Exam I Flashcards

Question

Embarrassingly parallel programming →

Answer 1

Large number of nearly identical jobs (easy) No communication between them (independent)

Answer 2

Communication between processors/nodes Transferring data takes time. Not always a speed up Time intensive and much more complex

Answer 3

Work divided into completely independent pieces with no need for communication between them Each processors works on its own section w/o needing to check in with others Runs its own section of the data and then combines it ***highly efficient and scalable However…many system require communication for programming (more complex)

Answer 4

secure shell protocol allow remote access and file transfers on HPC systems ssh: secure gateway to remote computing resources Allows you to log into a remote machine and run commands as if you were sitting at the remote terminal Highly secure tunnel from your computer to the super computer (everything encrypted)

Answer 5

secure copy Allows you to securely copy file to and from a remote system (between local & remote systems) Specifies destination while remaining secure

Answer 6

prevent software conflict in HPC environments If all programs are available at the same time, they might conflict ** lets you set up with programs you'll run on the compute nodes Need to specify what software tools are required Selects software needed for a particular task to maintain order in complex envt

Answer 7

IDLE: node not being used MIXED: some processors are being used, some are idle ALLOCATED: node is fully busy

Answer 8

Allows users to monitor and manage their active HPC jobs Monitors work, manages resources, troubleshoot problems (realtime job visibility)

Answer 9

Must create a BASH script with certain elements: 1. Comments with instructions to SLURM 2. Module commands to load the software you need 3. UNIX commands to run the software

Answer 10

1. Job name 2. output filename 3. # of compute nodes 4. # tasks per node 5. time needed 6. cluster 7. partition

Answer 11

- computers improve through experience (data) - machines find patterns in the data --- use patterns to do something (prediction, classification, clustering) - automates/scales up tedious tasks

Answer 12

- programming with a list of specific instructions - learning through experience and finding patterns in the data to be used

Answer 13

- Improves diagnostics, binding predictions, and structure analysis - scales up repetitive tasks that require intelligence

Answer 14

Ex/ identifying biomarkers/symptoms → identifies patient disease Ex/ using crystal structure of protein → predict structure binding Ex/ use amino acid seq → predict secondary structure it belongs to

Answer 15

- uses labeled data to train predictive models - data is LABELED with features (right answers are known ahead of time) - learns to get the right label by matching patterns in the data ** some set of features (#) that describe real things

Answer 16

- learning discovers patterns without labeled data - NO LABELS - goal is NOT to map data features to right answers *** goal is to make sense of the data ex/ grouping data with similar characteristics

Answer 17

- separate training and testing ensures UNBIASED evaluation of models - BEFORE TRAINING = Set aside some data for testing - then use training data to teach ML model (new info) - then test the model on testing data set aside initially - training data trains - testing data tests/evaluates ML model's performance

Answer 18

- models capture noise instead of real patterns - TOO MANY FEATURES - data isn't truly representative - finds random patterns in training data that doesn't generalize all scenarios (doesn't predict accurately)

Answer 19

- models fail to learn from the data - not enough features that the learner can distinguish one set of descriptors from another

Answer 20

Categorize the data into a limited number of predefined classes Ex/ does molecule bind or not? Ex/ does brain have tumor type A, B, and C?

Answer 21

Continuous output (not a category) Ex/ binding affinity Ex/ predicted age of death based on certain conditions *** anything can be turned into a classification problem by setting a range/right answers

Answer 22

classify data through hierarchical rules The machine learns a series of nested decision rules ***SUPERVISED -- divides the data iteratively until it knows that category the examples go into

Answer 23

Advantage: people can understand the machine’s “thinking” by looking at the rules Disadvantage: overfitting is easy

Answer 24

Simplifies the use of many ML algorithms Implements so many machine-learning algorithms Done much to democratize ML [increased accessibility] Uses pip and conda to control software downloading

Answer 25

Support vector machines (SVM)classify data using optimal separating boundaries SUPERVISED ML used for classification and regression tasks GOAL ⇒ find line that maximizes margins (far away from lines as possible) ** line that can best separate 2 things/patterns in the data

Answer 26

Enhance SVM performance by transforming data Characteristics might lend themselves to a linear separation calculates similarity between 2 data points to take non-linear relationships and transforming them into linear ones Transform the data with a kernel function; you can get a better linear separation Separates data into patterns/relationships BC most data WILL NOT be easily seperated by lines **both supervised and unsupervised

Answer 27

Groups data points based on similarity UNSUPERVISED ML Given objects with n characteristics each, can we divide those objects into k clusters Specifies beforehand – then data tries to fit into the categories It's an iterative process that uses a mathematical distance measure to assign each data point to the nearest cluster centroid