12 mars 2021, 10:30-11:30
Séminaire en ligne : https://bbb.di.ens.fr/b/cam-pz6-kdj
Traversal algorithms and heuristics for reasoning over logic based knowledge graphs (Davide Benedetto)
Knowledge Graphs (KGs) provide a concise and intuitive abstraction for a variety of domains where edges capture the (potentially recursive) relationships between the entities. This is leading to the rise of systems and tools able to facilitate graph data modeling, processing, and analysis, with prominent AI companies developing core systems based on the property graph model.
In this context, Datalog-based languages are being re-discovered to be ductile to accomplish reasoning tasks over complex property graphs as they provide the essential elements to enable graph navigational operations.
The semantics of a Datalog program is usually specified in an operational way via the chase procedure. It entails multiple non-deterministic choices such as the rule application order and the fact binding order when multiple unification is possible. In state-of-the-art reasoners, chase-based procedures are
not directly adopted, but encoded in the form of engineered variations of the volcano iterator model and so essentially within a pipe-and-filters architecture, where nodes (filters) are relational algebra operators and edges (pipes) are dependency connections between the rules. Such (potentially cyclic) structures, known as access plans, need to be translated into reasoning plans, where abstract relational algebra operators are transformed into specific project, select and join implementations: many implementations of each operator exist and it is up to the optimizer to choose the best one in terms of execution cost.
Here I focus on cases where the Datalog reasoning process involves a graph traversal task and I investigate the connection between reasoning plans and graph traversal strategies. Then I move from the observation that the nondeterministic choices posed by the chase can be leveraged to control graph traversals —allowing to alternate breadth-first and depth-first strategies— and study the link of such choices with the reasoning plans. I will conclude that in plans, specific join implementations and rule prioritization policies reflect the nondeterministic choices and exploit them to guide graph traversals in modern reasoners. Specifically, I implemented the results in the Vadalog System, a state-of-the-art knowledge graph management systems and conduct experimental evaluation.
Evolution of AI (as of 2021) (Shrey Mishra)
AI itself is very vast and not be just Machine Learning(although majority of learning is) there are other aspects to it when building a custom solution and those solutions generally tend to revolve around other subtopics such as Discrete optimisation, meta heuristics, Deep learning and NLP.
With a lot of currently available tools and frameworks it can be a bit harder to keep track of the right set of subtasks/tools depending upon the problem statement. The talk will be divided into 3 parts:
1. Introduction to various problem statements where different subforms of AI can be used ? (Examples of topics/tools and how to get started ?)
2. What are recent developments in AI ? The AI timeline of Important papers and tools and how to use them ? (Examples of important papers in the AI community)
3. How did I implement some of these on various different tasks (My masters and previous background)?
I will try to include some code snippets on some of the topics that I have covered so far when trying to solve several real world problems revolving around sentiment analysis, image detection, optimisation, deploying machine learning models etc.