for debugging your system. Get started with Ch 13 (my slides on probabilistic reasoning) before we cover that material in class.

Phase 1: 1.1. (DUE: 3-1-2021-M-11:59pm) Implement first a layered DAG in an object-oriented fashion in any language of your choice. Your code will input: number of nodes, nodes as Boolean variables, arcs, and probability tables on nodes (prior probability on source nodes, conditional probabilities on other nodes). Add enough error checking facilities to ensure valid input, e.g., probability distribution not adding up to one. Limits on the architecture: nodes 7-10, depth 3-5, arcs 7-15. Phase 1.1 submission on Canvas, source code of the graph object, well commented to understand the design. Optional: an one page design of the object.

1.2. (DUE: 4-5-2021-M-W-11:59pm) The graph object should be able to input these values and answer queries by computing over the probabilities from the tables on nodes. Inferencing algorithms are in Ch14.1-2, Russell-Norvig Text 3rd Ed. (Ch 13 for 4th ed). There are 3 algorithms, recursive, variable elimination or dynamic programming, and Stochastic. I suggest, you code the first algorithm. You must follow the book’s Enumeration-Ask algorithm, with possible “memoisation” if you know it. Using downloaded code is ok as long as it follows the book’s algorithm, but follow my guidelines as above, refer-modify-describe-own.

Phase 1.2 submission on Canvas, source code of the graph object (update, if you want, from the previous phase) plus the inferencing method, well commented to understand the design. Debug with the Burglary example computing P(j | m). Optional: Output the call tree as in Homework-2.

Phase 2 (April 25-Sun by 9pm, on canvas, code+ brief design/description of 3 KBs + 3x3 query answer screen shot): Design three separate knowledge bases as asked above, and answer minimum three queries on each of the 3 knowledge bases. Extend your last report with any change in the Bayesian Networks. Screen dump, with possible explanations.

The network may be abstract, and values may be hypothetical, although examples from real life would be nice to have! This is a real piece of AI.

UG REGULAR GROUPS

Alexander Esenwein aesenwein2018@my.fit.edu

Hussein Okasha hokasha2016@my.fit.edu

Ryan Schwieterman rschwieterma2018@my.fit.edu

Joshua Quinto jquinto2018@my.fit.edu

Justyn Diaz diazj2016@my.fit.edu

Nick Cottrell ncottrell2019@my.fit.edu

Daniel Wall dwall2018@my.fit.edu

Caelan Shoop cshoop2018@my.fit.edu

Appleby, Colbi J. cappleby2018@my.fit.edu

Price, Tyler T. tprice2015@my.fit.edu

Burns, Calvin J. cburns2017@my.fit.edu

Cepeda, Carlos ccepeda2018@my.fit.edu

Dewey, John C. jdewey2018@my.fit.edu

Diaz, Justyn A. diazj2016@my.fit.edu

Grondin, Kevin M. kgrondin2018

Ze Cao zcao2018

Lerner, Zev zlerner2018

Binsalman, Saeed Omar A sbinsalman2019

-----------------------

|[pic] |[pic] |

Layered networks. The left ones are DAGs, but the right one is not a DAG. Your project should follow the DAG architecture with limits mentioned in the project description below. Your project is NOT on neural network.

|[pic] |

How Probability tables look like on nodes.

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download