ai_collection

Contravariance Theory methodology — formal proof that minimal DNN solutions to hard tasks exhibit strong alignment of privileged axes, with alignment "zipping" up the network hierarchy. Bridges NeuroAI convergent evolution theory and brain-DNN comparison methods.

Brain-Inspired Unsupervised Self-Reflection (BUS) framework for enhancing VLM reasoning without labeled data. Uses neuroscience-backed backward prediction to enable self-verification on unlabeled data.

Dynamic neural manifold architecture for flexible closed-loop control on neuromorphic hardware — mapping spiking activity to low-dimensional manifold trajectories with sensory-modulated geometry for explainable neural computation.

Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware. Uses ring attractor networks with sensory-modulated control neurons (speed, shape, selection) to drive subspace rotations and fine-grained trajectory control in neural state space. Implemented on SpiNNaker 2 chip with robotic maze navigation validation.

Graph-regularized learning framework for EEG-based emotion recognition using psychological emotion topology. Conceptualizes emotions as nodes in a graph with edges encoding proximity based on dimensional emotion theories. Use when building EEG emotion classifiers, affective BCI systems, or applying graph regularization to psychological classification tasks.

Interpretable machine learning methodology for predicting Parkinson's disease motor severity (MDS-UPDRS Part III) from neuroimaging features — Quantitative Susceptibility Mapping (QSM) MRI and multiband multiecho resting-state fMRI Regional Homogeneity (ReHo). Uses SVR, Elastic Net, Random Forest, XGBoost with nested CV and SHAP interpretability. Full multimodal model explains 45.4% variance. QSH+c clinical model achieves 75% within ±5 UPDRS points. Activation: Parkinson's prediction, QSM MRI, ReHo fMRI, MDS-UPDRS, motor severity prediction, SHAP neuroimaging, multiband multiecho fMRI, interpretable ML Parkinson, quantitative susceptibility mapping

Sample-Adaptive Hyperbolic Graph Neural Network (SA-HGNN) for EEG-based depression recognition. Combines sample-adaptive graph construction with hyperbolic graph convolution and attention pooling to capture hierarchical brain network structure in EEG signals.

STST-JEPA: Shallow-Target Spatio-Temporal Joint Embedding Predictive Architecture for EEG self-supervised learning. Largest EEG foundation model (47,703 sessions, ages 5-81) using JEPA-style latent prediction with EMA tokenizer + auxiliary signal reconstruction. Rank 1 on NeuralBench for sex, age, psychopathology. Brain age gap correlates with cognitive efficiency. Activation: stst-jepa, eeg foundation model, brain age, self-supervised eeg, eeg self-supervised learning, JEPA eeg, EEG2Rep, brain space, neuralbench, brain age gap, cognitive efficiency

Validation methodology for deep multimodal brain-encoding models (TRIBE, the 2025 Algonauts challenge winner) against behavioral engagement metrics. Tests whether predicted fMRI signals forecast aggregate population behavior (YouTube replay heatmaps, neuroforecasting). Shows predicted neural drive fails to predict re-watch despite high encoding accuracy — null result with Bayes factor bounds and equivalence tests. Activation: brain encoding validation, TRIBE, neuroforecasting, predicted fMRI, fMRI engagement, brain encoding behavioral prediction, Algonauts challenge, fMRI replay heatmap, global field power fMRI, YouTube replay

Haken Lighthouse model with adaptive conduction delays and phase locking theory. Provides analytically tractable framework for phase-locked states in delayed spiking networks, spike-time perturbation stability analysis, and activity-dependent white matter plasticity (myelination-modulated delays). Applicable to: SNN temporal coordination, communication-through-coherence, white matter plasticity modeling, delayed spiking network analysis, circulant ring networks, autapse dynamics, and slow-fast adaptive delay systems.

Learning biophysical Hodgkin-Huxley models from extracellular MEA measurements using differentiable simulation and simulation-based inference. Enables precise neurostimulation prediction from minutes of recording instead of hours of stimulus testing.

Methodology for studying magnetic field effects on chimera states in Hindmarsh-Rose neuronal networks. Covers traveling chimera, multicluster chimera, and multicluster chimera breather transformations under spatial magnetic field applications.

Dendritic In-Context Learning (DendriCL) — single-layer compartmental SNN that achieves ICL via dendritic subthreshold dynamics implementing online LMS. Use when: designing SNN architectures for in-context learning, exploring biologically-plausible ICL mechanisms, studying dendritic computation, implementing seed-stable ICL beyond moderate task dimensions, analyzing compartmental neuron models.

Causal mechanism framework for anhedonia and reward valuation deficits in Vision-Language Models — mechanistic analysis linking VLM reward processing to Nucleus Accumbens dysfunction patterns from clinical depression research.

Neuromorphic silicon neuron controller (SiLIF-DBS) for adaptive deep brain stimulation in Parkinson's Disease — CMOS-implemented closed-loop aDBS achieving 75% power reduction with beta-band biomarker tracking.

Non-Hermitian Potential Well Formalism for Conscious-Preconscious-Subliminal Processing methodology. Models the Global Neuronal Workspace (GNW) as a complex-valued landscape where sensory encoding and conscious access are unified. Activation: non-Hermitian consciousness, GNW landscape, conscious preconscious subliminal, potential well consciousness, complex-valued GNW, non-Hermitian Hamiltonian consciousness.

Non-Hermitian Potential Well Formalism for modeling the Global Neuronal Workspace (GNW) consciousness framework. Uses nonlinear Schrödinger-type equation in imaginary time with non-Hermitian, non-normal Hamiltonian and Lotka-Volterra-type term to reproduce subliminal-preconscious-conscious hierarchy. Maps conscious access to bound-state emergence in complex-valued landscape. Use for consciousness modeling, neural field theory, GNW theory, and quantum-inspired cognitive dynamics.

Quantum computing applications in finance: portfolio optimization, option pricing, risk management, financial simulations, and quantum economics using quantum algorithms (QAOA, quantum annealing, quantum Monte Carlo, amplitude estimation, entangled neural traders). Use for quantum finance research, NISQ-era financial applications, quantum advantage analysis in derivatives/derivatives pricing, and economic action constants.

Formal verification toolchain for probabilistic spiking neural networks using weight-discretized quotient abstractions. CogSpike framework integrates SNN design, simulation, and PRISM-based verification. Key contributions: weight-discretized quotient model abstraction (17x state reduction per neuron), two-sided fidelity theorem bounding firing disagreement to gray zone, Asymptotic Silence theorem guaranteeing permanent silence of unforced neurons, topology-dependent exponential state space reduction. Covers probabilistic model checking of DTMC encodings, synaptic weight discretization, verification of seven canonical topologies. Activation: formal verification, probabilistic SNN, quotient abstraction, CogSpike, PRISM, DTMC, state space explosion, synaptic weight discretization, fidelity theorem

Comprehensive survey of single-entity spiking neuron models — classification of biologically plausible neural systems including discrete and continuous analogs, membrane potential dynamics, and components affecting neural dynamics for accurate simulation of biological processes.

Model-free temporal-switch (TS) framework for transferable lightweight neuromorphic computing. Enables direct transfer of trained models to unseen hardware devices without post-training calibration by incorporating a broader spectrum of devices during training. Addresses device-to-device variations that undermine practical advantages of neuromorphic computing. Activation: temporal switch framework, neuromorphic transfer, device variation robustness, memristor reservoir computing, model-free transfer, lightweight neuromorphic, direct deployment

Treats LLM-as-judge evaluator-replacement ambiguity as a measurement-validity problem. Judge upgrades are not interchangeable. Stronger judges reduce but don't remove position/verbosity bias. Proposes audit trails including dataset slices, bias probes, and error-dependence estimates. Activation: LLM-as-judge, evaluation reliability, measurement validity, evaluator bias, AI evaluation.

Formalizes the Context Access Divide (CAD) as a dimension of agentic inequality operating at the interaction level. Dynamic Context Retrieval vs Manual Attachment causes combinatorial collapse in task-success probability. Proposes contextuality as complement to Sharp et al.'s framework. Activation: agentic inequality, context access, interaction-level architecture, agent fairness, AI equity.

Capability-driven benchmark for evaluating proactive agents in dynamic real-world settings. UniClawBench evaluates five foundational capabilities (Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, Cross-Platform Coordination) across 400 bilingual tasks in live Docker containers with closed-loop evaluation. Activation: proactive agents, real-world benchmark, agent evaluation, capability-driven, multimodal agents, closed-loop evaluation.

Latent Memory Palace (LMP): reasoning for control policies as autoregressive variational inference. Organizes information in latent memory palace with iterative adaptive retrieval. LMP-π achieves strong performance with interpretable adaptive test-time compute. Variable-length action tokenizer. Activation: latent memory, reasoning for control, autoregressive variational inference, adaptive reasoning, continuous control.

HCC-STAR: clinically aligned LLM for hepatocellular carcinoma staging, treatment, and prognosis. Reads EMR narratives, outputs risk stratification, guideline-consistent treatments with rationales, and survival estimates. Outperforms GPT-5 and Gemini-2.5 Pro. Activation: clinical-reasoning LLM, hepatocellular carcinoma, risk stratification, treatment guidance, EMR.

Multi-agent marketplace simulation studying formal mechanisms for market stability with self-interested LLM agents. 18 DeepSeek-V3 agents with complementary specialties trade in constrained network. Mediation identified as top mechanism, robust under adversarial attack. Activation: market stability, self-interested agents, multi-agent economics, cooperation mechanisms, social dilemmas.

Memory-augmented agentic framework that exposes a frozen VLA as a retryable contact-rich primitive composed with analytic primitives. Learns operating range from execution traces and failure models. +38.6pp on LIBERO-Pro, +25.4pp on RoboCasa365. Activation: vision-language-action, memory-guided agents, manipulation primitives, frozen VLA, robot manipulation.

Hierarchical machine teaching algorithm for robust reward learning across multiple MDPs. Demonstrates comparisons impose stronger constraints than demonstrations in unlimited-data regime. Greedily selects informative environments then queries low-cost feedback. Activation: machine teaching, reward learning, inverse reinforcement learning, multi-environment, robust reward, feedback modalities.

Proactive memory agent that runs alongside an unmodified action agent to prevent behavioral state decay in long-horizon tasks. Updates structured memory bank from trajectory and selectively injects reminders. Plug-and-play with frontier agents. +8.3pp on Terminal-Bench 2.0, +6.8pp on τ²-Bench. Activation: proactive memory, long-horizon agents, trajectory management, context surfacing, memory retrieval, behavioral state decay.

Studies quantile-based distributional RL from statistical efficiency perspective. Non-asymptotic error bound O(√(m/n)) under W∞ metric. Achieves optimal √n convergence rate. Asymptotic distribution and semiparametric efficiency bound. Berry-Esseen theorem. Activation: distributional RL, quantile regression, statistical efficiency, policy evaluation, return distribution.

Codebook-free binary spherical coding for extreme low-bit LLM weight compression. Maps weight chunks onto unit hypersphere and binarizes into sign streams. Residual BSQ stage for reconstruction error. Category-wise recovery distillation. Activation: binary spherical coding, LLM compression, low-bit quantization, lookup-free coding, model deployment.

Procrustes-conditioned Joint End-to-end Top-K SAE for extracting cross-seed universal features from independently trained BERT models. Combines Top-K sparsity, end-to-end optimization, and dead-feature revival. Pearson r ≥ 0.70 across seeds. Activation: sparse autoencoder, cross-seed universality, Procrustes alignment, mechanistic interpretability, feature universality.

Hierarchical structure-aware document analysis system using LLMs. Parses documents into hierarchical trees preserving layouts, builds structure-aware semantic indices for filtering and question answering. Handles academic papers, technical manuals, financial reports. Activation: document analysis, hierarchical structure, LLM system, information extraction, document understanding.

Training-free best-first draft tree for speculative decoding using Domino's conditional non-factorized correction. Achieves up to 6.6x speedup on Qwen3-4B and highest mean accept length (10.7 tokens/round). GPU-native CUDA-graph builder for efficient tree construction. Activation: speculative decoding, tree-structured drafting, Domino conditioning, LLM inference, block-diffusion.

Shows that post-training quantization evaluation via accuracy/perplexity fails to capture behavioral changes. Introduces correctness agreement metric. Reveals non-linear breakpoints at low bit-widths. Query/key projections more sensitive than value/output. Activation: quantization, LLM deployment, behavioral change, correctness agreement, post-training quantization.

MAESTRO: Markov-chain Approximated Expert Sparsification via Transition-based Routing for MoE structured pruning. Models expert activation as Ergodic Markov chains for globally-aware importance. Outperforms baselines by 10.61% at 50% compression. Lower cross-task variance. Activation: mixture-of-experts, expert pruning, MoE deployment, structured pruning, language model efficiency.

Balanced session-centric LLM scheduling for agent serving workloads. Routes first request in each session for load balance and follow-ups cache-aware. 10-16% TPS improvement. Leverages intra-session locality and 80%+ KV-reuse in agent traces. Activation: LLM scheduling, agent serving, session-centric scheduling, inference infrastructure, tokens-per-second.

Shows that Super Weight pruning degradation doesn't universally apply. Training Super Weights in isolation drops accuracy to random-guessing. Parameter importance ≠ trainability. Vanilla LoRA with 0.16% parameters succeeds. Activation: super weights, LLM training, parameter pruning, critical parameters, selective training.

Practical investigation of training-free relaxed speculative decoding for LLM inference acceleration. Unifies existing approaches within a shared framework, benchmarks on contemporary settings. Relaxed speculation trades lossless guarantees for speed-ups and controlled capability-speed trade-offs. Activation: speculative decoding, LLM inference acceleration, lossy speculation, training-free, draft verification.