1 of 17

HW3 Discussion

2 of 17

Content

  1. Model Track (2)
  2. Benchmark Track (2)
  3. Peer review
  4. Group Discussion

3 of 17

Model track - Jeong Jun Lee

Analysis of AGI - [K-KP-E]

Knowledge

Knowledge

Program

Environment

Json file

JavaScript

Program

Othello Env

4 of 17

Model track - Jeong Jun Lee

Analysis of AGI - [LLM]

Information

Distribution

Imbalance

Language

Limitation

Limitations of LLM

Lack of

Compression

Reasoning in continuous latent space

5 of 17

Model track - Jeong Jun Lee

Sketch of AGI core capability - [Generalization]

Hopfield Networks

  • Gradient descent to store patterns

  • By minimizing an Energy function

  • Use attention, but still rely on fixed weights
  • Compute connections

with other nodes implicitly

  • Store dynamic information

  • Fundamentally tend to learn

6 of 17

Model track - Jeong Jun Lee

Dynamic Node Architecture

Sketch of AGI core capability - [Generalization]

Node A

Node B

Node C

  • Node ∈ [-1, 1] updated every iteration
  • No node can be unique
  • Each neuron maintains dynamic state
  • Connection form based on mutual information
  • Real-time learning capability

7 of 17

Model track - Byeong Chang Kim

Analysis of AGI - [DIKM]

“Human-Like” & “Intelligent”

Data: Observation in market A and B.

Information: Milk is 300 won, cereal is 1000 won in A. 500, 1000 in B.

Knowledge: Milk is cheaper in A than B.

Wisdom: Coffee would be cheaper in

8 of 17

Model track - Byeong Chang Kim

Analysis of AGI - [Dreamer]

Phase 1: Collects and learns current world dynamics

Phase 2: Uses learned time-series-based world model to sequentially predict value, reward, and action

- Learns policy and critic networks without actual world interaction

Use Dream/Imagination of Humans

9 of 17

Model track - Byeong Chang Kim

Analysis of AGI - [Dreamer]

1. Manipulated Phenomenal World

2. Absence of Dialectical Thinking

3. Processing Time Limitations

Limitation

Generative Method – GAN

Contrastive Method – Positive/Negative

10 of 17

Model track - Byeong Chang Kim

Meta-cognitive – “Know what I know and I don’t know”

Sketch of AGI core capability - [Dreamer + AMAGO2]

Maintain Dreamer’s human-like characteristics

Utilize AMAGO2’s meta-cognitive capabilities (Attention)

Contrastive encoder for counter problem processing

- capability to handle dialectical thesis, antithesis, and synthesis

11 of 17

Benchmark track - Byunghwa Yoo & Jung Min Kim

Key Characteristics for AGI Evaluation

1. Generalization Ability

2. Multi-Modality

3. Cognitive Competence Implementation

4. Continual Learning

5. Usage and Prior Knowledge Integration

1.Generalization Ability (Multi-task)

2. Human-like

3. Reasoning capability

4. Application

5. Creativity

12 of 17

Benchmark track - Byunghwa Yoo & Jung Min Kim

Current Benchmark’s Limitation

- Lack of modality

- Not continual learning

- If the model train with new one, we can’t check the ability of it already had.

- Measure only properties that it want to measure

- It become task-specific while we try to solve ARC with AI. (learn every patterns used in ARC)

- Get multiple multi-modal benchmark.

To evaluate AGI, all benchmark’s environment should be different.

- Continual learning with training set with multiple benchmark.

- Get IQ-like test, and reasoning task.

- Use human’s test.

By gathering them up, we can avoid task-specific benchmark.

13 of 17

Benchmark track - Byunghwa Yoo & Jung Min Kim

Othello’s Good Point

- Prototype of AGI Benchmark

An environment in which the early stages of AGI can be tested in a specific environment

- Explainability

This is discrete grid environment. So we can know what is changed.

- Try to evaluate “Human-like” action.

We can make our own strategy and play.

- With other environments, model can upgrade their own strategy to win.

14 of 17

Benchmark track - Byunghwa Yoo & Jung Min Kim

Improvement Plan

1. Highly restricted environment

- DL based and RL based approach are impossible for this benchmark. (Due to time constraint) – Make environment that can support DL.

2. No prior-knowledge

- Model can’t use their prior knowledge when they encounter the new task

- Making a DB can be helpful to utilize prior knowledge.

1. Rules are simple

- Mono Color game

- Call only coordinate

- Combine with other games (YINSH)

- 1v1 to multiple player

15 of 17

Peer review

- You can take the HW3 with your name on the post-it and read it.

- Please read the answer sheet carefully and return it to the owner when you finished reading it.

- Discuss with author about your idea.

~3:20 pm.

16 of 17

Group discussion

17 of 17

Key Discussions

- Evaluate Snakebench or other benchmark as an AGI benchmark. (B)

- Discuss about question 2 from this HW3. Evaluate that features are well-aligned with your question 1. (B)

- Consider whether current methods of processing and representing knowledge are sufficient. (M)

- Examine whether scaling up is truly the answer (M)

- Explore how we might measure “AGI-tic” using methods other than accuracy. (M)