Public library

I have a public Zotero library where I archive and catalogue the papers I read, contact me if you wish to be added to the library. Follows a catalogue of the tags that I use.

Tags

We have several classes of tags, each identified by its token cardinality, ranging from one (single-tag or 1-tag) to n (n-tag). Single-tags are comprised of a single phrase (possibly with spaces), such as survey or use case and have no tag terminator. Conversely, n-tags (2-tag, 3-tag, …) are a sequence of colon-separated tags comprised of a prefix and suffix:

1
prefix:suffix

prefix is a general tag, e.g. data:, scope:, while suffix is a more specific tag, e.g. local, text, feature importance. They are defined in BNF form.

Example:

1
2
Prefix scope:
Suffix local | global | local to global

generates tags

1
2
3
4
scope:local
scope:global
scope:local to global
scope:sub-global

n-tags allow for varying levels of specificity and compositionality. Moreover, when dealing with unclear or overly general papers the prefix can be omitted. For instance, if we are dealing with a paper on explanatory interactive learning (prefix xil:) and the type of explanation provided to the user is not specified we can just use the tag xil:. On the other hand, if the paper specifies it, we can insert it as a suffix, e.g. xil:feature:importance, xil:decision rules.

Tag usage across collections: the alg tag

Some tags belong to some collections, e.g. the gen:autoencoder tag for the ml/generative models/autoencoder collection, but they can (and are) used in other collections too, provided they are prefixed by a alg: prefix. The alg: prefix indicates that the given tag is not the focus of the paper, rather an accessory algorithm/technique/feature.

Say we have a paper on local explainability leveraging autoencoders (AE) like ABELE: an obvious tag to use would be gen:autoencoder. Still, the paper merely leverages existing autoencoder literature to define a XAI algorithm. If we were to use gen:autoencoder here, we could not tell apart papers on autoencoders (tag gen:autoencoder) from papers merely leveraging autoencoders (tag alg:gen:autoencoder). To separate the two classes of papers, we simply add a alg: prefix to the gen:autoencoder tag.

General tags

As seen in “Tag usage across collections: the alg tag”, alg: is a key tag used across collections to add accessory information. Some tags do not belong to any collection, thus we use them as 2-tags with an alg:

Very cool/important papers

star

Papers on software tools/frameworks

software

Theoretical papers

theoretical

Explanation derived/inspired from…?

alg:

1
shapley values | shap | lime

Algorithm based on…?

alg:

1
backpropagation | input perturbation | knowledge graph

XAI: must-have tags!

Papers in any XAI in collection. Should not use a alg: prefix when using outside these collections.

What is the explanation scope?

scope:

1
local | global | local to global

Data: what’s the model input data?

data:

1
any | text | image | tabular | sequence | shapelet | graph | health

Model: is the XAI algorithm model-agnostic? Is it explaining a specific model? A model for a specific task?

explains:

1
2
3
4
5
6
7
8
9
10
11
model-agnostic | by design |
linear model |
nn | rnn | cnn | transformer | neurons |
svm |
latent space |
language model | qa |
clustering |
ensemble | forest |
kb:completion
information retriever | recommendation system |
reinforcement agent

• kb:completion explains a knowledge base completion model

What kind of explanation is provided?

Is it a counterfactual? Feature importance (saliency and word importance included)? Does it explain the interaction between features? Does it provide the importance of the neurons in a neural model? Does it provide decision rules (decision trees included)?

explanation:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
aut:grammar | aut:automaton |
caption |
counterfactual |
decision rules | oblique decision rules |
feature:importance | feature:interaction |
graph nodes | graph edges |
input importance |
knowledge graph nodes | knowledge graph relations |
multi-modal |
natural language |
neuron importance |
prototypes |
proof |
shapelet |
structural equations |
taxonomy |
visualization

• aut:automaton an automaton
• aut:grammar a (first-order, context-free, etc.) grammar
• caption caption for an image
• counterfactual a counterfactual
• decision rules
• feature:importance what important are the input features?
• feature:interaction how are sets of features correlated?
• prototypes a (set of) prototypical example(s)
• graph nodes nodes in a graph
• graph edges edges in a graph
• knowledge graph nodes nodes in a knowledge graph
• knowledge graph relations edges in a knowledge graph
• latent dimension
• multi-modal multi-modal explanations for reinforcement agents
• natural language a natural language explanation
• neuron importance important or significant neurons
• structural equations feature equations in causal models
• taxonomy

Task: what task is the black box solving?

Given that most tasks are n-ary classification, we omit the tag for classification problems.

task:

1
2
3
4
5
6
7
8
9
10
11
benchmarking |
dfa distillation |
entailment trees |
interactive xai |
masking |
multiclass classification |
probing | prompting | prompt tuning | knowledge editing |
natural language inference | qa | fact checking |
regression | ordinal regression |
st graphs |
kg-to-text | concepts-to-text

• benchmarking generic papers on benchmarking (XAI included)
• dfa distillation distilling a DFA from a model
• entailment trees models who construct proof-like natural language trees to reach a goal
• interactive xai XAI with interactive human intervention
• prompting probing masked language models for knowledge
• probing generic papers on probing techniques
• prompt tuning papers on injecting knowledge in language models through prompts
• knowledge editing papers on editing beliefs of language models
• kg-to-text papers on generating text from knowledge graphs
• concepts-to-text papers on generating text from given concepts
• st graphs structured graphs of belief

Explanations: definitions & desiderata and evaluations & benchmarking

exp:

1
definition | validation | benchmark | critique

• definition what is an explanation?
• validation how to validate explanations?
• benchmark best explanations?
• critique a critique of explanation algorithms

Discrete spaces: automata, languages and symbols

Papers on grammars, deterministic finite state automata (DFA), epsilon deterministic finite state automata (eps-DFA), regular and context-free grammars

aut:

1
2
3
grammar | grammar:regular | grammar:context-free | grammar:temporal |
automata:dfa | automata:multiset | automata:symbolic |
symbolic learning

• grammar
• regular grammar regular grammars
• context-free grammar context free grammars
• temporal grammar temporal grammars
• dfa deterministic finite state automata
• eps-dfa epsilon deterministic finite state automata
• sym-dfa symbolic dfa
• symbolic learning learning (from) discrete symbols

Logic

Papers on logic programming, logic-powered models and logic programs inductors.

log:

1
2
3
4
5
6
7
markov network | logic network |
answer set programming | logic programming | inductive logic programming |
inductive logic programming:meta |
theorem proving |
reasoning |
answer set programming |
adaptive | fuzzy

• markov network markov networks
• logic network neural models performing logic operations
• logic programming logic programming
• inductive logic programming learning logic programs from data
• inductive logic programming learning programs of logic programs from data
• theorem proving can we prove a theorem?
• reasoning reasoning in neural models
• adaptive adaptive logic
• fuzzy fuzzy logic

Argumentation

Papers on argumentation.

arg:

Polyhedra and combinatorics

Papers on polyhedra and combinatorics: search algorithm, CSP, SAT, MAXSAT and Set Cover problem, Mixed Integer Programming and Polyhedral learning

cmb:

1
2
3
4
5
6
7
search |
csp |
mip | ilp |
optimal transport |
sat | maxsat |
set cover |
polyhedral learning | polytope | polytope approximation

• search search in a combinatorial space
• csp constraint satisfaction problems
• optimal transport optimal transport problems
• mip mixed integer programming problems
• ilp integer linear programming problem
• sat satisfiability problems
• maxsat maximum satisfiability problems
• set cover set coverage problems
• polyhedral learning learning polyhedra

Computer vision

cv:

1
segmentation | object detection | style transfer

• segmentation finding meaningful parts of an image
• object detection detect objects in images
• style transfer applying stylistic patterns to images

Sequence

seq:

1
sax

• sax the SAX family of shapelet discretization algorithms

Clustering

clus:

1
clustering | pruning

• clustering learning a clustering
• pruning merging clusters (mainly in hierarchical clusterings)

Paper type: surveys, benchmarks and theoretical papers

Use tags survey, benchmark and theoretical appropriately.

Adversarial

Papers on adversarial attack/defense in ml/adversarial

Attack or defense?

adv:

1
attack | defense

• attack adversarial attacks
• defense defense against adversarial attacks

Latent space

Papers on embedding, latent space representation, traversal and (dis)entanglement.

lat:

1
2
3
4
dimensionality reduction |
disentanglement |
embedding | representation |
traversal | interpolation

• dimensionality reduction dimensionality reduction
• disentanglement to each latent dimension its own meaning
• embedding learning an unsupervised embedding (latent) representation
• representation learning a latent representation
• traversal percolating the latent space
• interpolation latent space interpolation

Generative models

Papers on generative models (AE, VAE, GAN, flow models, ecc) in /ml/generative models.

What’s the model class?

gen:

1
2
3
4
5
6
autoencoder | variational autoencoder |
conditioned |
flow |
gan |
process |
other

• autoencoder autoencoders
• variational autoencoder variational autoencoders
• flow flow models
• gan generative adversarial networks
• process stochastic processes (mainly gaussian)
• other other ad hoc generative models

Knowledge bases

kb:

1
2
3
4
5
common sense |
completion |
embedding |
inject | linking
mining

• common sense knowledge bases on common sense reasoning
• completion infer missing edges in a given knowledge graph
• embedding creating embeddings for a knowledge graph nodes and edges
• inject enrich model training with a given knowledge base
• linking map tokens to entities in the knowledge base/graph
• mining distilling a knowledge base from data/models

Explainable Interactive Learning (XIL)

Explainable Interactive Learning (XIL) inserts a human-in-the-loop in the training procedure of a machine learning model by periodically presenting the model to the human and prompting an explainable correction

xil:

1
2
3
4
5
counterfactuals |
decision rules | decision making |
feature:importance |
prototypes |
?

• feature:importance training black box models with explainable human intervention in the form of feature importance measures
• decision rules training black box models with explainable human intervention in the form of decision rules
• decision making humans and algorithms collaborating at prediction time
• prototypes training black box models with explainable human intervention in the form of prototypes selection
• counterfactuals training black box models with explainable human intervention in the form of counterfactual decision rules
• ? unclear

Machine Learning & Pattern Recognition

mlpr:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
backpropagation |
boosting |
boundary tree |
differentiable tree |
capsule network |
concept learning |
decision tree | forest |
distillation |
program synthesis |
graphical model
regularization | constraints |
transfer learning |
som |
training

• model training training
• distillation distill models in other models
• concept learning learning higher level features
• program synthesis papers on synthesizing programs from data
• regularization regularization
• constraints learning the set of constraints inferred by a model/constraining a model behavior
• som the self-organizing map
• transfer learning leverage existing models in different domains/datasets/etc

Computational Mathematics

Maths, duh

cm:

1
matrix factorization | piecewise linear models | voronoi diagram

• matrix factorization factorizing matrices
• piecewise linear models on piecewise linear models
• voronoi diagram on voronoi diagrams

Optimization

opt:

1
submodular optimization | genetic programming

Statistics

Ususally prefixed by alg: as the literature is not focused on statistics

stat:

1
bayes | shapley | bandits | mutual information

• shapley leveraging the Shapley values
• bandits bandits selection algorithm
• mutual information

Natural Language Processing (NLP)

nlp:

1
2
3
attention | transformer |
language model | perplexity |
qa | summarization | text similarity

• attention defining and leveraging the attention mechanism
• transformer defining and leveraging the multi-head attention mechanism for Transformers
• qa question answering

Fairness

Papers on fairness in /fairness on defining fairness, learning fair models or remove unfair behavior from trained unfair models

fairness:

1
definition | learning | cleaning

• definition formally defining a fairness measure
• learning learning fair models
• cleaning removing unfair behavior from given models