- 전체보기(71)
- paper-review(55)
- Digital Pathology(24)
- survival analysis(15)
- ssl(14)
- study(13)
- Representation Learning(10)
- math(9)
- Neuroscience(8)
- RNN(6)
- compositionality(4)
- multiple-instance-learning(4)
- bayesian(4)
- RL(3)
- Graph(3)
- memory(3)
- cognitive-map(3)
- psychophysics(2)
- XAI(2)
- curriculum learning(1)
- Book review(1)
- language(1)
[Paper Review] Signatures of Criticality in Efficient Coding Networks
This paper studies two big ideas in neuroscience: criticality and efficient coding
[Paper Review] Recurrence resonance - noise-enhanced dynamics in recurrent neural networks
random noise enhances exploring multiple attractors
[Paper Review] Fragmentation of grid cell maps in a multicompartment environment
In this paper, they recorded neural activity in grid cells and place cells when rats ran through a hairpin maze.
[Probabilistic Machine Learning] Hidden Markov Model (forward-backward algorithm)
HMM (forward-backward algorithm) study and implementation
[Paper Review] Emergence of Hidden Capabilities: Exploring Learning Dynamics in Concept Space
concept space, signal strength, learning dynamics and trajectories, phenomenological model, sudden transition in learning phase, underspecification
[Paper Review] Approximation and Optimization Theory for Linear Continuous-Time Recurrent Neural Networks
curse of memory: the necessity of memory decay for approximation
[Paper Review] Understanding and Controlling Memory in Recurrent Neural Networks
curriculum learning, attractor formation
[Paper Review] A complexity-based theory of compositionality
tried to quantify the concept of compositionality through the lens of Kolmogorov complexity, algorithmic information quantity
[Paper Review] The Simplicity Bias in Multi-Task RNNs: Shared Attractors, Reuse of Dynamics, and Geometric Representation
they observed the fundamentals of RNN behavior - simplicity bias, a tendency to share attractors and reuse dynamics.
[Paper Review] Flexible multitask computation in recurrent networks utilizes shared dynamical motifs
task period dynamics, representation, task compositionality
[Paper Review] Task representations in neural networks trained to perform many cognitive tasks
task representation, compositionality
[Paper Review] Decision making, movement planning and statistical decision theory & Statistical decision theory and the selection of rapid, goal-directed movements
Decision making, motor task, uncertainty 관련해서 2000년대 꽤나 현재 연구의 초석이 되는 두 논문을 간단히 정리한 내용이다.
[Paper Review] Signal-dependent noise determines motor planning
Motor planning 시 (control) signal-dependent noise 와 함께 final position variance 를 minimize 하는 방향으로 trajectory 가 optimize 된다는 이론을 설명한다.
[Paper Review] Transferring structural knowledge across cognitive maps in humans and models
cognitive map 을 학습하고 이용하는 과정에서 abstract structural knowledge 를 추출하고 이를 다른 task 에 transfer 한다는 것을 모델링, 행동 실험 결과로 보인 논문이다.
[Paper Review] The Tolman-Eichenbaum Machine: Unifying Space and Relational Memory through Generalization in the Hippocampal Formation
structural abstraction, sensory experience 의 conjuction 으로써 relational memory 를 표현하는 model (TEM) 을 통해 타 모델에선 불가능했던 분석을 수행하고 remapping 과 관련된 새로운 관찰 얻음
[Probabilistic Machine Learning] Gaussian filtering and smoothing, part 2
Kalman Smoother, Unscented Kalman Filter (with sigma points)
[Paper Review] Mildly Conservative Q-learning for Offline Reinforcement Learning
MCQ 방식은 기존 방식에서의 value function underestimation (pessimism) 을 보완하고자 OOD action 에 대해 pseudo target value 를 할당하여 model 이 suboptimal traj 에 stitch 하도록 함
[Paper Review] Offline Reinforcement Learning with Implicit Q-Learning
Implicit Q-Learning (IQL) 은 in distribution action Q value 의 상위 expectile 을 예측하는 model 을 통해 offline RL 의 distributional shift 를 피하고자 한 model-free 방식.
[Paper Review] Conservative Q-Learning for Offline Reinforcement Learning
Conservative Q-Learning (model-free offline reinforcement learning)
[Probabilistic Machine Learning] Gaussian filtering and smoothing, part 1
Probabilistic Machine Learning 의 Chapter 8 인 Gaussian filtering and smoothing 에 대해 정리하고자 한다. 이번 포스트에는 8.4 Inference based on the unscented transform 까