Knowledge Vault 2/74 - ICLR 2014-2023
Manuela Veloso ICLR 2021 - Invited Talk - AI in Finance: Scope and Examples
<Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:

graph LR classDef main fill:#f9f9f9, stroke:#333, stroke-width:1px, font-weight:bold, font-size:14px; classDef veloso fill:#d4f9d4, stroke:#333, stroke-width:1px, font-weight:bold, font-size:14px; classDef ai fill:#d4d4f9, stroke:#333, stroke-width:1px, font-weight:bold, font-size:14px; classDef finance fill:#f9d4d4, stroke:#333, stroke-width:1px, font-weight:bold, font-size:14px; classDef data fill:#f9f9d4, stroke:#333, stroke-width:1px, font-weight:bold, font-size:14px; A[Manuela Veloso
ICLR 2021] --> B[Veloso: JPMorgan AI Head,
CMU Professor 1] A --> C[AI: capture intelligence in algorithms 2] C --> D[Veloso: integrate perception, cognition, action 3] A --> E[AI in finance: exciting,
varied, data-rich 4] E --> F[JPMorgan: AI research, applied teams 5] F --> G[JPMorgan AI goals: predict, liberate,
eradicate crime, ethics 6] A --> H[Trading: visual plots guide decisions 7] H --> I[Classify time series images:
buy/no-buy signals 8] H --> J[Predict time series: image
autoencoder 9] A --> K[Multi-agent markets: reinforcement learning 10] K --> L[Market maker strategies: learned parameters 11] K --> M[Calibrate simulations to real
data: reinforcement learning 12] A --> N[AI finance areas: discovery, crime,
compliance, ethics 13] N --> O[Standardize data enables machine learning 14] O --> P[NLP extracts units, scales
for standard representation 15] A --> Q[AI finance: learn, reason
from varied data 16] Q --> R[DocuBot: interactive report generation 17] R --> S[DocuBot: learns language, generation
through interaction 18] A --> T[Crypto price classification: same
as other assets 19] T --> U[Time series code to
be open sourced 20] A --> V[Technical analysis, charting: visual reasoning 21] V --> W[Time series images: novel
AI representation 22] A --> X[Multi-modal data: capture dynamic
correlations visually 23] X --> Y[Complex agents: parameterized types,
behavior learning 24] A --> Z[AI translates data: tables,
visuals, language 25] Z --> AA[Images, numbers: different strengths,
similar prediction 26] Z --> AB[Temporal images, trajectories can
simplify learning 27] A --> AC[End-to-end AI: feasible through
modular decomposition 28] AC --> AD[Docubot: component reuse for
domain-specific generation 29] A --> AE[Theme: representation learning extracts,
translates knowledge 30] class A main; class B,D veloso; class C,E,F,G,N,Q,V,W,X,Y,Z,AA,AB,AC,AD,AE ai; class H,I,J,K,L,M,T,U finance; class O,P,R,S data;


1.-Dr. Manuela Veloso is head of JPMorgan AI Research Group and Herbert A. Simon Professor at Carnegie Mellon University.

2.-AI aims to capture every aspect of intelligence within an algorithm, including perception, cognition, and action.

3.-Dr. Veloso's research focuses on integrating perception, cognition and action, such as in autonomous robots she developed at CMU.

4.-AI in finance is exciting due to the variety of business areas and enormous amount of data produced.

5.-JPMorgan has AI research and applied AI teams working across business units to bring AI transformation to the company.

6.-AI research goals at JPMorgan include predicting economic systems, liberating data safely, eradicating financial crime, and establishing ethical AI.

7.-Financial trading decisions are guided by visual plots of time series data on asset values changing over time.

8.-Classifying time series as images of buy/no-buy signals using neural networks outperforms using the raw numerical data.

9.-Neural networks can predict future portions of a time series by representing it as an image and training an autoencoder.

10.-Over-the-counter markets with multiple interacting agents can be simulated using multi-agent reinforcement learning.

11.-Market maker agent strategies are represented by parameters like risk aversion and connectivity to investors, which are learned through simulation.

12.-Calibrating simulation parameters to match real market data is itself posed as a reinforcement learning problem.

13.-AI in finance covers many areas like information discovery, customer experience, financial crime prevention, regulation compliance, and ethics.

14.-Standardizing the representation of heterogeneous financial data enables easier application of machine learning techniques.

15.-Natural language processing extracts information like units and scales from titles and headers to convert data to a standard representation.

16.-AI in finance makes heavy use of learning from and reasoning about data represented in various forms.

17.-The DocuBot system automatically generates reports, slides, charts from data through interactive learning from user instructions and feedback.

18.-DocuBot continuously learns and improves its language understanding and document generation abilities through interaction, without relying on large training datasets.

19.-Applying image-based classification to time series like cryptocurrency prices follows the same principles as for other financial assets.

20.-Code for time series image classification is planned to be open sourced after further polish and parameterization.

21.-Technical analysis and charting are commonly used by human traders to visually reason about financial time series data.

22.-Representing time series as images for AI is a novel contribution highlighting the importance of representation learning in AI.

23.-Multi-modal data like video can capture dynamic correlations between multiple time series in a visual representation.

24.-Agents in complex domains can be represented by parameterized types, with behavior learning occurring in the parameter space.

25.-AI enables automatically translating between different data representations, like from tables to visualizations and natural language insights.

26.-Image and numerical representations have different strengths and biases, but can achieve similar predictive performance with proper techniques.

27.-Capturing temporal data as images or trajectories can simplify learning compared to sequence models like LSTMs in some cases.

28.-End-to-end AI systems are appealing but made more feasible by decomposing the problem and reusing modular components.

29.-Docubot exemplifies component reuse by applying the core engine to generate different types of documents with domain-specific templates.

30.-The overarching theme is the power of representation learning to extract and translate knowledge between modalities in AI systems.

Knowledge Vault built byDavid Vivancos 2024