SECTION / 01
Projects.
Software and research projects.
56 ENTRIES
01BiasLoopA diffusion-model diagnostic that repeatedly feeds generated images back through the same model, then tracks how visual trajectories and image embeddings shift with loop strength.02Continual CLIPA continual-learning extension for CLIP-like models that adds visual concepts from a small set of examples by updating latent representations instead of fine-tuning the model weights.032D Drawing MachineA DIY pen plotter built around Arduino-controlled stepper motors, with Python tools for manual control, image-to-line conversion and G-code generation.04Drift Model SimulatorAn interactive 2D simulation of the drifting field from Generative Modeling via Drifting. Data particles attract generated samples while generated particles repel one another, making the paper’s particle dynamics visible in real time.05FractalTokenizerAn adaptive image tokenizer that combines corpus-level variable-byte-pair encoding with fractal quadtree partitioning. It creates variable-size patches intended to balance spatial detail with the information each patch carries.06SceneGraph-LLMGenerates scene graphs from images by combining LLM-created synthetic supervision with a vision-language model for object relationships. The output represents detected entities and their relationships as structured graph data.07RGB-D SLAM with RealSense D415A real-time RGB-D mapping and camera-tracking project using an Intel RealSense D415. The system combines color and depth observations to estimate camera motion and build a 3D map.08PruneForge: Zero-Training LLM Pruning for AgentsPruneForge searches for a task-specific, smaller language model by removing MLP channels, attention heads and layers while checking output similarity against a small prompt set. The pruning loop uses no fine-tuning and saves an inspectable pruning recipe.09Region-Coarsening Image EmbeddingsA controlled image-tokenization study that merges spatially connected superpixel regions and measures when Vision Transformer embeddings begin to drift. The repository compares representation stability across encoders and coarsening levels.10NER-LLM: Label-Free, Zero-Shot Named-Entity RecognitionA zero-shot named-entity recognition pipeline that creates synthetic examples with larger language models, then adapts a compact Qwen model with LoRA or DoRA to return entities as structured JSON.11Graphs as Scalable Symbolic Structures for Natural TextA text-to-graph-to-text workflow for extracting entities and relations from open-domain prose. It evaluates graph structure and reconstructed text to explore when symbolic graphs can support scalable reasoning over natural language.12BitDepthUpscaler: Safe Residual DequantizationA perceptual dequantization framework that reconstructs 16-bit-per-channel images from 8-bit inputs while respecting the original quantization intervals. It targets smoother tone edits and reduced false contouring without inventing values inconsistent with the source.13Orthogonality-ExplainabilityAn explainability study that constrains a selected convolutional layer to an orthogonal filter basis and examines whether its channels behave as more independent visual concepts. The implementation uses manifold-aware optimization and compares concept-level explanations.14Self-Explaining Classifiers with Sparse Autoencoders and LLMsA classifier explanation pipeline that fits a sparse autoencoder to an intermediate CNN feature map, turns learned activations into human-readable concepts, and uses those concepts to explain individual predictions.15SketchRecognition: Sketch-to-Face Retrieval and ReconstructionA sketch-to-face retrieval system that maps forensic-style sketches into a frozen ArcFace embedding space. The resulting sketch encoder can query an existing face-embedding index without requiring the gallery to be re-encoded.16LipReading with 3D Face Keypoints and Cross-Attention AdaptersA visual speech-recognition prototype that conditions a frozen causal language model on normalized 3D face landmarks. Gated cross-attention adapters inject the landmark sequence as conditioning tokens for lip-reading.17Mutual Information Adaptive Label SmoothingInformation Matching adapts label smoothing to the normalized mutual information between an original image and its augmented view. The experiments test whether this signal improves image-classification robustness under blur, noise, dropout and other degradations.18Diffusion Beam SearchAn exploration of search-based sampling for diffusion models. Best-of-N and beam-search variants branch across candidate denoising paths, then compare candidates to investigate whether search can improve generated image quality.19Chain-of-Debate on LLMA multi-agent reasoning system in which a leader coordinates specialist agents, asks follow-up questions and synthesizes competing arguments into a final answer. The repository implements a structured debate workflow for language-model agents.20GPT Image CorrectorA transformer-based image-to-image model for correcting visual artifacts in images generated by ChatGPT Image Generator, including warm color casts and excessive sharpness. The repository presents paired before-and-after examples.21Robot Dog Meta VR ControllerA virtual-reality teleoperation demonstration for a quadruped robot. The project connects a Meta Quest controller and a Unity scene to a Python control loop for remote robot movement.22Quadruped Robot Object Detection Person FollowingA quadruped-robot demo that uses YOLOv11 detections to identify a person and drive a follow behavior in real time. The project combines onboard perception with mobile-robot control.23Robot Arm VR Remote ControlA remote-control prototype that maps Oculus VR controller input through Unity, Python and Arduino to a physical robotic arm. It explores low-latency interaction between a virtual controller and hardware.24Reinforcement Learning: Proximal Policy OptimizationA from-scratch Proximal Policy Optimization implementation trained on Flappy Bird, with Snake explored as a next game environment. The project makes the policy-learning loop tangible through game play and agent behavior.25Conceptual Framework for Learning and CognitionA formal framework for describing how a system acquires information, forms knowledge and uses that knowledge to interpret and act. The project organizes learning and cognition concepts into a unified model.26ViT-AssemblerA self-supervised Vision Transformer learns image representations by solving a jigsaw-style pretext task: image patches are shuffled and the model predicts their original positions. The repository includes the model overview and training experiments.27CERMatchCERMatch is a Python utility for evaluating OCR output with Character Error Rate and matching predicted words to ground-truth words. It provides word-level matching alongside aggregate recognition-error measurements.28PixLore: A Dataset-Driven Approach to Rich Image CaptioningPixLore investigates richer image captions by fine-tuning BLIP-2 with LoRA on a curated and augmented dataset. The project asks whether smaller specialist captioners can contribute complementary detail to image understanding.29EchoSight: Unified Embedding Space for Audio and VideoEchoSight explores a shared embedding space for audio and video, inspired by CLIP-style contrastive learning. The goal is to represent paired sound and visual content for cross-modal search and related retrieval tasks.30VQVAE Toy ImplementationA compact PyTorch study implementation of a Vector-Quantized Variational Autoencoder, trained on FFHQ face images. It exposes the encoder, codebook and decoder path through a small, inspectable model.31Masked Autoencoders Toy ImplementationA small implementation of Masked Autoencoders for inspecting the self-supervised image-reconstruction workflow. Random image patches are hidden from the encoder and reconstructed by the decoder.32Denoising Diffusion Toy ImplementationA hands-on PyTorch implementation of denoising diffusion with a U-Net and attention. It trains on image datasets and exposes the forward-noise and reverse-generation process for experimentation.33Fish Pond Augmented RealityAn interactive augmented-reality fish pond projected onto a floor or surface. A camera and laser-based input detect movement, letting visitors disturb the projected fish and water.34Self-Supervised Learning with VigRegA minimal PyTorch implementation of VICReg, a self-supervised learning method that trains image representations using invariance, variance and covariance regularization.35Vision Transformers PyTorch Toy ImplementationA from-scratch PyTorch implementation of the original Vision Transformer for image classification. It breaks the architecture into patch embedding, transformer encoder and classification-head components.36Homeostatic Neural NetworksA toy implementation of homeostatic neural networks that adapts internal activity in response to distribution or concept shift, following the paper “Need Is All You Need.”372D Neural Fields RenderingAn image-fitting experiment with 2D neural fields: a coordinate-based model learns image color from spatial position, while input gradients reveal how rapidly the represented signal changes.38Whack-a-Mole Augmented Reality GameA projector-and-depth-camera augmented-reality game that places interactive targets on the floor and lets players hit them in a physical play space.39Drone Person FollowingAn early person-following prototype built around the DJI Tello mini-drone. It combines basic computer-vision tracking with drone movement commands to keep a person in view.40AdaIN Style TransferA Keras implementation of Arbitrary Style Transfer in Real-Time using Adaptive Instance Normalization. It transfers feature statistics from a style image into content features to create stylized outputs.41Unsupervised Domain AdaptationA PyTorch reproduction of Unsupervised Domain Adaptation by Backpropagation. A gradient-reversal layer trains a feature encoder to classify source labels while reducing the feature signal that reveals domain identity.423×3 Rubik’s Cube Augmented Reality SolverA real-time augmented-reality Rubik’s Cube solver that detects cube colors, estimates the cube state and overlays a 3D solution representation. It was built as a computer-vision exercise without deep learning.43PokemonGANA Wasserstein GAN with gradient penalty trained to generate new Pokémon-like designs. The repository documents sample generations from the trained model.44Super Resolution SRGANA flower-image super-resolution experiment using a GAN with perceptual loss. The model compares a low-resolution input with its generated high-resolution result.45Augmented Reality 3D Model Scene PlacementA markerless augmented-reality demo that uses AKAZE keypoints and homography to keep a 3D model anchored to a scene as the camera moves and returns to the same view.46Snake Game in BASICA compact Snake game written in MM-Basic for the Maximite ColorMax 2. The repository preserves the source for a small retro-computing game.47Reverse Quantization in Language Model PretrainingA GPT-2-sized study of whether a BF16 finishing phase can recover validation quality after simulated NVFP4 training. On four H100 GPUs, a single-seed experiment found partial recovery: switching to BF16 after 80% of training recovered about 55% of the measured loss gap, but remained behind the full-BF16 baseline. NVFP4 is simulated here, and the result still needs multi-seed replication.48Latent ChessA work-in-progress investigation into learned latent representations for chess decisions and searchless play. The intended approach is mathematically motivated, but the current system does not work and available benchmark artifacts do not establish playing strength. The accessible project folders contain checkpoints, datasets and outputs, but no implementation source.49Distilled Routed Top-K Sparse AutoencodersAn experimental routed sparse autoencoder and Neuron Atlas for probing Gemma and CLIP features. Early causal ablations are mixed: several language features move one next-token score in the expected direction, while most selected vision features have no measurable effect. Treat the labels and explanations as hypotheses, not validated interpretations.50Brainz: EEG Experiments and Creative ToolsAn experimental EEG toolkit built around a BrainAccess MAXI headset: signal-quality checks, motor and affect experiments, a multi-scale brain-state map, generative piano, and an EEG-triggered Arduino lamp prototype. Single-participant results remain weak and do not support reliable cognitive or emotional inference.51A Hierarchical Minecraft AgentA research fork of MineStudio exploring a three-system Minecraft agent: verifier-backed mechanics, STEVE-1 motor control, and higher-level task planning. In a paired 100-episode stone-acquisition screen, one targeted checkpoint reached 66% success versus 37% for the vanilla reference. Later router and PPO attempts regressed, the live full hierarchy remains unvalidated, and the latest iron-pickaxe run completed none. A work in progress with mixed results.52Neuron Manifold ExplorerA research toolkit for mapping visual-feature responses across realistic generated-image modes and matched semantic interventions. A saved ImageNet-100 study probes ResNet-50 layer4.2 channel 899 with 800 FLUX.2 Klein image pairs; natural-image generalization remains unfinished, and strong color and lighting controls limit texture-only conclusions.53LTX-2.5 Video Restoration and DepixelationFine-tunes a frozen LTX-2.5 video model with rank-64 LoRA on FFHQ and COCO restoration trajectories. The completed FFHQ run improved four metrics on 10 held-out faces, while a one-image COCO check showed mixed signals: better LPIPS and MS-SSIM, but worse pixel-error metrics. Broader restoration quality and identity preservation remain unproven.54Fractal Merge CriticA sequential rate–distortion policy that learns where and when to merge tokens in frozen transformers, using multi-round damage estimates and a bounded learned merge operator. A small CLIP pilot showed better ranking at aggressive compression but regressed at mild compression, found no calibrated safe merges, and added more overhead than the default policy could recover. The tested Qwen pilot underperformed cosine. Scale and deployment gates remain red.55FractalTokenizer: Adaptive Image TokenizationResearch code exploring image tokens allocated by estimated information. A one-seed v2 ImageNet100 run reached 75.02% top-1 at 128 tokens, below a 256-token grid baseline at 76.32%; a later modified-density rerun scored 73.52%. The deterministic v3 tokenizer has synthetic tests, but no completed matched-rate downstream result. A rate-distortion advantage has not been established.56Feature GenesisA research pipeline for tracing sparse neural features through training and testing their causes with deterministic replay. Its saved QuickDraw ResNet-34 run passed the artifact gate and a three-feature circuit pilot, while example labels remain provisional and replication across model seeds and data domains is unfinished.